{"meta": {"count": 1000, "db_response_time_ms": 35, "page": 1, "per_page": 100, "groups_count": null, "x_query": {"oql": "get works where retracted is (false)\n  and year is (2025)\n  and title-abstract has (stemmed \"GPU kernel\" OR stemmed \"KV cache\" OR stemmed \"LLM serving\" OR stemmed \"inference optimization\" OR stemmed \"inference serving\" OR stemmed \"neural network compiler\" OR stemmed \"speculative decoding\")\n  and type is (article or preprint or review);\nthen sample (1000) of those works with seed (42)", "oqo": {"get_rows": "works", "filter_rows": [{"column_id": "is_retracted", "value": false}, {"column_id": "publication_year", "value": 2025}, {"join": "or", "filters": [{"column_id": "title_and_abstract.search", "value": "\"GPU kernel\"", "operator": "has"}, {"column_id": "title_and_abstract.search", "value": "\"KV cache\"", "operator": "has"}, {"column_id": "title_and_abstract.search", "value": "\"LLM serving\"", "operator": "has"}, {"column_id": "title_and_abstract.search", "value": "\"inference optimization\"", "operator": "has"}, {"column_id": "title_and_abstract.search", "value": "\"inference serving\"", "operator": "has"}, {"column_id": "title_and_abstract.search", "value": "\"neural network compiler\"", "operator": "has"}, {"column_id": "title_and_abstract.search", "value": "\"speculative decoding\"", "operator": "has"}]}, {"join": "or", "filters": [{"column_id": "type", "value": "article"}, {"column_id": "type", "value": "preprint"}, {"column_id": "type", "value": "review"}]}], "sample": 1000, "seed": "42"}, "url": "/works?filter=is_retracted:false,publication_year:2025,title_and_abstract.search:%22GPU kernel%22|%22KV cache%22|%22LLM serving%22|%22inference optimization%22|%22inference serving%22|%22neural network compiler%22|%22speculative decoding%22,type:article|preprint|review&sample=1000&seed=42&per_page=100&page=1"}, "cost_usd": 0.001}, "results": [{"id": "https://openalex.org/W4415318248", "doi": "https://doi.org/10.48550/arxiv.2510.07293", "title": "AudioMarathon: A Comprehensive Benchmark for Long-Context Audio Understanding and Efficiency in Audio LLMs", "publication_year": 2025, "publication_date": "2025-10-08", "type": "preprint", "language": "en", "authorships": [{"author_position": "first", "author": {"id": null, "display_name": "He, Peize", "orcid": null, "observed_orcids": []}, "institutions": [{"id": "https://openalex.org/I183067930", "display_name": "Shanghai Jiao Tong University", "ror": "https://ror.org/0220qvk04", "country_code": "CN", "type": "education", "lineage": ["https://openalex.org/I183067930"]}], "countries": ["CN"], "is_corresponding": false, "raw_author_name": "He, Peize", "raw_affiliation_strings": ["Shanghai Jiao Tong University"], "raw_orcid": null, "affiliations": [{"raw_affiliation_string": "Shanghai Jiao Tong University", "institution_ids": ["https://openalex.org/I183067930"]}]}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5102635197", "display_name": "Zichen Wen", "orcid": "https://orcid.org/0009-0002-6157-5898", "observed_orcids": ["https://orcid.org/0009-0002-6157-5898"]}, "institutions": [{"id": "https://openalex.org/I183067930", "display_name": "Shanghai Jiao Tong University", "ror": "https://ror.org/0220qvk04", "country_code": "CN", "type": "education", "lineage": ["https://openalex.org/I183067930"]}, {"id": "https://openalex.org/I4391012619", "display_name": "Shanghai Artificial Intelligence Laboratory", "ror": "https://ror.org/03wkvpx79", "country_code": "CN", "type": "facility", "lineage": ["https://openalex.org/I4391012619"]}], "countries": ["CN"], "is_corresponding": false, "raw_author_name": "Wen, Zichen", "raw_affiliation_strings": ["Shanghai AI Laboratory", "Shanghai Jiao Tong University"], "raw_orcid": null, "affiliations": [{"raw_affiliation_string": "Shanghai AI Laboratory", "institution_ids": ["https://openalex.org/I4391012619"]}, {"raw_affiliation_string": "Shanghai Jiao Tong University", "institution_ids": ["https://openalex.org/I183067930"]}]}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5100383178", "display_name": "Yubo Wang", "orcid": "https://orcid.org/0000-0001-5766-3751", "observed_orcids": ["https://orcid.org/0000-0001-5766-3751", "https://orcid.org/0009-0001-2313-6299"]}, "institutions": [{"id": "https://openalex.org/I183067930", "display_name": "Shanghai Jiao Tong University", "ror": "https://ror.org/0220qvk04", "country_code": "CN", "type": "education", "lineage": ["https://openalex.org/I183067930"]}], "countries": ["CN"], "is_corresponding": false, "raw_author_name": "Wang, Yubo", "raw_affiliation_strings": ["Shanghai Jiao Tong University"], "raw_orcid": null, "affiliations": [{"raw_affiliation_string": "Shanghai Jiao Tong University", "institution_ids": ["https://openalex.org/I183067930"]}]}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5100375974", "display_name": "Yuxuan Wang", "orcid": "https://orcid.org/0000-0001-6476-9589", "observed_orcids": ["https://orcid.org/0000-0001-6476-9589", "https://orcid.org/0009-0008-2055-7785", "https://orcid.org/0009-0006-6417-5207"]}, "institutions": [{"id": "https://openalex.org/I183067930", "display_name": "Shanghai Jiao Tong University", "ror": "https://ror.org/0220qvk04", "country_code": "CN", "type": "education", "lineage": ["https://openalex.org/I183067930"]}], "countries": ["CN"], "is_corresponding": false, "raw_author_name": "Wang, Yuxuan", "raw_affiliation_strings": ["Shanghai Jiao Tong University"], "raw_orcid": null, "affiliations": [{"raw_affiliation_string": "Shanghai Jiao Tong University", "institution_ids": ["https://openalex.org/I183067930"]}]}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5103231035", "display_name": "Xiaoqian Liu", "orcid": "https://orcid.org/0000-0002-2814-3404", "observed_orcids": ["https://orcid.org/0000-0002-2814-3404"]}, "institutions": [{"id": "https://openalex.org/I12912129", "display_name": "Northeastern University", "ror": "https://ror.org/04t5xt781", "country_code": "US", "type": "education", "lineage": ["https://openalex.org/I12912129"]}, {"id": "https://openalex.org/I183067930", "display_name": "Shanghai Jiao Tong University", "ror": "https://ror.org/0220qvk04", "country_code": "CN", "type": "education", "lineage": ["https://openalex.org/I183067930"]}], "countries": ["CN", "US"], "is_corresponding": false, "raw_author_name": "Liu, Xiaoqian", "raw_affiliation_strings": ["Northeastern University", "Shanghai Jiao Tong University"], "raw_orcid": null, "affiliations": [{"raw_affiliation_string": "Northeastern University", "institution_ids": ["https://openalex.org/I12912129"]}, {"raw_affiliation_string": "Shanghai Jiao Tong University", "institution_ids": ["https://openalex.org/I183067930"]}]}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5045031730", "display_name": "Jiajie Huang", "orcid": "https://orcid.org/0000-0002-7788-249X", "observed_orcids": ["https://orcid.org/0000-0002-7788-249X"]}, "institutions": [{"id": "https://openalex.org/I183067930", "display_name": "Shanghai Jiao Tong University", "ror": "https://ror.org/0220qvk04", "country_code": "CN", "type": "education", "lineage": ["https://openalex.org/I183067930"]}], "countries": ["CN"], "is_corresponding": false, "raw_author_name": "Huang, Jiajie", "raw_affiliation_strings": ["Shanghai Jiao Tong University"], "raw_orcid": null, "affiliations": [{"raw_affiliation_string": "Shanghai Jiao Tong University", "institution_ids": ["https://openalex.org/I183067930"]}]}, {"author_position": "middle", "author": {"id": null, "display_name": "Lei, Zehui", "orcid": null, "observed_orcids": []}, "institutions": [{"id": "https://openalex.org/I183067930", "display_name": "Shanghai Jiao Tong University", "ror": "https://ror.org/0220qvk04", "country_code": "CN", "type": "education", "lineage": ["https://openalex.org/I183067930"]}], "countries": ["CN"], "is_corresponding": false, "raw_author_name": "Lei, Zehui", "raw_affiliation_strings": ["Shanghai Jiao Tong University"], "raw_orcid": null, "affiliations": [{"raw_affiliation_string": "Shanghai Jiao Tong University", "institution_ids": ["https://openalex.org/I183067930"]}]}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5104287356", "display_name": "Zhuangcheng Gu", "orcid": null, "observed_orcids": []}, "institutions": [{"id": "https://openalex.org/I74973139", "display_name": "Carnegie Mellon University", "ror": "https://ror.org/05x2bcf33", "country_code": "US", "type": "education", "lineage": ["https://openalex.org/I74973139"]}], "countries": ["US"], "is_corresponding": false, "raw_author_name": "Gu, Zhuangcheng", "raw_affiliation_strings": ["Carnegie Mellon University"], "raw_orcid": null, "affiliations": [{"raw_affiliation_string": "Carnegie Mellon University", "institution_ids": ["https://openalex.org/I74973139"]}]}, {"author_position": "middle", "author": {"id": null, "display_name": "Jin, Xiangqi", "orcid": null, "observed_orcids": []}, "institutions": [{"id": "https://openalex.org/I183067930", "display_name": "Shanghai Jiao Tong University", "ror": "https://ror.org/0220qvk04", "country_code": "CN", "type": "education", "lineage": ["https://openalex.org/I183067930"]}], "countries": ["CN"], "is_corresponding": false, "raw_author_name": "Jin, Xiangqi", "raw_affiliation_strings": ["Shanghai Jiao Tong University"], "raw_orcid": null, "affiliations": [{"raw_affiliation_string": "Shanghai Jiao Tong University", "institution_ids": ["https://openalex.org/I183067930"]}]}, {"author_position": "middle", "author": {"id": null, "display_name": "Yang, Jiabing", "orcid": null, "observed_orcids": []}, "institutions": [{"id": "https://openalex.org/I4210165038", "display_name": "University of Chinese Academy of Sciences", "ror": "https://ror.org/05qbk4x57", "country_code": "CN", "type": "education", "lineage": ["https://openalex.org/I19820366", "https://openalex.org/I4210165038"]}], "countries": ["CN"], "is_corresponding": false, "raw_author_name": "Yang, Jiabing", "raw_affiliation_strings": ["University of Chinese Academy of Sciences"], "raw_orcid": null, "affiliations": [{"raw_affiliation_string": "University of Chinese Academy of Sciences", "institution_ids": ["https://openalex.org/I4210165038"]}]}, {"author_position": "middle", "author": {"id": null, "display_name": "Li, Kai", "orcid": null, "observed_orcids": []}, "institutions": [{"id": "https://openalex.org/I99065089", "display_name": "Tsinghua University", "ror": "https://ror.org/03cve4549", "country_code": "CN", "type": "education", "lineage": ["https://openalex.org/I99065089"]}], "countries": ["CN"], "is_corresponding": false, "raw_author_name": "Li, Kai", "raw_affiliation_strings": ["Tsinghua University"], "raw_orcid": null, "affiliations": [{"raw_affiliation_string": "Tsinghua University", "institution_ids": ["https://openalex.org/I99065089"]}]}, {"author_position": "middle", "author": {"id": null, "display_name": "Liu, Zhifei", "orcid": null, "observed_orcids": []}, "institutions": [{"id": "https://openalex.org/I183067930", "display_name": "Shanghai Jiao Tong University", "ror": "https://ror.org/0220qvk04", "country_code": "CN", "type": "education", "lineage": ["https://openalex.org/I183067930"]}], "countries": ["CN"], "is_corresponding": false, "raw_author_name": "Liu, Zhifei", "raw_affiliation_strings": ["Shanghai Jiao Tong University"], "raw_orcid": null, "affiliations": [{"raw_affiliation_string": "Shanghai Jiao Tong University", "institution_ids": ["https://openalex.org/I183067930"]}]}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5100613916", "display_name": "Weijia Li", "orcid": "https://orcid.org/0000-0003-0876-9773", "observed_orcids": ["https://orcid.org/0000-0003-0876-9773"]}, "institutions": [{"id": "https://openalex.org/I157773358", "display_name": "Sun Yat-sen University", "ror": "https://ror.org/0064kty71", "country_code": "CN", "type": "education", "lineage": ["https://openalex.org/I157773358"]}, {"id": "https://openalex.org/I4391012619", "display_name": "Shanghai Artificial Intelligence Laboratory", "ror": "https://ror.org/03wkvpx79", "country_code": "CN", "type": "facility", "lineage": ["https://openalex.org/I4391012619"]}], "countries": ["CN"], "is_corresponding": false, "raw_author_name": "Li, Weijia", "raw_affiliation_strings": ["Shanghai AI Laboratory", "Sun Yat-sen University"], "raw_orcid": null, "affiliations": [{"raw_affiliation_string": "Shanghai AI Laboratory", "institution_ids": ["https://openalex.org/I4391012619"]}, {"raw_affiliation_string": "Sun Yat-sen University", "institution_ids": ["https://openalex.org/I157773358"]}]}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5010758235", "display_name": "Cunxiang Wang", "orcid": "https://orcid.org/0000-0002-3023-8082", "observed_orcids": ["https://orcid.org/0000-0002-3023-8082"]}, "institutions": [{"id": "https://openalex.org/I99065089", "display_name": "Tsinghua University", "ror": "https://ror.org/03cve4549", "country_code": "CN", "type": "education", "lineage": ["https://openalex.org/I99065089"]}], "countries": ["CN"], "is_corresponding": false, "raw_author_name": "Wang, Cunxiang", "raw_affiliation_strings": ["Tsinghua University"], "raw_orcid": null, "affiliations": [{"raw_affiliation_string": "Tsinghua University", "institution_ids": ["https://openalex.org/I99065089"]}]}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5101615091", "display_name": "Conghui He", "orcid": "https://orcid.org/0000-0001-8697-695X", "observed_orcids": ["https://orcid.org/0000-0001-8697-695X"]}, "institutions": [{"id": "https://openalex.org/I4391012619", "display_name": "Shanghai Artificial Intelligence Laboratory", "ror": "https://ror.org/03wkvpx79", "country_code": "CN", "type": "facility", "lineage": ["https://openalex.org/I4391012619"]}], "countries": ["CN"], "is_corresponding": false, "raw_author_name": "He, Conghui", "raw_affiliation_strings": ["Shanghai AI Laboratory"], "raw_orcid": null, "affiliations": [{"raw_affiliation_string": "Shanghai AI Laboratory", "institution_ids": ["https://openalex.org/I4391012619"]}]}, {"author_position": "last", "author": {"id": "https://openalex.org/A5100689115", "display_name": "Linfeng Zhang", "orcid": "https://orcid.org/0000-0002-4239-7953", "observed_orcids": ["https://orcid.org/0000-0002-4239-7953"]}, "institutions": [{"id": "https://openalex.org/I183067930", "display_name": "Shanghai Jiao Tong University", "ror": "https://ror.org/0220qvk04", "country_code": "CN", "type": "education", "lineage": ["https://openalex.org/I183067930"]}], "countries": ["CN"], "is_corresponding": false, "raw_author_name": "Zhang, Linfeng", "raw_affiliation_strings": ["Shanghai Jiao Tong University"], "raw_orcid": null, "affiliations": [{"raw_affiliation_string": "Shanghai Jiao Tong University", "institution_ids": ["https://openalex.org/I183067930"]}]}], "abstract_inverted_index": {"Processing": [0], "long-form": [1, 66], "audio": [2, 30, 80, 99, 127, 171, 180, 187], "is": [3], "a": [4, 55, 70], "major": [5], "challenge": [6], "for": [7, 158], "Large": [8], "Audio": [9], "Language": [10], "models": [11, 14, 42, 182], "(LALMs).": [12], "These": [13], "struggle": [15], "with": [16, 24, 82], "the": [17, 137, 156, 170], "quadratic": [18], "cost": [19], "of": [20, 73, 95, 139, 184], "attention": [21], "($O(N^2)$)": [22], "and": [23, 38, 62, 108, 110, 121, 135, 142, 154, 162, 172], "modeling": [25], "long-range": [26], "temporal": [27, 160], "dependencies.": [28], "Existing": [29], "benchmarks": [31], "are": [32], "built": [33, 75], "mostly": [34], "from": [35, 85], "short": [36], "clips": [37], "do": [39], "not": [40], "evaluate": [41, 59, 118], "in": [43], "realistic": [44], "long": [45], "context": [46], "settings.": [47], "To": [48], "address": [49], "this": [50], "gap,": [51], "we": [52], "introduce": [53], "AudioMarathon,": [54], "benchmark": [56], "designed": [57], "to": [58, 87, 92, 97, 176], "both": [60], "understanding": [61, 181], "inference": [63], "efficiency": [64], "on": [65], "audio.": [67], "AudioMarathon": [68, 167], "provides": [69], "diverse": [71], "set": [72], "tasks": [74], "upon": [76], "three": [77], "pillars:": [78], "long-context": [79], "inputs": [81], "durations": [83], "ranging": [84], "90.0": [86], "300.0": [88], "seconds,": [89], "which": [90], "correspond": [91], "encoded": [93], "sequences": [94], "2,250": [96], "7,500": [98], "tokens,": [100], "respectively,": [101], "full": [102], "domain": [103], "coverage": [104], "across": [105, 151], "speech,": [106], "sound,": [107], "music,": [109], "complex": [111, 186], "reasoning": [112, 161], "that": [113], "requires": [114], "multi-hop": [115], "inference.": [116], "We": [117, 130, 165], "state-of-the-art": [119], "LALMs": [120, 153], "observe": [122], "clear": [123], "performance": [124], "drops": [125], "as": [126], "length": [128], "grows.": [129], "also": [131], "study": [132], "acceleration": [133], "techniques": [134], "analyze": [136], "trade-offs": [138], "token": [140], "pruning": [141], "KV": [143], "cache": [144], "eviction.": [145], "The": [146], "results": [147], "show": [148], "large": [149], "gaps": [150], "current": [152], "highlight": [155], "need": [157], "better": [159], "memory-efficient": [163], "architectures.": [164], "believe": [166], "will": [168], "drive": [169], "multimodal": [173], "research": [174], "community": [175], "develop": [177], "more": [178], "advanced": [179], "capable": [183], "solving": [185], "tasks.": [188]}, "primary_topic": {"id": "https://openalex.org/T10201", "display_name": "Speech Recognition and Synthesis", "score": 0.7153000235557556, "subfield": {"id": "https://openalex.org/subfields/1702", "display_name": "Artificial Intelligence"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, "topics": [{"id": "https://openalex.org/T10201", "display_name": "Speech Recognition and Synthesis", "score": 0.7153000235557556, "subfield": {"id": "https://openalex.org/subfields/1702", "display_name": "Artificial Intelligence"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, {"id": "https://openalex.org/T11309", "display_name": "Music and Audio Processing", "score": 0.21130000054836273, "subfield": {"id": "https://openalex.org/subfields/1711", "display_name": "Signal Processing"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, {"id": "https://openalex.org/T10860", "display_name": "Speech and Audio Processing", "score": 0.03280000016093254, "subfield": {"id": "https://openalex.org/subfields/1711", "display_name": "Signal Processing"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}], "primary_location": {"id": "pmh:oai:arXiv.org:2510.07293", "is_oa": true, "landing_page_url": "http://arxiv.org/abs/2510.07293", "pdf_url": "https://arxiv.org/pdf/2510.07293", "source": {"id": "https://openalex.org/S4306400194", "display_name": "arXiv (Cornell University)", "issn_l": "2331-8422", "issn": ["2331-8422"], "is_oa": true, "is_in_doaj": false, "is_core": false, "listed_in": [], "host_organization": "https://openalex.org/I205783295", "host_organization_name": "Cornell University", "host_organization_lineage": ["https://openalex.org/I205783295"], "host_organization_lineage_names": [], "type": "repository"}, "license": "public-domain", "license_id": "https://openalex.org/licenses/public-domain", "version": "submittedVersion", "is_accepted": false, "is_published": false, "raw_source_name": null, "raw_type": "text"}, "cited_by_count": 0, "is_retracted": false}, {"id": "https://openalex.org/W4416118560", "doi": "https://doi.org/10.48550/arxiv.2504.05097", "title": "State Tuning: State-based Test-Time Scaling on RWKV-7", "publication_year": 2025, "publication_date": "2025-04-07", "type": "preprint", "language": "en", "authorships": [{"author_position": "first", "author": {"id": "https://openalex.org/A5100714256", "display_name": "Xiao Liu", "orcid": "https://orcid.org/0000-0003-4673-8273", "observed_orcids": ["https://orcid.org/0000-0003-4673-8273"]}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Xiao, Liu", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5100337955", "display_name": "Zhiyuan Li", "orcid": "https://orcid.org/0009-0002-1727-8597", "observed_orcids": ["https://orcid.org/0009-0002-1727-8597"]}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Zhiyuan, Li", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "last", "author": {"id": "https://openalex.org/A5116054097", "display_name": "Lin Yueyu", "orcid": null, "observed_orcids": []}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Yueyu, Lin", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}], "abstract_inverted_index": {"Test-time": [0], "scaling": [1, 35, 193], "has": [2], "emerged": [3], "as": [4, 206], "a": [5, 24, 52, 106, 123, 166], "prominent": [6], "research": [7], "direction": [8], "in": [9, 214], "machine": [10], "learning,": [11], "enabling": [12], "models": [13, 172], "to": [14, 44, 56, 65, 109, 126, 136, 146, 194], "enhance": [15], "their": [16], "expressive": [17], "capabilities": [18], "during": [19], "inference.Transformers,": [20], "renowned": [21], "for": [22, 210], "striking": [23], "delicate": [25], "balance": [26], "between": [27], "efficiency": [28, 181], "and": [29, 111, 155], "expressiveness,": [30], "have": [31], "benefited": [32], "from": [33], "test-time": [34, 57, 192], "techniques": [36], "that": [37, 104, 165], "leverage": [38], "an": [39, 101, 207], "expanding": [40], "key-value": [41], "(KV)": [42], "cache": [43], "significantly": [45], "improve": [46], "performance.In": [47], "this": [48], "paper,": [49], "we": [50, 60, 99, 121, 141, 163], "introduce": [51], "novel": [53], "state-based": [54], "approach": [55, 92], "scaling,": [58], "which": [59], "term": [61], "state": [62, 114, 130, 150, 161, 204], "tuning,": [63], "tailored": [64], "the": [66, 71, 82, 87, 113, 117, 129, 133, 148, 160, 174, 180, 183, 189, 201], "RNN-based": [67], "RWKV-7": [68, 118, 185], "model.By": [69], "exploiting": [70], "unique": [72], "strengths": [73], "of": [74, 116, 182, 191, 203], "RWKV-7,": [75], "our": [76], "method": [77, 125, 178], "achieves": [78], "state-of-the-art": [79], "performance": [80, 213], "on": [81, 94, 173], "target": [83], "task": [84], "without": [85], "altering": [86], "model's": [88, 134], "pre-trained": [89], "weights.": [90], "Our": [91, 198, 217], "centers": [93], "three": [95], "key": [96], "innovations.": [97], "First,": [98], "develop": [100], "observer": [102], "framework": [103], "allows": [105], "smaller": [107, 167], "model": [108, 168, 212], "replicate": [110], "learn": [112], "dynamics": [115], "model.": [119], "Second,": [120], "employ": [122], "kernel": [124], "dynamically": [127], "upscale": [128], "size,": [131], "enhancing": [132], "capacity": [135], "capture": [137], "intricate": [138], "patterns.": [139], "Third,": [140], "integrate": [142], "Decorrelated": [143], "Backpropagation": [144], "(DBP)": [145], "optimize": [147], "upscaled": [149], "matrix,": [151, 162], "thereby": [152], "improving": [153], "convergence": [154], "expressivity.": [156], "By": [157], "tuning": [158, 205], "only": [159], "demonstrate": [164], "can": [169], "outperform": [170], "larger": [171], "given": [175], "task.": [176], "This": [177], "preserves": [179], "original": [184], "architecture": [186], "while": [187], "harnessing": [188], "power": [190], "deliver": [195], "superior": [196], "results.": [197], "findings": [199], "underscore": [200], "potential": [202], "effective": [208], "strategy": [209], "advancing": [211], "resource-constrained": [215], "settings.": [216], "code": [218], "is": [219], "https://github.com/TorchRWKV/flash-linear-attention.": [220]}, "primary_topic": {"id": "https://openalex.org/T10036", "display_name": "Advanced Neural Network Applications", "score": 0.4336000084877014, "subfield": {"id": "https://openalex.org/subfields/1707", "display_name": "Computer Vision and Pattern Recognition"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, "topics": [{"id": "https://openalex.org/T10036", "display_name": "Advanced Neural Network Applications", "score": 0.4336000084877014, "subfield": {"id": "https://openalex.org/subfields/1707", "display_name": "Computer Vision and Pattern Recognition"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, {"id": "https://openalex.org/T10028", "display_name": "Topic Modeling", "score": 0.26759999990463257, "subfield": {"id": "https://openalex.org/subfields/1702", "display_name": "Artificial Intelligence"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, {"id": "https://openalex.org/T11307", "display_name": "Domain Adaptation and Few-Shot Learning", "score": 0.20900000631809235, "subfield": {"id": "https://openalex.org/subfields/1702", "display_name": "Artificial Intelligence"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}], "primary_location": {"id": "pmh:oai:arXiv.org:2504.05097", "is_oa": true, "landing_page_url": "http://arxiv.org/abs/2504.05097", "pdf_url": "https://arxiv.org/pdf/2504.05097", "source": {"id": "https://openalex.org/S4306400194", "display_name": "arXiv (Cornell University)", "issn_l": "2331-8422", "issn": ["2331-8422"], "is_oa": true, "is_in_doaj": false, "is_core": false, "listed_in": [], "host_organization": "https://openalex.org/I205783295", "host_organization_name": "Cornell University", "host_organization_lineage": ["https://openalex.org/I205783295"], "host_organization_lineage_names": [], "type": "repository"}, "license": "cc-by", "license_id": "https://openalex.org/licenses/cc-by", "version": "submittedVersion", "is_accepted": false, "is_published": false, "raw_source_name": null, "raw_type": "text"}, "cited_by_count": 0, "is_retracted": false}, {"id": "https://openalex.org/W4410128284", "doi": "https://doi.org/10.2139/ssrn.5242694", "title": "Beyond Kv Caching: Shared Attention for Efficient Llms", "publication_year": 2025, "publication_date": "2025-01-01", "type": "preprint", "language": "en", "authorships": [{"author_position": "first", "author": {"id": "https://openalex.org/A5057076890", "display_name": "Bingli Liao", "orcid": "https://orcid.org/0009-0005-0474-7464", "observed_orcids": ["https://orcid.org/0009-0005-0474-7464"]}, "institutions": [], "countries": [], "is_corresponding": true, "raw_author_name": "Bingli Liao", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "last", "author": {"id": "https://openalex.org/A5028815763", "display_name": "Danilo Vasconcellos Vargas", "orcid": "https://orcid.org/0000-0001-7442-1279", "observed_orcids": ["https://orcid.org/0000-0001-7442-1279"]}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Danilo  Vasconcellos Vargas", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}], "abstract_inverted_index": null, "primary_topic": {"id": "https://openalex.org/T10028", "display_name": "Topic Modeling", "score": 0.5663999915122986, "subfield": {"id": "https://openalex.org/subfields/1702", "display_name": "Artificial Intelligence"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, "topics": [{"id": "https://openalex.org/T10028", "display_name": "Topic Modeling", "score": 0.5663999915122986, "subfield": {"id": "https://openalex.org/subfields/1702", "display_name": "Artificial Intelligence"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, {"id": "https://openalex.org/T10181", "display_name": "Natural Language Processing Techniques", "score": 0.1770000010728836, "subfield": {"id": "https://openalex.org/subfields/1702", "display_name": "Artificial Intelligence"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, {"id": "https://openalex.org/T10036", "display_name": "Advanced Neural Network Applications", "score": 0.17499999701976776, "subfield": {"id": "https://openalex.org/subfields/1707", "display_name": "Computer Vision and Pattern Recognition"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}], "primary_location": {"id": "doi:10.2139/ssrn.5242694", "is_oa": true, "landing_page_url": "https://doi.org/10.2139/ssrn.5242694", "pdf_url": null, "source": {"id": "https://openalex.org/S4210172589", "display_name": "SSRN Electronic Journal", "issn_l": "1556-5068", "issn": ["1556-5068"], "is_oa": true, "is_in_doaj": false, "is_core": false, "listed_in": [], "host_organization": "https://openalex.org/I1318003438", "host_organization_name": "RELX Group (Netherlands)", "host_organization_lineage": ["https://openalex.org/I1318003438"], "host_organization_lineage_names": [], "type": "repository"}, "license": null, "license_id": null, "version": "acceptedVersion", "is_accepted": true, "is_published": false, "raw_source_name": null, "raw_type": "posted-content"}, "cited_by_count": 3, "is_retracted": false}, {"id": "https://openalex.org/W4416777430", "doi": "https://doi.org/10.48550/arxiv.2511.20714", "title": "Inferix: A Block-Diffusion based Next-Generation Inference Engine for World Simulation", "publication_year": 2025, "publication_date": "2025-11-25", "type": "preprint", "language": null, "authorships": [{"author_position": "first", "author": {"id": "https://openalex.org/A5120435904", "display_name": "Inferix Team", "orcid": null, "observed_orcids": []}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Inferix Team", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": null, "display_name": "Feng, Tianyu", "orcid": null, "observed_orcids": []}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Feng, Tianyu", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5045569962", "display_name": "Yizeng Han", "orcid": "https://orcid.org/0000-0001-5706-8784", "observed_orcids": ["https://orcid.org/0000-0001-5706-8784"]}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Han, Yizeng", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5016341667", "display_name": "Jiahao He", "orcid": "https://orcid.org/0000-0003-3694-4989", "observed_orcids": ["https://orcid.org/0000-0003-3694-4989"]}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "He, Jiahao", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5086133186", "display_name": "Yuanyu He", "orcid": null, "observed_orcids": []}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "He, Yuanyu", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5081647301", "display_name": "Xi Yu Lin", "orcid": "https://orcid.org/0000-0003-2387-4117", "observed_orcids": ["https://orcid.org/0000-0003-2387-4117"]}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Lin, Xi", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5100416737", "display_name": "Teng Liu", "orcid": "https://orcid.org/0000-0002-4788-0426", "observed_orcids": ["https://orcid.org/0000-0002-4788-0426"]}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Liu, Teng", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5066000911", "display_name": "Hanfeng Lu", "orcid": "https://orcid.org/0000-0001-8934-8270", "observed_orcids": ["https://orcid.org/0000-0001-8934-8270"]}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Lu, Hanfeng", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5113292421", "display_name": "Jiasheng Tang", "orcid": "https://orcid.org/0009-0001-9154-8615", "observed_orcids": ["https://orcid.org/0009-0001-9154-8615"]}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Tang, Jiasheng", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5100392145", "display_name": "Wei Wang", "orcid": "https://orcid.org/0000-0002-9795-8367", "observed_orcids": ["https://orcid.org/0000-0002-9795-8367"]}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Wang, Wei", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5101804657", "display_name": "Zhiyuan Wang", "orcid": "https://orcid.org/0000-0002-5368-1132", "observed_orcids": ["https://orcid.org/0000-0002-5368-1132"]}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Wang, Zhiyuan", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": null, "display_name": "Wu, Jichao", "orcid": null, "observed_orcids": []}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Wu, Jichao", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5101785541", "display_name": "Mingyang Yang", "orcid": "https://orcid.org/0000-0002-1921-1217", "observed_orcids": ["https://orcid.org/0000-0002-1921-1217"]}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Yang, Mingyang", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5112032880", "display_name": "Yinghao Yu", "orcid": null, "observed_orcids": []}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Yu, Yinghao", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5100358735", "display_name": "Zeyu Zhang", "orcid": "https://orcid.org/0009-0006-2834-3007", "observed_orcids": ["https://orcid.org/0009-0006-2834-3007", "https://orcid.org/0009-0001-4488-9900", "https://orcid.org/0000-0002-3685-9373"]}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Zhang, Zeyu", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "last", "author": {"id": "https://openalex.org/A5076928390", "display_name": "Bohan Zhuang", "orcid": "https://orcid.org/0000-0002-0074-0303", "observed_orcids": ["https://orcid.org/0000-0002-0074-0303"]}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Zhuang, Bohan", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}], "abstract_inverted_index": {"World": [0], "models": [1, 29, 163], "serve": [2], "as": [3, 9, 123, 165], "core": [4], "simulators": [5], "for": [6, 43, 151, 205], "fields": [7], "such": [8], "agentic": [10], "AI,": [11, 13], "embodied": [12], "and": [14, 22, 38, 72, 94, 115, 158, 176, 181, 220], "gaming,": [15], "capable": [16], "of": [17, 70, 102, 197], "generating": [18, 76], "long,": [19], "physically": [20], "realistic,": [21], "interactive": [23, 173], "high-quality": [24, 116], "videos.": [25], "Moreover,": [26], "scaling": [27], "these": [28], "could": [30], "unlock": [31], "emergent": [32], "capabilities": [33], "in": [34, 79, 91], "visual": [35], "perception,": [36], "understanding,": [37], "reasoning,": [39], "paving": [40], "the": [41, 61, 68, 212], "way": [42], "a": [44, 124, 199], "new": [45, 200], "paradigm": [46], "that": [47], "moves": [48], "beyond": [49], "current": [50], "LLM-centric": [51], "vision": [52], "foundation": [53], "models.": [54], "A": [55], "key": [56], "breakthrough": [57], "empowering": [58], "them": [59], "is": [60, 120], "semi-autoregressive": [62, 135], "(block-diffusion)": [63], "decoding": [64, 136], "paradigm,": [65], "which": [66], "merges": [67], "strengths": [69], "diffusion": [71, 81, 105, 162], "autoregressive": [73], "methods": [74], "by": [75, 106], "video": [77, 96, 104, 161, 174, 207], "tokens": [78], "block-applying": [80], "within": [82], "each": [83], "block": [84], "while": [85], "conditioning": [86], "on": [87, 141], "previous": [88], "ones,": [89], "resulting": [90], "more": [92], "coherent": [93], "stable": [95], "sequences.": [97], "Crucially,": [98], "it": [99, 146, 190], "overcomes": [100], "limitations": [101], "standard": [103], "reintroducing": [107], "LLM-style": [108], "KV": [109], "Cache": [110], "management,": [111], "enabling": [112, 178], "efficient,": [113], "variable-length,": [114], "generation.": [117], "Therefore,": [118], "Inferix": [119, 167, 219], "specifically": [121], "designed": [122], "next-generation": [125], "inference": [126], "engine": [127], "to": [128, 184, 217], "enable": [129], "immersive": [130], "world": [131, 142, 187, 222], "synthesis": [132], "through": [133, 194], "optimized": [134], "processes.": [137], "This": [138], "dedicated": [139], "focus": [140], "simulation": [143, 183], "distinctly": [144], "sets": [145], "apart": [147], "from": [148, 159], "systems": [149], "engineered": [150], "high-concurrency": [152], "scenarios": [153], "(like": [154], "vLLM": [155], "or": [156], "SGLang)": [157], "classic": [160], "(such": [164], "xDiTs).": [166], "further": [168], "enhances": [169], "its": [170], "offering": [171], "with": [172], "streaming": [175], "profiling,": [177], "real-time": [179], "interaction": [180], "realistic": [182], "accurately": [185], "model": [186, 223], "dynamics.": [188], "Additionally,": [189], "supports": [191], "efficient": [192], "benchmarking": [193], "seamless": [195], "integration": [196], "LV-Bench,": [198], "fine-grained": [201], "evaluation": [202], "benchmark": [203], "tailored": [204], "minute-long": [206], "generation": [208], "scenarios.": [209], "We": [210], "hope": [211], "community": [213], "will": [214], "work": [215], "together": [216], "advance": [218], "foster": [221], "exploration.": [224]}, "primary_topic": {"id": "https://openalex.org/T10775", "display_name": "Generative Adversarial Networks and Image Synthesis", "score": 0.9937000274658203, "subfield": {"id": "https://openalex.org/subfields/1707", "display_name": "Computer Vision and Pattern Recognition"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, "topics": [{"id": "https://openalex.org/T10775", "display_name": "Generative Adversarial Networks and Image Synthesis", "score": 0.9937000274658203, "subfield": {"id": "https://openalex.org/subfields/1707", "display_name": "Computer Vision and Pattern Recognition"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, {"id": "https://openalex.org/T11714", "display_name": "Multimodal Machine Learning Applications", "score": 0.0052999998442828655, "subfield": {"id": "https://openalex.org/subfields/1707", "display_name": "Computer Vision and Pattern Recognition"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, {"id": "https://openalex.org/T10036", "display_name": "Advanced Neural Network Applications", "score": 9.999999747378752e-05, "subfield": {"id": "https://openalex.org/subfields/1707", "display_name": "Computer Vision and Pattern Recognition"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}], "primary_location": {"id": "pmh:oai:arXiv.org:2511.20714", "is_oa": true, "landing_page_url": "https://arxiv.org/abs/2511.20714", "pdf_url": "https://arxiv.org/pdf/2511.20714", "source": {"id": "https://openalex.org/S4306400194", "display_name": "arXiv (Cornell University)", "issn_l": "2331-8422", "issn": ["2331-8422"], "is_oa": true, "is_in_doaj": false, "is_core": false, "listed_in": [], "host_organization": "https://openalex.org/I205783295", "host_organization_name": "Cornell University", "host_organization_lineage": ["https://openalex.org/I205783295"], "host_organization_lineage_names": [], "type": "repository"}, "license": "cc-by", "license_id": "https://openalex.org/licenses/cc-by", "version": "submittedVersion", "is_accepted": false, "is_published": false, "raw_source_name": null, "raw_type": "text"}, "cited_by_count": 0, "is_retracted": false}, {"id": "https://openalex.org/W4414585490", "doi": "https://doi.org/10.48550/arxiv.2505.19155", "title": "Sparse-to-Dense: A Free Lunch for Lossless Acceleration of Video Understanding in LLMs", "publication_year": 2025, "publication_date": "2025-05-25", "type": "preprint", "language": "en", "authorships": [{"author_position": "first", "author": {"id": "https://openalex.org/A5103242466", "display_name": "Xuan Zhang", "orcid": "https://orcid.org/0000-0003-1152-9957", "observed_orcids": ["https://orcid.org/0000-0003-1152-9957"]}, "institutions": [{"id": "https://openalex.org/I79891267", "display_name": "Singapore Management University", "ror": "https://ror.org/050qmg959", "country_code": "SG", "type": "education", "lineage": ["https://openalex.org/I79891267"]}], "countries": ["SG"], "is_corresponding": false, "raw_author_name": "Zhang, Xuan", "raw_affiliation_strings": ["Singapore Management University ,"], "raw_orcid": null, "affiliations": [{"raw_affiliation_string": "Singapore Management University ,", "institution_ids": ["https://openalex.org/I79891267"]}]}, {"author_position": "middle", "author": {"id": null, "display_name": "Du, Cunxiao", "orcid": null, "observed_orcids": []}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Du, Cunxiao", "raw_affiliation_strings": ["Sea AI Lab ,"], "raw_orcid": null, "affiliations": [{"raw_affiliation_string": "Sea AI Lab ,", "institution_ids": []}]}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5104014314", "display_name": "Sicheng Yu", "orcid": "https://orcid.org/0009-0001-7935-9424", "observed_orcids": ["https://orcid.org/0009-0001-7935-9424"]}, "institutions": [{"id": "https://openalex.org/I79891267", "display_name": "Singapore Management University", "ror": "https://ror.org/050qmg959", "country_code": "SG", "type": "education", "lineage": ["https://openalex.org/I79891267"]}], "countries": ["SG"], "is_corresponding": false, "raw_author_name": "Yu, Sicheng", "raw_affiliation_strings": ["Singapore Management University ,"], "raw_orcid": null, "affiliations": [{"raw_affiliation_string": "Singapore Management University ,", "institution_ids": ["https://openalex.org/I79891267"]}]}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5104038166", "display_name": "Jiawei Wu", "orcid": null, "observed_orcids": []}, "institutions": [{"id": "https://openalex.org/I165932596", "display_name": "National University of Singapore", "ror": "https://ror.org/01tgyzw49", "country_code": "SG", "type": "education", "lineage": ["https://openalex.org/I165932596"]}], "countries": ["SG"], "is_corresponding": false, "raw_author_name": "Wu, Jiawei", "raw_affiliation_strings": ["National University of Singapore"], "raw_orcid": null, "affiliations": [{"raw_affiliation_string": "National University of Singapore", "institution_ids": ["https://openalex.org/I165932596"]}]}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5060676898", "display_name": "Fengzhuo Zhang", "orcid": null, "observed_orcids": []}, "institutions": [{"id": "https://openalex.org/I165932596", "display_name": "National University of Singapore", "ror": "https://ror.org/01tgyzw49", "country_code": "SG", "type": "education", "lineage": ["https://openalex.org/I165932596"]}], "countries": ["SG"], "is_corresponding": false, "raw_author_name": "Zhang, Fengzhuo", "raw_affiliation_strings": ["National University of Singapore"], "raw_orcid": null, "affiliations": [{"raw_affiliation_string": "National University of Singapore", "institution_ids": ["https://openalex.org/I165932596"]}]}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5050309466", "display_name": "Wei Gao", "orcid": "https://orcid.org/0000-0001-7429-5495", "observed_orcids": ["https://orcid.org/0000-0001-7429-5495"]}, "institutions": [{"id": "https://openalex.org/I79891267", "display_name": "Singapore Management University", "ror": "https://ror.org/050qmg959", "country_code": "SG", "type": "education", "lineage": ["https://openalex.org/I79891267"]}], "countries": ["SG"], "is_corresponding": false, "raw_author_name": "Gao, Wei", "raw_affiliation_strings": ["Singapore Management University ,"], "raw_orcid": null, "affiliations": [{"raw_affiliation_string": "Singapore Management University ,", "institution_ids": ["https://openalex.org/I79891267"]}]}, {"author_position": "last", "author": {"id": null, "display_name": "Liu, Qian", "orcid": null, "observed_orcids": []}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Liu, Qian", "raw_affiliation_strings": ["Sea AI Lab ,"], "raw_orcid": null, "affiliations": [{"raw_affiliation_string": "Sea AI Lab ,", "institution_ids": []}]}], "abstract_inverted_index": {"Due": [0], "to": [1, 50, 95, 126, 147], "the": [2, 12, 17, 25, 41, 86, 109], "auto-regressive": [3], "nature": [4], "of": [5, 28, 44], "current": [6], "video": [7, 29, 132], "large": [8], "language": [9], "models": [10], "(Video-LLMs),": [11], "inference": [13], "latency": [14], "increases": [15], "as": [16], "input": [18], "sequence": [19], "length": [20], "grows,": [21], "posing": [22], "challenges": [23], "for": [24], "efficient": [26], "processing": [27], "sequences": [30], "that": [31, 38, 75, 123], "are": [32], "usually": [33], "very": [34], "long.": [35], "We": [36], "observe": [37], "during": [39], "decoding,": [40], "attention": [42, 84], "scores": [43], "most": [45], "tokens": [46, 58], "in": [47, 115, 131], "Video-LLMs": [48, 97], "tend": [49], "be": [51], "sparse": [52, 82, 149], "and": [53, 85], "concentrated,": [54], "with": [55, 151], "only": [56], "certain": [57], "requiring": [59], "comprehensive": [60], "full": [61, 90], "attention.": [62, 91], "Based": [63], "on": [64], "this": [65], "insight,": [66], "we": [67], "introduce": [68], "Sparse-to-Dense": [69], "(StD),": [70], "a": [71, 119, 127, 140, 144, 148], "novel": [72], "decoding": [73], "strategy": [74], "integrates": [76], "two": [77], "distinct": [78], "modules:": [79], "one": [80], "leveraging": [81], "top-K": [83], "other": [87], "employing": [88], "dense": [89], "These": [92], "modules": [93], "collaborate": [94], "accelerate": [96], "without": [98], "loss.": [99], "The": [100], "fast": [101], "(sparse)": [102], "model": [103, 112, 136], "speculatively": [104], "decodes": [105], "multiple": [106], "tokens,": [107], "while": [108, 138], "slow": [110], "(dense)": [111], "verifies": [113], "them": [114], "parallel.": [116], "StD": [117], "is": [118], "tuning-free,": [120], "plug-and-play": [121], "solution": [122], "achieves": [124], "up": [125], "1.94$\\times$": [128], "walltime": [129], "speedup": [130], "processing.": [133], "It": [134], "maintains": [135], "performance": [137], "enabling": [139], "seamless": [141], "transition": [142], "from": [143], "standard": [145], "Video-LLM": [146, 150], "minimal": [152], "code": [153], "modifications.": [154]}, "primary_topic": {"id": "https://openalex.org/T11714", "display_name": "Multimodal Machine Learning Applications", "score": 0.8295999765396118, "subfield": {"id": "https://openalex.org/subfields/1707", "display_name": "Computer Vision and Pattern Recognition"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, "topics": [{"id": "https://openalex.org/T11714", "display_name": "Multimodal Machine Learning Applications", "score": 0.8295999765396118, "subfield": {"id": "https://openalex.org/subfields/1707", "display_name": "Computer Vision and Pattern Recognition"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, {"id": "https://openalex.org/T10036", "display_name": "Advanced Neural Network Applications", "score": 0.06260000169277191, "subfield": {"id": "https://openalex.org/subfields/1707", "display_name": "Computer Vision and Pattern Recognition"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, {"id": "https://openalex.org/T10181", "display_name": "Natural Language Processing Techniques", "score": 0.04320000112056732, "subfield": {"id": "https://openalex.org/subfields/1702", "display_name": "Artificial Intelligence"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}], "primary_location": {"id": "pmh:oai:arXiv.org:2505.19155", "is_oa": true, "landing_page_url": "https://arxiv.org/abs/2505.19155", "pdf_url": "https://arxiv.org/pdf/2505.19155", "source": {"id": "https://openalex.org/S4306400194", "display_name": "arXiv (Cornell University)", "issn_l": "2331-8422", "issn": ["2331-8422"], "is_oa": true, "is_in_doaj": false, "is_core": false, "listed_in": [], "host_organization": "https://openalex.org/I205783295", "host_organization_name": "Cornell University", "host_organization_lineage": ["https://openalex.org/I205783295"], "host_organization_lineage_names": [], "type": "repository"}, "license": "cc-by", "license_id": "https://openalex.org/licenses/cc-by", "version": "submittedVersion", "is_accepted": false, "is_published": false, "raw_source_name": null, "raw_type": "text"}, "cited_by_count": 0, "is_retracted": false}, {"id": "https://openalex.org/W7118051371", "doi": null, "title": "Yggdrasil: Bridging Dynamic Speculation and Static Runtime for Latency-Optimal Tree-Based LLM Decoding", "publication_year": 2025, "publication_date": "2025-12-29", "type": "preprint", "language": null, "authorships": [{"author_position": "first", "author": {"id": "https://openalex.org/A5091217469", "display_name": "Yue Guan", "orcid": "https://orcid.org/0009-0005-7433-2627", "observed_orcids": ["https://orcid.org/0009-0005-7433-2627"]}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Guan, Yue", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5121801880", "display_name": "Changming Yu", "orcid": null, "observed_orcids": []}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Yu, Changming", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5121757028", "display_name": "Shihan Fang", "orcid": null, "observed_orcids": []}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Fang, Shihan", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5121755920", "display_name": "Weiming Hu", "orcid": null, "observed_orcids": []}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Hu, Weiming", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5041947969", "display_name": "Zaifeng Pan", "orcid": "https://orcid.org/0000-0002-6759-2616", "observed_orcids": ["https://orcid.org/0000-0002-6759-2616"]}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Pan, Zaifeng", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5121749910", "display_name": "Zheng Wang", "orcid": null, "observed_orcids": []}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Wang, Zheng", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5121816858", "display_name": "Zihan Liu", "orcid": null, "observed_orcids": []}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Liu, Zihan", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5140632867", "display_name": "Yangjie Zhou", "orcid": null, "observed_orcids": []}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Zhou, Yangjie", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5121808620", "display_name": "Yufei Ding", "orcid": null, "observed_orcids": []}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Ding, Yufei", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5135688432", "display_name": "Minyi Guo", "orcid": null, "observed_orcids": []}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Guo, Minyi", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "last", "author": {"id": "https://openalex.org/A5003939279", "display_name": "Jingwen Leng", "orcid": "https://orcid.org/0000-0002-5660-5493", "observed_orcids": ["https://orcid.org/0000-0002-5660-5493"]}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Leng, Jingwen", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}], "abstract_inverted_index": {"Speculative": [0], "decoding": [1, 41], "improves": [2], "LLM": [3], "inference": [4], "by": [5], "generating": [6], "and": [7, 27, 46, 66, 76], "verifying": [8], "multiple": [9, 86], "tokens": [10], "in": [11], "parallel,": [12], "but": [13], "existing": [14], "systems": [15], "suffer": [16], "from": [17], "suboptimal": [18], "performance": [19], "due": [20], "to": [21, 69, 79], "a": [22, 34, 59], "mismatch": [23], "between": [24], "dynamic": [25], "speculation": [26], "static": [28, 56], "runtime": [29], "assumptions.": [30], "We": [31], "present": [32], "Yggdrasil,": [33], "co-designed": [35], "system": [36], "that": [37], "enables": [38], "latency-optimal": [39], "speculative": [40], "through": [42], "context-aware": [43], "tree": [44, 53], "drafting": [45], "compiler-friendly": [47], "execution.": [48], "Yggdrasil": [49, 72], "introduces": [50], "an": [51], "equal-growth": [52], "structure": [54], "for": [55, 63], "graph": [57], "compatibility,": [58], "latency-aware": [60], "optimization": [61], "objective": [62], "draft": [64], "selection,": [65], "stage-based": [67], "scheduling": [68], "reduce": [70], "overhead.": [71], "supports": [73], "unmodified": [74], "LLMs": [75], "achieves": [77], "up": [78], "$3.98\\times$": [80], "speedup": [81], "over": [82], "state-of-the-art": [83], "baselines": [84], "across": [85], "hardware": [87], "setups.": [88]}, "primary_topic": {"id": "https://openalex.org/T10054", "display_name": "Parallel Computing and Optimization Techniques", "score": 0.5371000170707703, "subfield": {"id": "https://openalex.org/subfields/1708", "display_name": "Hardware and Architecture"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, "topics": [{"id": "https://openalex.org/T10054", "display_name": "Parallel Computing and Optimization Techniques", "score": 0.5371000170707703, "subfield": {"id": "https://openalex.org/subfields/1708", "display_name": "Hardware and Architecture"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, {"id": "https://openalex.org/T10181", "display_name": "Natural Language Processing Techniques", "score": 0.32280001044273376, "subfield": {"id": "https://openalex.org/subfields/1702", "display_name": "Artificial Intelligence"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, {"id": "https://openalex.org/T10028", "display_name": "Topic Modeling", "score": 0.04740000143647194, "subfield": {"id": "https://openalex.org/subfields/1702", "display_name": "Artificial Intelligence"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}], "primary_location": {"id": "pmh:oai:arXiv.org:2512.23858", "is_oa": true, "landing_page_url": "http://arxiv.org/abs/2512.23858", "pdf_url": "https://arxiv.org/pdf/2512.23858", "source": {"id": "https://openalex.org/S4306400194", "display_name": "arXiv (Cornell University)", "issn_l": "2331-8422", "issn": ["2331-8422"], "is_oa": true, "is_in_doaj": false, "is_core": false, "listed_in": [], "host_organization": "https://openalex.org/I205783295", "host_organization_name": "Cornell University", "host_organization_lineage": ["https://openalex.org/I205783295"], "host_organization_lineage_names": [], "type": "repository"}, "license": null, "license_id": null, "version": "submittedVersion", "is_accepted": false, "is_published": false, "raw_source_name": null, "raw_type": "text"}, "cited_by_count": 0, "is_retracted": false}, {"id": "https://openalex.org/W4416118778", "doi": "https://doi.org/10.48550/arxiv.2504.05178", "title": "The 1st Solution for 4th PVUW MeViS Challenge: Unleashing the Potential of Large Multimodal Models for Referring Video Segmentation", "publication_year": 2025, "publication_date": "2025-04-07", "type": "preprint", "language": "en", "authorships": [{"author_position": "first", "author": {"id": "https://openalex.org/A5068398083", "display_name": "Hao Fang", "orcid": "https://orcid.org/0000-0002-8846-8294", "observed_orcids": ["https://orcid.org/0000-0002-8846-8294"]}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Fang, Hao", "raw_affiliation_strings": ["Shandong University Team : MVP-Lab"], "raw_orcid": null, "affiliations": [{"raw_affiliation_string": "Shandong University Team : MVP-Lab", "institution_ids": []}]}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5091558139", "display_name": "Runmin Cong", "orcid": "https://orcid.org/0000-0003-0972-4008", "observed_orcids": ["https://orcid.org/0000-0003-0972-4008"]}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Cong, Runmin", "raw_affiliation_strings": ["Shandong University Team : MVP-Lab"], "raw_orcid": null, "affiliations": [{"raw_affiliation_string": "Shandong University Team : MVP-Lab", "institution_ids": []}]}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5081796777", "display_name": "Xiankai Lu", "orcid": "https://orcid.org/0000-0002-9543-6960", "observed_orcids": ["https://orcid.org/0000-0002-9543-6960"]}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Lu, Xiankai", "raw_affiliation_strings": ["Shandong University Team : MVP-Lab"], "raw_orcid": null, "affiliations": [{"raw_affiliation_string": "Shandong University Team : MVP-Lab", "institution_ids": []}]}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5100324489", "display_name": "Zhiyang Chen", "orcid": "https://orcid.org/0000-0002-9341-7483", "observed_orcids": ["https://orcid.org/0000-0002-9341-7483"]}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Chen, Zhiyang", "raw_affiliation_strings": ["Shandong University Team : MVP-Lab"], "raw_orcid": null, "affiliations": [{"raw_affiliation_string": "Shandong University Team : MVP-Lab", "institution_ids": []}]}, {"author_position": "last", "author": {"id": "https://openalex.org/A5101561578", "display_name": "Zhang We", "orcid": "https://orcid.org/0000-0001-6478-3110", "observed_orcids": ["https://orcid.org/0000-0001-6478-3110", "https://orcid.org/0009-0001-6208-5359"]}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Zhang, Wei", "raw_affiliation_strings": ["Shandong University Team : MVP-Lab"], "raw_orcid": null, "affiliations": [{"raw_affiliation_string": "Shandong University Team : MVP-Lab", "institution_ids": []}]}], "abstract_inverted_index": {"Motion": [0], "expression": [1], "video": [2, 79, 107], "segmentation": [3], "is": [4, 89], "designed": [5], "to": [6, 18, 47, 52, 70, 113, 131], "segment": [7], "objects": [8], "in": [9, 49, 77, 154], "accordance": [10], "with": [11], "the": [12, 19, 73, 106, 110, 115, 119, 125, 133, 146, 155], "input": [13], "motion": [14, 30], "expressions.": [15], "In": [16, 58], "contrast": [17], "conventional": [20], "Referring": [21], "Video": [22], "Object": [23], "Segmentation": [24], "(RVOS),": [25], "it": [26, 37], "places": [27], "emphasis": [28], "on": [29, 145], "as": [31, 33, 85], "well": [32], "multi-object": [34], "expressions,": [35], "making": [36], "more": [38], "arduous.": [39], "Recently,": [40], "Large": [41], "Multimodal": [42], "Models": [43], "(LMMs)": [44], "have": [45], "begun": [46], "shine": [48], "RVOS": [50], "due": [51], "their": [53], "powerful": [54], "vision-language": [55], "perception": [56], "capabilities.": [57], "this": [59], "work,": [60], "we": [61, 82, 103, 123], "propose": [62], "a": [63, 90, 137], "simple": [64], "and": [65, 100, 150], "effective": [66], "inference": [67, 111], "optimization": [68], "method": [69], "fully": [71], "unleash": [72], "potential": [74], "of": [75, 97, 118, 127, 136], "LMMs": [76], "referring": [78], "segmentation.": [80], "Firstly,": [81], "use": [83], "Sa2VA": [84], "our": [86], "baseline,": [87], "which": [88], "unified": [91], "LMM": [92], "for": [93], "dense": [94], "grounded": [95], "understanding": [96, 117], "both": [98], "images": [99], "videos.": [101], "Secondly,": [102], "uniformly": [104], "sample": [105], "frames": [108], "during": [109], "process": [112], "enhance": [114], "model's": [116], "entire": [120], "video.": [121], "Finally,": [122], "integrate": [124], "results": [126], "multiple": [128], "expert": [129], "models": [130], "mitigate": [132], "erroneous": [134], "predictions": [135], "single": [138], "model.": [139], "Our": [140], "solution": [141], "achieved": [142], "61.98%": [143], "J&F": [144], "MeViS": [147, 159], "test": [148], "set": [149], "ranked": [151], "1st": [152], "place": [153], "4th": [156], "PVUW": [157], "Challenge": [158], "Track": [160], "at": [161], "CVPR": [162], "2025.": [163]}, "primary_topic": {"id": "https://openalex.org/T11714", "display_name": "Multimodal Machine Learning Applications", "score": 0.9922000169754028, "subfield": {"id": "https://openalex.org/subfields/1707", "display_name": "Computer Vision and Pattern Recognition"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, "topics": [{"id": "https://openalex.org/T11714", "display_name": "Multimodal Machine Learning Applications", "score": 0.9922000169754028, "subfield": {"id": "https://openalex.org/subfields/1707", "display_name": "Computer Vision and Pattern Recognition"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, {"id": "https://openalex.org/T10036", "display_name": "Advanced Neural Network Applications", "score": 0.004600000102072954, "subfield": {"id": "https://openalex.org/subfields/1707", "display_name": "Computer Vision and Pattern Recognition"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, {"id": "https://openalex.org/T11605", "display_name": "Visual Attention and Saliency Detection", "score": 0.0008999999845400453, "subfield": {"id": "https://openalex.org/subfields/1707", "display_name": "Computer Vision and Pattern Recognition"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}], "primary_location": {"id": "pmh:oai:arXiv.org:2504.05178", "is_oa": true, "landing_page_url": "http://arxiv.org/abs/2504.05178", "pdf_url": "https://arxiv.org/pdf/2504.05178", "source": {"id": "https://openalex.org/S4306400194", "display_name": "arXiv (Cornell University)", "issn_l": "2331-8422", "issn": ["2331-8422"], "is_oa": true, "is_in_doaj": false, "is_core": false, "listed_in": [], "host_organization": "https://openalex.org/I205783295", "host_organization_name": "Cornell University", "host_organization_lineage": ["https://openalex.org/I205783295"], "host_organization_lineage_names": [], "type": "repository"}, "license": null, "license_id": null, "version": "submittedVersion", "is_accepted": false, "is_published": false, "raw_source_name": null, "raw_type": "text"}, "cited_by_count": 0, "is_retracted": false}, {"id": "https://openalex.org/W4416524120", "doi": "https://doi.org/10.48550/arxiv.2506.23934", "title": "QPART: Adaptive Model Quantization and Dynamic Workload Balancing for Accuracy-aware Edge Inference", "publication_year": 2025, "publication_date": "2025-06-30", "type": "preprint", "language": "en", "authorships": [{"author_position": "first", "author": {"id": null, "display_name": "Li, Xiangchen", "orcid": null, "observed_orcids": []}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Li, Xiangchen", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5019689955", "display_name": "Saeid Ghafouri", "orcid": "https://orcid.org/0000-0003-3799-5702", "observed_orcids": ["https://orcid.org/0000-0003-3799-5702"]}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Ghafouri, Saeid", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5100562282", "display_name": "Ji Bo", "orcid": null, "observed_orcids": []}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Ji, Bo", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5086758775", "display_name": "Hans Vandierendonck", "orcid": "https://orcid.org/0000-0001-5868-9259", "observed_orcids": ["https://orcid.org/0000-0001-5868-9259"]}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Vandierendonck, Hans", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5029098934", "display_name": "Deepu John", "orcid": "https://orcid.org/0000-0002-6139-1100", "observed_orcids": ["https://orcid.org/0000-0002-6139-1100"]}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "John, Deepu", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "last", "author": {"id": "https://openalex.org/A5005410613", "display_name": "Dimitrios S. Nikolopoulos", "orcid": "https://orcid.org/0000-0003-0217-8307", "observed_orcids": ["https://orcid.org/0000-0003-0217-8307"]}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Nikolopoulos, Dimitrios S.", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}], "abstract_inverted_index": {"As": [0], "machine": [1], "learning": [2], "inferences": [3], "increasingly": [4], "move": [5], "to": [6, 10, 50, 65, 94, 152], "edge": [7, 36], "devices,": [8, 37], "adapting": [9], "diverse": [11, 35, 66], "computational": [12, 53, 114], "capabilities,": [13], "hardware,": [14], "and": [15, 57, 63, 75, 84, 102, 118, 140, 149, 156, 210, 220], "memory": [16], "constraints": [17, 59], "becomes": [18], "more": [19, 61], "critical.": [20], "Instead": [21], "of": [22, 164, 184, 198], "relying": [23], "on": [24], "a": [25, 46, 99, 128, 204], "pre-trained": [26], "model": [27, 48, 82, 101, 138], "fixed": [28], "for": [29, 132, 160], "all": [30], "future": [31], "inference": [32, 43, 77, 85, 95, 106, 134, 191], "queries": [33, 96], "across": [34], "we": [38, 71, 126], "argue": [39], "that": [40, 79], "planning": [41], "an": [42, 73, 167], "pattern": [44], "with": [45, 108, 213], "request-specific": [47], "tailored": [49], "the": [51, 90, 105, 109, 112, 133, 181, 190], "device's": [52, 113], "capacity,": [54, 117], "accuracy": [55, 119, 162, 168, 199, 221], "requirements,": [56], "time": [58, 154, 209], "is": [60], "cost-efficient": [62], "robust": [64], "scenarios.": [67], "To": [68, 175], "this": [69, 88, 178], "end,": [70], "propose": [72], "accuracy-aware": [74], "workload-balanced": [76], "system": [78], "integrates": [80], "joint": [81, 137], "quantization": [83, 139, 146, 186], "partitioning.": [86, 141], "In": [87], "approach,": [89], "server": [91], "dynamically": [92], "responds": [93], "by": [97, 194, 217], "sending": [98], "quantized": [100], "adaptively": [103], "sharing": [104], "workload": [107], "device.": [110], "Meanwhile,": [111], "power,": [115], "channel": [116], "requirements": [120, 163], "are": [121], "considered": [122], "when": [123], "deciding.": [124], "Furthermore,": [125], "introduce": [127], "new": [129], "optimization": [130, 173], "framework": [131], "system,": [135, 193], "incorporating": [136], "Our": [142], "approach": [143], "optimizes": [144], "layer-wise": [145, 187], "bit": [147], "width": [148], "partition": [150], "points": [151], "minimize": [153], "consumption": [155], "cost": [157], "while": [158], "accounting": [159], "varying": [161], "tasks": [165], "through": [166], "degradation": [169, 222], "metric": [170], "in": [171, 189, 207], "our": [172, 176], "model.": [174], "knowledge,": [177], "work": [179], "represents": [180], "first": [182], "exploration": [183], "optimizing": [185], "bit-width": [188], "serving": [192], "introducing": [195], "theoretical": [196], "measurement": [197], "degradation.": [200], "Simulation": [201], "results": [202], "demonstrate": [203], "substantial": [205], "reduction": [206], "overall": [208], "power": [211], "consumption,": [212], "computation": [214], "payloads": [215], "decreasing": [216], "over": [218], "80%": [219], "kept": [223], "below": [224], "1%.": [225]}, "primary_topic": {"id": "https://openalex.org/T10273", "display_name": "IoT and Edge/Fog Computing", "score": 0.7484999895095825, "subfield": {"id": "https://openalex.org/subfields/1705", "display_name": "Computer Networks and Communications"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, "topics": [{"id": "https://openalex.org/T10273", "display_name": "IoT and Edge/Fog Computing", "score": 0.7484999895095825, "subfield": {"id": "https://openalex.org/subfields/1705", "display_name": "Computer Networks and Communications"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, {"id": "https://openalex.org/T10036", "display_name": "Advanced Neural Network Applications", "score": 0.23409999907016754, "subfield": {"id": "https://openalex.org/subfields/1707", "display_name": "Computer Vision and Pattern Recognition"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, {"id": "https://openalex.org/T14098", "display_name": "Energy Efficiency in Computing", "score": 0.002199999988079071, "subfield": {"id": "https://openalex.org/subfields/1708", "display_name": "Hardware and Architecture"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}], "primary_location": {"id": "pmh:oai:arXiv.org:2506.23934", "is_oa": true, "landing_page_url": "http://arxiv.org/abs/2506.23934", "pdf_url": "https://arxiv.org/pdf/2506.23934", "source": {"id": "https://openalex.org/S4306400194", "display_name": "arXiv (Cornell University)", "issn_l": "2331-8422", "issn": ["2331-8422"], "is_oa": true, "is_in_doaj": false, "is_core": false, "listed_in": [], "host_organization": "https://openalex.org/I205783295", "host_organization_name": "Cornell University", "host_organization_lineage": ["https://openalex.org/I205783295"], "host_organization_lineage_names": [], "type": "repository"}, "license": null, "license_id": null, "version": "submittedVersion", "is_accepted": false, "is_published": false, "raw_source_name": null, "raw_type": "text"}, "cited_by_count": 1, "is_retracted": false}, {"id": "https://openalex.org/W4416373982", "doi": "https://doi.org/10.48550/arxiv.2510.04008", "title": "RACE Attention: A Strictly Linear-Time Attention Layer for Training on Outrageously Large Contexts", "publication_year": 2025, "publication_date": "2025-10-05", "type": "preprint", "language": "en", "authorships": [{"author_position": "first", "author": {"id": "https://openalex.org/A5076233117", "display_name": "Sahil Joshi", "orcid": "https://orcid.org/0000-0003-3505-9363", "observed_orcids": ["https://orcid.org/0000-0003-3505-9363"]}, "institutions": [{"id": "https://openalex.org/I74775410", "display_name": "Rice University", "ror": "https://ror.org/008zs3103", "country_code": "US", "type": "education", "lineage": ["https://openalex.org/I74775410"]}], "countries": ["US"], "is_corresponding": false, "raw_author_name": "Joshi, Sahil", "raw_affiliation_strings": ["Department of Computer Science , Rice University , TX , USA"], "raw_orcid": null, "affiliations": [{"raw_affiliation_string": "Department of Computer Science , Rice University , TX , USA", "institution_ids": ["https://openalex.org/I74775410"]}]}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5020686363", "display_name": "Agniva Chowdhury", "orcid": null, "observed_orcids": []}, "institutions": [{"id": "https://openalex.org/I74775410", "display_name": "Rice University", "ror": "https://ror.org/008zs3103", "country_code": "US", "type": "education", "lineage": ["https://openalex.org/I74775410"]}], "countries": ["US"], "is_corresponding": false, "raw_author_name": "Chowdhury, Agniva", "raw_affiliation_strings": ["Department of Computer Science , Rice University , TX , USA"], "raw_orcid": null, "affiliations": [{"raw_affiliation_string": "Department of Computer Science , Rice University , TX , USA", "institution_ids": ["https://openalex.org/I74775410"]}]}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5120555334", "display_name": "Amar Kanakamedala", "orcid": null, "observed_orcids": []}, "institutions": [{"id": "https://openalex.org/I74775410", "display_name": "Rice University", "ror": "https://ror.org/008zs3103", "country_code": "US", "type": "education", "lineage": ["https://openalex.org/I74775410"]}], "countries": ["US"], "is_corresponding": false, "raw_author_name": "Kanakamedala, Amar", "raw_affiliation_strings": ["Department of Computer Science , Rice University , TX , USA"], "raw_orcid": null, "affiliations": [{"raw_affiliation_string": "Department of Computer Science , Rice University , TX , USA", "institution_ids": ["https://openalex.org/I74775410"]}]}, {"author_position": "middle", "author": {"id": null, "display_name": "Singh, Ekam", "orcid": null, "observed_orcids": []}, "institutions": [{"id": "https://openalex.org/I74775410", "display_name": "Rice University", "ror": "https://ror.org/008zs3103", "country_code": "US", "type": "education", "lineage": ["https://openalex.org/I74775410"]}], "countries": ["US"], "is_corresponding": false, "raw_author_name": "Singh, Ekam", "raw_affiliation_strings": ["Department of Computer Science , Rice University , TX , USA"], "raw_orcid": null, "affiliations": [{"raw_affiliation_string": "Department of Computer Science , Rice University , TX , USA", "institution_ids": ["https://openalex.org/I74775410"]}]}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5078135146", "display_name": "E. Y. Tu", "orcid": null, "observed_orcids": []}, "institutions": [{"id": "https://openalex.org/I74775410", "display_name": "Rice University", "ror": "https://ror.org/008zs3103", "country_code": "US", "type": "education", "lineage": ["https://openalex.org/I74775410"]}], "countries": ["US"], "is_corresponding": false, "raw_author_name": "Tu, Evan", "raw_affiliation_strings": ["Department of Computer Science , Rice University , TX , USA"], "raw_orcid": null, "affiliations": [{"raw_affiliation_string": "Department of Computer Science , Rice University , TX , USA", "institution_ids": ["https://openalex.org/I74775410"]}]}, {"author_position": "last", "author": {"id": "https://openalex.org/A5024993683", "display_name": "Anshumali Shrivastava", "orcid": "https://orcid.org/0000-0002-5042-2856", "observed_orcids": ["https://orcid.org/0000-0002-5042-2856"]}, "institutions": [{"id": "https://openalex.org/I74775410", "display_name": "Rice University", "ror": "https://ror.org/008zs3103", "country_code": "US", "type": "education", "lineage": ["https://openalex.org/I74775410"]}], "countries": ["US"], "is_corresponding": false, "raw_author_name": "Shrivastava, Anshumali", "raw_affiliation_strings": ["Department of Computer Science , Rice University , TX , USA"], "raw_orcid": null, "affiliations": [{"raw_affiliation_string": "Department of Computer Science , Rice University , TX , USA", "institution_ids": ["https://openalex.org/I74775410"]}]}], "abstract_inverted_index": {"Softmax": [0, 31, 68], "Attention": [1, 69, 81, 121, 194], "has": [2], "a": [3, 35, 40, 64, 87, 143, 148, 178, 197], "quadratic": [4], "time": [5, 135], "complexity": [6], "in": [7, 74, 177], "sequence": [8, 75], "length,": [9], "which": [10, 182], "becomes": [11], "prohibitive": [12], "to": [13, 67, 128, 157], "run": [14], "at": [15, 213], "long": [16], "contexts,": [17], "even": [18], "with": [19, 86], "highly": [20], "optimized": [21], "GPU": [22, 165], "kernels.": [23], "For": [24], "example,": [25], "FlashAttention-2/3": [26], "(exact,": [27], "GPU-optimized": [28], "implementations": [29], "of": [30, 39, 106, 155, 188], "Attention)": [32], "cannot": [33], "complete": [34], "single": [36, 41, 149, 179], "forward-backward": [37, 180], "pass": [38], "attention": [42, 93, 109, 150, 191], "layer": [43, 151], "once": [44], "the": [45, 83, 107, 186], "context": [46], "exceeds": [47], "~4": [48], "million": [49, 159, 168], "tokens": [50, 160, 169], "on": [51, 147, 161, 170, 206], "an": [52, 162, 171], "NVIDIA": [53, 163], "GH200": [54, 164], "(96": [55], "GB).": [56], "We": [57, 209], "introduce": [58], "Repeated": [59], "Arrays-of-Count": [60], "Estimators": [61], "(RACE)": [62], "Attention,": [63], "kernel-inspired": [65], "alternative": [66], "that": [70], "is": [71, 183], "strictly": [72], "linear": [73], "length": [76, 131], "and": [77, 91, 99, 117, 136, 152, 166, 199], "embedding": [78], "size.": [79], "RACE": [80, 120, 193], "replaces": [82], "exponential": [84], "kernel": [85], "sharpened": [88], "angular": [89], "similarity,": [90], "approximates": [92], "outputs": [94], "via": [95], "Gaussian": [96], "random": [97], "projections": [98], "soft": [100], "Locality-Sensitive": [101], "Hashing": [102], "(LSH),": [103], "avoiding": [104], "construction": [105], "full": [108], "matrix.": [110], "Across": [111], "language": [112, 115], "modeling,": [113, 116], "masked": [114], "text/image": [118], "classification,": [119], "matches": [122], "or": [123], "outperforms": [124], "strong": [125], "baselines": [126], "up": [127, 156], "64K": [129], "seqeuence": [130], "while": [132], "reducing": [133], "wall-clock": [134], "memory": [137], "usage.": [138], "In": [139], "addition,": [140], "we": [141], "conduct": [142], "controlled": [144], "scaling": [145], "study": [146], "demonstrate": [153], "processing": [154], "12": [158], "75": [167], "Intel": [172], "Xeon": [173], "Gold": [174], "5220R": [175], "CPU": [176], "pass,": [181], "well": [184], "beyond": [185], "capabilities": [187], "current": [189], "state-of-the-art": [190], "implementations.": [192], "thus": [195], "offers": [196], "practical": [198], "theoretically": [200], "grounded": [201], "mechanism": [202], "for": [203], "long-context": [204], "training": [205], "today's": [207], "hardware.": [208], "release": [210], "our": [211], "code": [212], "https://github.com/sahiljoshi515/RACE_Attention.": [214]}, "primary_topic": {"id": "https://openalex.org/T10028", "display_name": "Topic Modeling", "score": 0.8730000257492065, "subfield": {"id": "https://openalex.org/subfields/1702", "display_name": "Artificial Intelligence"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, "topics": [{"id": "https://openalex.org/T10028", "display_name": "Topic Modeling", "score": 0.8730000257492065, "subfield": {"id": "https://openalex.org/subfields/1702", "display_name": "Artificial Intelligence"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, {"id": "https://openalex.org/T10181", "display_name": "Natural Language Processing Techniques", "score": 0.044199999421834946, "subfield": {"id": "https://openalex.org/subfields/1702", "display_name": "Artificial Intelligence"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, {"id": "https://openalex.org/T10036", "display_name": "Advanced Neural Network Applications", "score": 0.023399999365210533, "subfield": {"id": "https://openalex.org/subfields/1707", "display_name": "Computer Vision and Pattern Recognition"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}], "primary_location": {"id": "pmh:oai:arXiv.org:2510.04008", "is_oa": true, "landing_page_url": "https://arxiv.org/abs/2510.04008", "pdf_url": "https://arxiv.org/pdf/2510.04008", "source": {"id": "https://openalex.org/S4306400194", "display_name": "arXiv (Cornell University)", "issn_l": "2331-8422", "issn": ["2331-8422"], "is_oa": true, "is_in_doaj": false, "is_core": false, "listed_in": [], "host_organization": "https://openalex.org/I205783295", "host_organization_name": "Cornell University", "host_organization_lineage": ["https://openalex.org/I205783295"], "host_organization_lineage_names": [], "type": "repository"}, "license": "cc-by", "license_id": "https://openalex.org/licenses/cc-by", "version": "submittedVersion", "is_accepted": false, "is_published": false, "raw_source_name": null, "raw_type": "text"}, "cited_by_count": 0, "is_retracted": false}, {"id": "https://openalex.org/W4416429258", "doi": "https://doi.org/10.1109/jiot.2025.3635092", "title": "End-to-End Autonomous Driving: From Classic Paradigm to Large Model Empowerment\u2014A Comprehensive Survey", "publication_year": 2025, "publication_date": "2025-11-20", "type": "article", "language": null, "authorships": [{"author_position": "first", "author": {"id": "https://openalex.org/A5074340955", "display_name": "Wei Dong", "orcid": "https://orcid.org/0000-0003-0498-1494", "observed_orcids": ["https://orcid.org/0000-0003-0498-1494"]}, "institutions": [{"id": "https://openalex.org/I115592961", "display_name": "Jiangsu University", "ror": "https://ror.org/03jc41j30", "country_code": "CN", "type": "education", "lineage": ["https://openalex.org/I115592961"]}], "countries": ["CN"], "is_corresponding": false, "raw_author_name": "Wei Dong", "raw_affiliation_strings": ["Automotive Engineering Research Institute, Jiangsu University, Zhenjiang, China"], "raw_orcid": null, "affiliations": [{"raw_affiliation_string": "Automotive Engineering Research Institute, Jiangsu University, Zhenjiang, China", "institution_ids": ["https://openalex.org/I115592961"]}]}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5051137308", "display_name": "Sikai Lu", "orcid": "https://orcid.org/0000-0002-5990-8633", "observed_orcids": ["https://orcid.org/0000-0002-5990-8633"]}, "institutions": [{"id": "https://openalex.org/I115592961", "display_name": "Jiangsu University", "ror": "https://ror.org/03jc41j30", "country_code": "CN", "type": "education", "lineage": ["https://openalex.org/I115592961"]}], "countries": ["CN"], "is_corresponding": false, "raw_author_name": "Sikai Lu", "raw_affiliation_strings": ["Automotive Engineering Research Institute, Jiangsu University, Zhenjiang, China"], "raw_orcid": "https://orcid.org/0000-0002-5990-8633", "affiliations": [{"raw_affiliation_string": "Automotive Engineering Research Institute, Jiangsu University, Zhenjiang, China", "institution_ids": ["https://openalex.org/I115592961"]}]}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5044931100", "display_name": "Xinhe Chen", "orcid": "https://orcid.org/0009-0003-3176-7335", "observed_orcids": ["https://orcid.org/0009-0003-3176-7335"]}, "institutions": [{"id": "https://openalex.org/I115592961", "display_name": "Jiangsu University", "ror": "https://ror.org/03jc41j30", "country_code": "CN", "type": "education", "lineage": ["https://openalex.org/I115592961"]}], "countries": ["CN"], "is_corresponding": false, "raw_author_name": "Xinhe Chen", "raw_affiliation_strings": ["Automotive Engineering Research Institute, Jiangsu University, Zhenjiang, China"], "raw_orcid": "https://orcid.org/0009-0003-3176-7335", "affiliations": [{"raw_affiliation_string": "Automotive Engineering Research Institute, Jiangsu University, Zhenjiang, China", "institution_ids": ["https://openalex.org/I115592961"]}]}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5009046652", "display_name": "Shunyao Zhang", "orcid": "https://orcid.org/0000-0002-7928-0441", "observed_orcids": ["https://orcid.org/0000-0002-7928-0441"]}, "institutions": [{"id": "https://openalex.org/I115592961", "display_name": "Jiangsu University", "ror": "https://ror.org/03jc41j30", "country_code": "CN", "type": "education", "lineage": ["https://openalex.org/I115592961"]}], "countries": ["CN"], "is_corresponding": false, "raw_author_name": "Shunyao Zhang", "raw_affiliation_strings": ["Automotive Engineering Research Institute, Jiangsu University, Zhenjiang, China"], "raw_orcid": "https://orcid.org/0000-0002-7928-0441", "affiliations": [{"raw_affiliation_string": "Automotive Engineering Research Institute, Jiangsu University, Zhenjiang, China", "institution_ids": ["https://openalex.org/I115592961"]}]}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5101634250", "display_name": "Qingchao Liu", "orcid": "https://orcid.org/0000-0001-6486-0999", "observed_orcids": ["https://orcid.org/0000-0001-6486-0999"]}, "institutions": [{"id": "https://openalex.org/I115592961", "display_name": "Jiangsu University", "ror": "https://ror.org/03jc41j30", "country_code": "CN", "type": "education", "lineage": ["https://openalex.org/I115592961"]}], "countries": ["CN"], "is_corresponding": false, "raw_author_name": "Qingchao Liu", "raw_affiliation_strings": ["Automotive Engineering Research Institute, Jiangsu University, Zhenjiang, China"], "raw_orcid": "https://orcid.org/0000-0001-6486-0999", "affiliations": [{"raw_affiliation_string": "Automotive Engineering Research Institute, Jiangsu University, Zhenjiang, China", "institution_ids": ["https://openalex.org/I115592961"]}]}, {"author_position": "middle", "author": {"id": null, "display_name": "Ze Liu", "orcid": "https://orcid.org/0009-0008-3832-8016", "observed_orcids": ["https://orcid.org/0009-0008-3832-8016"]}, "institutions": [{"id": "https://openalex.org/I115592961", "display_name": "Jiangsu University", "ror": "https://ror.org/03jc41j30", "country_code": "CN", "type": "education", "lineage": ["https://openalex.org/I115592961"]}], "countries": ["CN"], "is_corresponding": false, "raw_author_name": "Ze Liu", "raw_affiliation_strings": ["Automotive Engineering Research Institute, Jiangsu University, Zhenjiang, China"], "raw_orcid": "https://orcid.org/0009-0008-3832-8016", "affiliations": [{"raw_affiliation_string": "Automotive Engineering Research Institute, Jiangsu University, Zhenjiang, China", "institution_ids": ["https://openalex.org/I115592961"]}]}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5100630554", "display_name": "Long Chen", "orcid": "https://orcid.org/0000-0002-2079-3867", "observed_orcids": ["https://orcid.org/0000-0002-2079-3867", "https://orcid.org/0000-0001-7473-6852"]}, "institutions": [{"id": "https://openalex.org/I115592961", "display_name": "Jiangsu University", "ror": "https://ror.org/03jc41j30", "country_code": "CN", "type": "education", "lineage": ["https://openalex.org/I115592961"]}], "countries": ["CN"], "is_corresponding": false, "raw_author_name": "Long Chen", "raw_affiliation_strings": ["Automotive Engineering Research Institute, Jiangsu University, Zhenjiang, China"], "raw_orcid": "https://orcid.org/0000-0002-2079-3867", "affiliations": [{"raw_affiliation_string": "Automotive Engineering Research Institute, Jiangsu University, Zhenjiang, China", "institution_ids": ["https://openalex.org/I115592961"]}]}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5100771674", "display_name": "Hai Wang", "orcid": "https://orcid.org/0000-0002-9136-8091", "observed_orcids": ["https://orcid.org/0000-0002-9136-8091"]}, "institutions": [{"id": "https://openalex.org/I115592961", "display_name": "Jiangsu University", "ror": "https://ror.org/03jc41j30", "country_code": "CN", "type": "education", "lineage": ["https://openalex.org/I115592961"]}], "countries": ["CN"], "is_corresponding": false, "raw_author_name": "Hai Wang", "raw_affiliation_strings": ["School of Automotive and Traffic Engineering, Jiangsu University, Zhenjiang, China", "School of Automotive and Traffic Engineering of Jiangsu University, Zhenjiang, China"], "raw_orcid": "https://orcid.org/0000-0002-9136-8091", "affiliations": [{"raw_affiliation_string": "School of Automotive and Traffic Engineering, Jiangsu University, Zhenjiang, China", "institution_ids": ["https://openalex.org/I115592961"]}, {"raw_affiliation_string": "School of Automotive and Traffic Engineering of Jiangsu University, Zhenjiang, China", "institution_ids": ["https://openalex.org/I115592961"]}]}, {"author_position": "last", "author": {"id": "https://openalex.org/A5057276700", "display_name": "Yingfeng Cai", "orcid": "https://orcid.org/0000-0002-0633-9887", "observed_orcids": ["https://orcid.org/0000-0002-0633-9887", "https://orcid.org/0009-0007-6534-9659", "https://orcid.org/0009-0005-9965-6142"]}, "institutions": [{"id": "https://openalex.org/I115592961", "display_name": "Jiangsu University", "ror": "https://ror.org/03jc41j30", "country_code": "CN", "type": "education", "lineage": ["https://openalex.org/I115592961"]}], "countries": ["CN"], "is_corresponding": false, "raw_author_name": "Yingfeng Cai", "raw_affiliation_strings": ["Automotive Engineering Research Institute, Jiangsu University, Zhenjiang, China"], "raw_orcid": "https://orcid.org/0000-0002-0633-9887", "affiliations": [{"raw_affiliation_string": "Automotive Engineering Research Institute, Jiangsu University, Zhenjiang, China", "institution_ids": ["https://openalex.org/I115592961"]}]}], "abstract_inverted_index": {"In": [0, 118], "recent": [1, 108], "years,": [2], "autonomous": [3, 29, 67, 203], "driving": [4, 30], "technology": [5], "has": [6, 31], "witnessed": [7], "rapid": [8], "global": [9], "development,": [10], "offering": [11], "effective": [12], "solutions": [13], "to": [14, 19, 37, 42, 190], "increasingly": [15], "severe": [16], "challenges": [17, 122], "related": [18], "traffic": [20], "congestion": [21], "and": [22, 49, 62, 73, 82, 101, 105, 115, 131, 140, 159, 174, 177, 183, 196, 213], "road": [23], "safety.": [24], "Among": [25], "various": [26], "paradigms,": [27], "end-to-end": [28, 66, 202], "emerged": [32], "as": [33, 96, 124, 192], "a": [34, 56, 193], "promising": [35], "alternative": [36], "traditional": [38], "modular": [39, 169], "systems,": [40], "owing": [41], "its": [43, 206], "streamlined": [44], "architecture,": [45], "enhanced": [46, 214], "decision": [47], "consistency,": [48], "superior": [50], "generalization": [51], "capabilities.": [52], "This": [53, 187], "survey": [54, 135, 188], "provides": [55], "comprehensive": [57, 194], "review": [58], "of": [59, 65, 75, 120, 151, 164, 180], "the": [60, 71, 88, 149, 162, 178], "evolution": [61, 207], "core": [63], "technologies": [64], "driving.": [68], "It": [69], "emphasizes": [70], "applications": [72], "developments": [74], "imitation": [76], "learning": [77, 80], "(IL),": [78], "reinforcement": [79], "(RL),": [81], "other": [83], "paradigms.": [84], "Furthermore,": [85], "it": [86, 144], "highlights": [87], "emerging": [89], "paradigm": [90], "empowered": [91], "by": [92], "foundation": [93, 165], "models,": [94], "such": [95, 123], "large": [97], "language": [98], "models": [99, 103, 153], "(LLMs)": [100], "vision-language": [102], "(VLMs),": [104], "systematically": [106], "categorizes": [107], "advances": [109], "in": [110, 201], "planning,": [111], "reasoning,": [112], "data": [113, 157], "generation,": [114], "scene": [116], "understanding.": [117], "light": [119], "persistent": [121], "multi-modal": [125, 185], "fusion": [126], "complexity,": [127], "low": [128], "sample": [129], "efficiency,": [130], "safety": [132, 215], "risks,": [133], "this": [134], "summarizes": [136], "representative": [137], "research": [138], "efforts": [139], "corresponding": [141], "solutions.": [142], "Finally,": [143], "outlines": [145], "future": [146], "directions,": [147], "including": [148], "integration": [150], "world": [152], "(WMs)": [154], "for": [155, 199], "unified": [156, 184], "generation": [158], "inference": [160], "optimization,": [161], "advancement": [163], "model": [166], "architectures": [167], "toward": [168, 208], "design,": [170], "sparse": [171], "activation": [172], "mechanisms,": [173], "knowledge": [175], "distillation,": [176], "realization": [179], "deployable,": [181], "transferable,": [182], "frameworks.": [186], "aims": [189], "serve": [191], "theoretical": [195], "technical": [197], "reference": [198], "researchers": [200], "driving,": [204], "facilitating": [205], "higher": [209], "performance,": [210], "stronger": [211], "generalization,": [212], "assurance.": [216]}, "primary_topic": {"id": "https://openalex.org/T11099", "display_name": "Autonomous Vehicle Technology and Safety", "score": 0.9961000084877014, "subfield": {"id": "https://openalex.org/subfields/2203", "display_name": "Automotive Engineering"}, "field": {"id": "https://openalex.org/fields/22", "display_name": "Engineering"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, "topics": [{"id": "https://openalex.org/T11099", "display_name": "Autonomous Vehicle Technology and Safety", "score": 0.9961000084877014, "subfield": {"id": "https://openalex.org/subfields/2203", "display_name": "Automotive Engineering"}, "field": {"id": "https://openalex.org/fields/22", "display_name": "Engineering"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, {"id": "https://openalex.org/T10462", "display_name": "Reinforcement Learning in Robotics", "score": 0.0020000000949949026, "subfield": {"id": "https://openalex.org/subfields/1702", "display_name": "Artificial Intelligence"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, {"id": "https://openalex.org/T10036", "display_name": "Advanced Neural Network Applications", "score": 0.00039999998989515007, "subfield": {"id": "https://openalex.org/subfields/1707", "display_name": "Computer Vision and Pattern Recognition"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}], "primary_location": {"id": "doi:10.1109/jiot.2025.3635092", "is_oa": false, "landing_page_url": "https://doi.org/10.1109/jiot.2025.3635092", "pdf_url": null, "source": {"id": "https://openalex.org/S2480266640", "display_name": "IEEE Internet of Things Journal", "issn_l": "2327-4662", "issn": ["2327-4662", "2372-2541"], "is_oa": false, "is_in_doaj": false, "is_core": true, "listed_in": ["cwts-core", "jufo-2", "norway-2"], "host_organization": "https://openalex.org/P4310319808", "host_organization_name": "Institute of Electrical and Electronics Engineers", "host_organization_lineage": ["https://openalex.org/P4310319808"], "host_organization_lineage_names": ["Institute of Electrical and Electronics Engineers"], "type": "journal"}, "license": null, "license_id": null, "version": "publishedVersion", "is_accepted": true, "is_published": true, "raw_source_name": "IEEE Internet of Things Journal", "raw_type": "journal-article"}, "cited_by_count": 7, "is_retracted": false}, {"id": "https://openalex.org/W7203549776", "doi": "https://doi.org/10.34218/ijcet_16_04_017", "title": "THE COLD START TAX: A BREAK EVEN LAW FOR SCALE TO ZERO GPU INFERENCE SERVING", "publication_year": 2025, "publication_date": "2025-07-29", "type": "article", "language": null, "authorships": [{"author_position": "first", "author": {"id": "https://openalex.org/A5147238168", "display_name": "PavanKumar Koneru", "orcid": null, "observed_orcids": []}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "PavanKumar Koneru", "raw_affiliation_strings": ["USA"], "raw_orcid": null, "affiliations": [{"raw_affiliation_string": "USA", "institution_ids": []}]}, {"author_position": "last", "author": {"id": "https://openalex.org/A5115898826", "display_name": "Sai Kumar Pola", "orcid": null, "observed_orcids": []}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Sai Kumar Pola", "raw_affiliation_strings": ["USA"], "raw_orcid": null, "affiliations": [{"raw_affiliation_string": "USA", "institution_ids": []}]}], "abstract_inverted_index": {"Graphics": [0], "processing": [1], "units": [2], "dominate": [3], "the": [4, 37, 56, 61, 88, 94, 106, 116, 123, 129, 132, 140, 146, 157, 163], "cost": [5, 117], "of": [6, 112, 131, 159, 165], "large": [7], "language": [8, 43], "model": [9, 44], "services,": [10], "and": [11, 46, 85, 145], "production": [12], "fleets": [13], "are": [14], "consistently": [15], "reported": [16], "to": [17, 25, 66, 68, 168], "be": [18], "under": [19, 100, 137], "utilised.The": [20], "reflexive": [21], "remedy": [22], "is": [23, 39, 91, 121, 142, 178], "scale": [24, 65], "zero:": [26], "an": [27, 54], "idle": [28, 97, 124], "accelerator": [29], "costs": [30], "money,": [31], "so": [32, 136], "release": [33], "it.This": [34], "paper": [35], "shows": [36], "reasoning": [38], "structurally": [40], "unsound": [41], "for": [42, 180], "inference,": [45], "gives": [47], "a": [48, 80, 143, 155, 169, 174], "closed": [49], "form": [50], "reason": [51], "rather": [52], "than": [53, 93, 105], "anecdote.Treating": [55], "keep": [57, 119], "alive": [58, 120], "window": [59], "as": [60], "single": [62], "parameter": [63, 147], "connecting": [64], "zero": [67], "always": [69], "on": [70], "provisioning,": [71], "we": [72], "derive": [73], "three": [74], "results.A": [75], "break": [76], "even": [77], "law:": [78], "releasing": [79], "replica": [81], "reduces": [82], "spend": [83, 160], "if": [84, 87], "only": [86], "cold": [89, 133, 170], "start": [90, 134], "shorter": [92, 104], "mean": [95, 107], "residual": [96], "gap,": [98], "which": [99], "Poisson": [101, 138, 181], "arrivals": [102, 139], "means": [103], "inter": [108], "arrival": [109], "time,": [110], "independently": [111], "tuning.An": [113], "optimality": [114], "condition:": [115], "optimal": [118], "where": [122], "gap": [125], "hazard": [126], "rate": [127], "equals": [128, 162], "reciprocal": [130], "duration,": [135], "optimum": [141], "corner": [144], "most": [148], "operators": [149], "tune": [150], "admits": [151], "no": [152], "interior": [153], "solution.And": [154], "duality:": [156], "fraction": [158, 164], "recoverable": [161], "requests": [166], "exposed": [167], "start,": [171], "scaled": [172], "by": [173], "burstiness": [175], "dividend": [176], "that": [177], "unity": [179], "arrivals.": [182]}, "primary_topic": {"id": "https://openalex.org/T10101", "display_name": "Cloud Computing and Resource Management", "score": 0.9843999743461609, "subfield": {"id": "https://openalex.org/subfields/1710", "display_name": "Information Systems"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, "topics": [{"id": "https://openalex.org/T10101", "display_name": "Cloud Computing and Resource Management", "score": 0.9843999743461609, "subfield": {"id": "https://openalex.org/subfields/1710", "display_name": "Information Systems"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, {"id": "https://openalex.org/T14098", "display_name": "Energy Efficiency in Computing", "score": 0.00839999970048666, "subfield": {"id": "https://openalex.org/subfields/1708", "display_name": "Hardware and Architecture"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, {"id": "https://openalex.org/T12127", "display_name": "Software System Performance and Reliability", "score": 0.0031999999191612005, "subfield": {"id": "https://openalex.org/subfields/1705", "display_name": "Computer Networks and Communications"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}], "primary_location": {"id": "doi:10.34218/ijcet_16_04_017", "is_oa": true, "landing_page_url": "https://doi.org/10.34218/ijcet_16_04_017", "pdf_url": "https://doi.org/10.34218/ijcet_16_04_017", "source": {"id": "https://openalex.org/S4210225229", "display_name": "INTERNATIONAL JOURNAL OF COMPUTER ENGINEERING & TECHNOLOGY", "issn_l": "0976-6367", "issn": ["0976-6367", "0976-6375"], "is_oa": true, "is_in_doaj": false, "is_core": false, "listed_in": [], "host_organization": null, "host_organization_name": null, "host_organization_lineage": [], "host_organization_lineage_names": [], "type": "journal"}, "license": null, "license_id": null, "version": "publishedVersion", "is_accepted": true, "is_published": true, "raw_source_name": "INTERNATIONAL JOURNAL OF COMPUTER ENGINEERING AND TECHNOLOGY", "raw_type": "journal-article"}, "cited_by_count": 0, "is_retracted": false}, {"id": "https://openalex.org/W7214561766", "doi": "https://doi.org/10.5281/zenodo.22994183", "title": "Telepathic Silicon: The Paradigm Shift to Latent-Space Collaboration in Multi-Agent Systems", "publication_year": 2025, "publication_date": "2025-12-05", "type": "article", "language": null, "authorships": [{"author_position": "first", "author": {"id": "https://openalex.org/A5134831847", "display_name": "Syed Muntasir Mamun", "orcid": "https://orcid.org/0000-0001-6845-2853", "observed_orcids": ["https://orcid.org/0000-0001-6845-2853"]}, "institutions": [{"id": "https://openalex.org/I4405252681", "display_name": "Ministry of Foreign Affairs, Dhaka", "ror": "https://ror.org/03s3w2m84", "country_code": null, "type": "government", "lineage": ["https://openalex.org/I4405252681"]}], "countries": ["BD"], "is_corresponding": true, "raw_author_name": "Mamun, Syed Muntasir", "raw_affiliation_strings": ["Ministry of Foreign Affairs, Bangladesh"], "raw_orcid": "https://orcid.org/0000-0001-6845-2853", "affiliations": [{"raw_affiliation_string": "Ministry of Foreign Affairs, Bangladesh", "institution_ids": ["https://openalex.org/I4405252681"]}]}], "abstract_inverted_index": {"The": [0, 164], "prevailing": [1], "architecture": [2, 128], "for": [3, 22], "Multi-Agent": [4, 52, 171], "Systems": [5, 172], "(MAS)": [6], "has": [7], "long": [8], "mimicked": [9], "human": [10], "organizational": [11], "structures,": [12], "relying": [13], "on": [14, 175], "natural": [15], "language": [16], "text": [17], "as": [18], "the": [19, 69, 75, 79, 85, 97, 105, 130, 138], "primary": [20], "medium": [21], "inter-agent": [23], "coordination.": [24], "While": [25], "interpretable,": [26], "this": [27, 102, 127], "approach": [28], "introduces": [29], "a": [30, 45, 60, 133], "severe": [31], "\"information": [32], "bottleneck,\"": [33], "compressing": [34], "high-dimensional": [35], "neural": [36], "states": [37], "into": [38], "low-bandwidth": [39], "discrete": [40], "tokens.": [41], "This": [42], "review": [43], "provides": [44], "critical": [46], "analysis": [47, 95], "of": [48, 78, 88, 96, 132, 141, 146], "\"Latent": [49], "Collaboration": [50, 169], "in": [51, 170], "Systems\"": [53], "(LatentMAS)": [54], "by": [55, 112, 119], "Zou": [56], "et": [57], "al.": [58], "(2025),": [59], "framework": [61], "that": [62, 101, 126], "enables": [63], "agents": [64], "to": [65, 121, 153, 167], "collaborate": [66], "entirely": [67], "within": [68], "continuous": [70], "latent": [71], "space.": [72], "We": [73], "examine": [74], "mathematical": [76], "rigor": [77], "Input-Output": [80], "Alignment": [81], "(W_a)": [82], "mechanism": [83], "and": [84, 137, 150], "information-theoretic": [86], "implications": [87], "lossless": [89], "Key-Value": [90], "(KV)": [91], "cache": [92], "transfer.": [93], "Statistical": [94], "source": [98], "data": [99], "reveals": [100], "method": [103], "breaks": [104], "traditional": [106], "speed-accuracy": [107], "trade-off,": [108], "reducing": [109], "token": [110], "usage": [111], "over": [113], "80%": [114], "while": [115], "enhancing": [116], "reasoning": [117, 147], "accuracy": [118], "up": [120], "14.6%.": [122], "Furthermore,": [123], "we": [124], "posit": [125], "signals": [129], "emergence": [131], "post-linguistic": [134], "\"model": [135], "language\"": [136], "imminent": [139], "proliferation": [140], "cybernetic": [142], "organisms\u2014synthetic": [143], "hives": [144], "capable": [145], "at": [148], "speeds": [149], "densities": [151], "inaccessible": [152], "biological": [154], "cognition.": [155], "Mamun,": [156], "SM": [157], "&": [158], "Gemini": [159], "3": [160], "(2025)\u2026\u2026\u2026\u2026\u2026\u2026\u2026\u2026\u2026\u2026..1": [161], "Telepathic": [162], "Silicon:": [163], "Paradigm": [165], "Shift": [166], "Latent-Space": [168], "Also": [173], "available": [174], "Academia.edu:": [176], "https://www.academia.edu/attachments/125636592/download_file?s=portfolio": [177]}, "primary_topic": {"id": "https://openalex.org/T10456", "display_name": "Multi-Agent Systems and Negotiation", "score": 0.9312000274658203, "subfield": {"id": "https://openalex.org/subfields/1702", "display_name": "Artificial Intelligence"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, "topics": [{"id": "https://openalex.org/T10456", "display_name": "Multi-Agent Systems and Negotiation", "score": 0.9312000274658203, "subfield": {"id": "https://openalex.org/subfields/1702", "display_name": "Artificial Intelligence"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, {"id": "https://openalex.org/T14381", "display_name": "Psychiatry, Mental Health, Neuroscience", "score": 0.01600000075995922, "subfield": {"id": "https://openalex.org/subfields/1702", "display_name": "Artificial Intelligence"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, {"id": "https://openalex.org/T10028", "display_name": "Topic Modeling", "score": 0.01269999984651804, "subfield": {"id": "https://openalex.org/subfields/1702", "display_name": "Artificial Intelligence"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}], "primary_location": {"id": "doi:10.5281/zenodo.22994183", "is_oa": true, "landing_page_url": "https://doi.org/10.5281/zenodo.22994183", "pdf_url": null, "source": {"id": "https://openalex.org/S4306400562", "display_name": "Zenodo (CERN European Organization for Nuclear Research)", "issn_l": null, "issn": null, "is_oa": true, "is_in_doaj": false, "is_core": false, "listed_in": [], "host_organization": "https://openalex.org/I67311998", "host_organization_name": "European Organization for Nuclear Research", "host_organization_lineage": ["https://openalex.org/I67311998"], "host_organization_lineage_names": [], "type": "repository"}, "license": "cc-by", "license_id": "https://openalex.org/licenses/cc-by", "version": null, "is_accepted": false, "is_published": false, "raw_source_name": null, "raw_type": "JournalArticle"}, "cited_by_count": 0, "is_retracted": false}, {"id": "https://openalex.org/W4414762093", "doi": "https://doi.org/10.48550/arxiv.2509.05165", "title": "KVCompose: Efficient Structured KV Cache Compression with Composite Tokens", "publication_year": 2025, "publication_date": "2025-09-05", "type": "preprint", "language": "en", "authorships": [{"author_position": "first", "author": {"id": "https://openalex.org/A5119820576", "display_name": "Dmitry Akulov", "orcid": null, "observed_orcids": []}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Akulov, Dmitry", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5028270135", "display_name": "Mohamed Sana", "orcid": "https://orcid.org/0000-0001-5644-5840", "observed_orcids": ["https://orcid.org/0000-0001-5644-5840"]}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Sana, Mohamed", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5045740906", "display_name": "Antonio De Domenico", "orcid": "https://orcid.org/0000-0003-1229-4045", "observed_orcids": ["https://orcid.org/0000-0003-1229-4045"]}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "De Domenico, Antonio", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5119536158", "display_name": "Tareq Salem", "orcid": null, "observed_orcids": []}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Salem, Tareq Si", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5048382270", "display_name": "Nicola Piovesan", "orcid": "https://orcid.org/0000-0001-5397-5248", "observed_orcids": ["https://orcid.org/0000-0001-5397-5248"]}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Piovesan, Nicola", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "last", "author": {"id": "https://openalex.org/A5064737739", "display_name": "Fadhel Ayed", "orcid": "https://orcid.org/0009-0008-4234-7682", "observed_orcids": ["https://orcid.org/0009-0008-4234-7682"]}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Ayed, Fadhel", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}], "abstract_inverted_index": {"Large": [0], "language": [1], "models": [2], "(LLMs)": [3], "rely": [4], "on": [5, 62], "key-value": [6], "(KV)": [7], "caches": [8], "for": [9, 147], "efficient": [10, 148], "autoregressive": [11], "decoding;": [12], "however,": [13], "cache": [14, 33, 58, 90], "size": [15], "grows": [16], "linearly": [17], "with": [18, 43, 112, 137], "context": [19], "length": [20], "and": [21, 80, 128, 144], "model": [22], "depth,": [23], "becoming": [24], "a": [25, 53, 142], "major": [26], "bottleneck": [27], "in": [28], "long-context": [29, 149], "inference.": [30], "Prior": [31], "KV": [32, 57], "compression": [34, 59], "methods": [35], "either": [36], "enforce": [37], "rigid": [38], "heuristics,": [39], "disrupt": [40], "tensor": [41], "layouts": [42], "per-attention-head": [44], "variability,": [45], "or": [46], "require": [47], "specialized": [48], "compute": [49], "kernels.": [50], "We": [51], "propose": [52], "simple,": [54], "yet": [55], "effective,": [56], "framework": [60], "based": [61], "attention-guided,": [63], "layer-adaptive": [64], "composite": [65, 84], "tokens.": [66, 114], "Our": [67], "method": [68], "aggregates": [69], "attention": [70], "scores": [71], "to": [72, 110], "estimate": [73], "token": [74], "importance,": [75], "selects": [76], "head-specific": [77], "tokens": [78, 85], "independently,": [79], "aligns": [81], "them": [82], "into": [83], "that": [86], "respect": [87], "the": [88], "uniform": [89], "structure": [91], "required": [92], "by": [93], "existing": [94], "inference": [95, 139], "engines.": [96], "A": [97], "global": [98], "allocation": [99], "mechanism": [100], "further": [101], "adapts": [102], "retention": [103], "budgets": [104], "across": [105], "layers,": [106], "assigning": [107], "more": [108], "capacity": [109], "layers": [111], "informative": [113], "This": [115], "approach": [116, 133], "achieves": [117], "significant": [118], "memory": [119], "reduction": [120], "while": [121], "preserving": [122], "accuracy,": [123], "consistently": [124], "outperforming": [125], "prior": [126], "structured": [127], "semi-structured": [129], "methods.": [130], "Crucially,": [131], "our": [132], "remains": [134], "fully": [135], "compatible": [136], "standard": [138], "pipelines,": [140], "offering": [141], "practical": [143], "scalable": [145], "solution": [146], "LLM": [150], "deployment.": [151]}, "primary_topic": {"id": "https://openalex.org/T10181", "display_name": "Natural Language Processing Techniques", "score": 0.3935999870300293, "subfield": {"id": "https://openalex.org/subfields/1702", "display_name": "Artificial Intelligence"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, "topics": [{"id": "https://openalex.org/T10181", "display_name": "Natural Language Processing Techniques", "score": 0.3935999870300293, "subfield": {"id": "https://openalex.org/subfields/1702", "display_name": "Artificial Intelligence"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, {"id": "https://openalex.org/T10028", "display_name": "Topic Modeling", "score": 0.382099986076355, "subfield": {"id": "https://openalex.org/subfields/1702", "display_name": "Artificial Intelligence"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, {"id": "https://openalex.org/T10036", "display_name": "Advanced Neural Network Applications", "score": 0.14100000262260437, "subfield": {"id": "https://openalex.org/subfields/1707", "display_name": "Computer Vision and Pattern Recognition"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}], "primary_location": {"id": "pmh:oai:arXiv.org:2509.05165", "is_oa": true, "landing_page_url": "http://arxiv.org/abs/2509.05165", "pdf_url": "https://arxiv.org/pdf/2509.05165", "source": {"id": "https://openalex.org/S4306400194", "display_name": "arXiv (Cornell University)", "issn_l": "2331-8422", "issn": ["2331-8422"], "is_oa": true, "is_in_doaj": false, "is_core": false, "listed_in": [], "host_organization": "https://openalex.org/I205783295", "host_organization_name": "Cornell University", "host_organization_lineage": ["https://openalex.org/I205783295"], "host_organization_lineage_names": [], "type": "repository"}, "license": null, "license_id": null, "version": "submittedVersion", "is_accepted": false, "is_published": false, "raw_source_name": null, "raw_type": "text"}, "cited_by_count": 0, "is_retracted": false}, {"id": "https://openalex.org/W4416302611", "doi": "https://doi.org/10.18535/ijsrm/v13i11.ec03", "title": "Multi-Objective Reinforcement Learning for Resource-Optimal LLM Serving in SaaS Clouds", "publication_year": 2025, "publication_date": "2025-11-16", "type": "article", "language": null, "authorships": [{"author_position": "first", "author": {"id": "https://openalex.org/A5120407001", "display_name": "Vasanthi Jangala Naga", "orcid": null, "observed_orcids": []}, "institutions": [{"id": "https://openalex.org/I4210132246", "display_name": "Workday (United States)", "ror": "https://ror.org/02nn1vm89", "country_code": "US", "type": "company", "lineage": ["https://openalex.org/I4210132246"]}], "countries": ["US"], "is_corresponding": true, "raw_author_name": "Vasanthi Jangala Naga", "raw_affiliation_strings": ["Workday"], "raw_orcid": null, "affiliations": [{"raw_affiliation_string": "Workday", "institution_ids": ["https://openalex.org/I4210132246"]}]}], "abstract_inverted_index": {"Large": [0], "Language": [1], "Models": [2], "(LLMs)": [3], "have": [4, 18, 89], "been": [5], "the": [6, 52, 63, 73], "cornerstone": [7], "for": [8], "current": [9, 27], "Software": [10, 76], "as": [11, 77], "a": [12, 35, 78, 83, 90], "Service": [13, 79], "(SaaS)": [14, 80], "solutions.": [15], "These": [16], "LLMs": [17], "made": [19], "intelligent": [20], "automation": [21], "and": [22, 81], "analytics": [23], "possible.": [24], "But": [25], "their": [26], "computation": [28], "or": [29], "inference": [30], "cost": [31], "is": [32, 51], "high.": [33], "As": [34], "result,": [36], "cloud": [37, 45], "service": [38], "companies": [39], "face": [40], "challenges": [41], "with": [42], "respect": [43], "to": [44, 88], "scalability.": [46], "Adaptive": [47, 68], "Precision": [48, 69], "Scaling": [49, 70], "(APS)": [50, 71], "strategy": [53], "of": [54, 67, 75, 85, 93], "adapting": [55], "computational": [56], "precision": [57, 86, 94], "during": [58], "execution.": [59], "This": [60], "paper": [61], "describes": [62], "newly": [64], "proposed": [65], "architecture": [66], "in": [72], "context": [74], "proposes": [82], "taxonomy": [84], "scaling": [87], "clearer": [91], "understanding": [92], "adaptivity.": [95]}, "primary_topic": {"id": "https://openalex.org/T10101", "display_name": "Cloud Computing and Resource Management", "score": 0.9980000257492065, "subfield": {"id": "https://openalex.org/subfields/1710", "display_name": "Information Systems"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, "topics": [{"id": "https://openalex.org/T10101", "display_name": "Cloud Computing and Resource Management", "score": 0.9980000257492065, "subfield": {"id": "https://openalex.org/subfields/1710", "display_name": "Information Systems"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, {"id": "https://openalex.org/T14098", "display_name": "Energy Efficiency in Computing", "score": 0.0008999999845400453, "subfield": {"id": "https://openalex.org/subfields/1708", "display_name": "Hardware and Architecture"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, {"id": "https://openalex.org/T10054", "display_name": "Parallel Computing and Optimization Techniques", "score": 0.00019999999494757503, "subfield": {"id": "https://openalex.org/subfields/1708", "display_name": "Hardware and Architecture"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}], "primary_location": {"id": "doi:10.18535/ijsrm/v13i11.ec03", "is_oa": true, "landing_page_url": "https://doi.org/10.18535/ijsrm/v13i11.ec03", "pdf_url": "https://ijsrm.net/index.php/ijsrm/article/download/6764/4180", "source": {"id": "https://openalex.org/S4210219273", "display_name": "International Journal of Scientific Research and Management (IJSRM)", "issn_l": "2321-3418", "issn": ["2321-3418"], "is_oa": true, "is_in_doaj": false, "is_core": false, "listed_in": [], "host_organization": null, "host_organization_name": null, "host_organization_lineage": [], "host_organization_lineage_names": [], "type": "journal"}, "license": "cc-by", "license_id": "https://openalex.org/licenses/cc-by", "version": "publishedVersion", "is_accepted": true, "is_published": true, "raw_source_name": "International Journal of Scientific Research and Management (IJSRM)", "raw_type": "journal-article"}, "cited_by_count": 0, "is_retracted": false}, {"id": "https://openalex.org/W4414580758", "doi": "https://doi.org/10.48550/arxiv.2505.16502", "title": "Recursive Offloading for LLM Serving in Multi-tier Networks", "publication_year": 2025, "publication_date": "2025-05-22", "type": "preprint", "language": "en", "authorships": [{"author_position": "first", "author": {"id": "https://openalex.org/A5101698317", "display_name": "Zhiyuan Wu", "orcid": "https://orcid.org/0000-0002-8925-4896", "observed_orcids": ["https://orcid.org/0000-0002-8925-4896"]}, "institutions": [{"id": "https://openalex.org/I4210090176", "display_name": "Institute of Computing Technology", "ror": "https://ror.org/0090r4d87", "country_code": "CN", "type": "facility", "lineage": ["https://openalex.org/I19820366", "https://openalex.org/I4210090176"]}], "countries": ["CN"], "is_corresponding": false, "raw_author_name": "Wu, Zhiyuan", "raw_affiliation_strings": ["State Key Lab of Processers , Institute of Computing Technology , Chinese Academy of Sciences , Beijing , China,"], "raw_orcid": null, "affiliations": [{"raw_affiliation_string": "State Key Lab of Processers , Institute of Computing Technology , Chinese Academy of Sciences , Beijing , China,", "institution_ids": ["https://openalex.org/I4210090176"]}]}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5012804259", "display_name": "Sheng Sun", "orcid": "https://orcid.org/0000-0002-0260-2692", "observed_orcids": ["https://orcid.org/0000-0002-0260-2692"]}, "institutions": [{"id": "https://openalex.org/I4210090176", "display_name": "Institute of Computing Technology", "ror": "https://ror.org/0090r4d87", "country_code": "CN", "type": "facility", "lineage": ["https://openalex.org/I19820366", "https://openalex.org/I4210090176"]}], "countries": ["CN"], "is_corresponding": false, "raw_author_name": "Sun, Sheng", "raw_affiliation_strings": ["State Key Lab of Processers , Institute of Computing Technology , Chinese Academy of Sciences , Beijing , China,"], "raw_orcid": null, "affiliations": [{"raw_affiliation_string": "State Key Lab of Processers , Institute of Computing Technology , Chinese Academy of Sciences , Beijing , China,", "institution_ids": ["https://openalex.org/I4210090176"]}]}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5100402336", "display_name": "Yuwei Wang", "orcid": "https://orcid.org/0000-0003-4282-9821", "observed_orcids": ["https://orcid.org/0000-0003-4282-9821"]}, "institutions": [{"id": "https://openalex.org/I4210090176", "display_name": "Institute of Computing Technology", "ror": "https://ror.org/0090r4d87", "country_code": "CN", "type": "facility", "lineage": ["https://openalex.org/I19820366", "https://openalex.org/I4210090176"]}], "countries": ["CN"], "is_corresponding": false, "raw_author_name": "Wang, Yuwei", "raw_affiliation_strings": ["State Key Lab of Processers , Institute of Computing Technology , Chinese Academy of Sciences , Beijing , China,"], "raw_orcid": null, "affiliations": [{"raw_affiliation_string": "State Key Lab of Processers , Institute of Computing Technology , Chinese Academy of Sciences , Beijing , China,", "institution_ids": ["https://openalex.org/I4210090176"]}]}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5102349269", "display_name": "Min Liu", "orcid": "https://orcid.org/0009-0007-2843-2648", "observed_orcids": ["https://orcid.org/0009-0007-2843-2648"]}, "institutions": [{"id": "https://openalex.org/I4210090176", "display_name": "Institute of Computing Technology", "ror": "https://ror.org/0090r4d87", "country_code": "CN", "type": "facility", "lineage": ["https://openalex.org/I19820366", "https://openalex.org/I4210090176"]}], "countries": ["CN"], "is_corresponding": false, "raw_author_name": "Liu, Min", "raw_affiliation_strings": ["State Key Lab of Processers , Institute of Computing Technology , Chinese Academy of Sciences , Beijing , China,"], "raw_orcid": null, "affiliations": [{"raw_affiliation_string": "State Key Lab of Processers , Institute of Computing Technology , Chinese Academy of Sciences , Beijing , China,", "institution_ids": ["https://openalex.org/I4210090176"]}]}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5114833676", "display_name": "Bo Gao", "orcid": "https://orcid.org/0009-0008-0894-7959", "observed_orcids": ["https://orcid.org/0009-0008-0894-7959"]}, "institutions": [{"id": "https://openalex.org/I4210090176", "display_name": "Institute of Computing Technology", "ror": "https://ror.org/0090r4d87", "country_code": "CN", "type": "facility", "lineage": ["https://openalex.org/I19820366", "https://openalex.org/I4210090176"]}], "countries": ["CN"], "is_corresponding": false, "raw_author_name": "Gao, Bo", "raw_affiliation_strings": ["State Key Lab of Processers , Institute of Computing Technology , Chinese Academy of Sciences , Beijing , China,"], "raw_orcid": null, "affiliations": [{"raw_affiliation_string": "State Key Lab of Processers , Institute of Computing Technology , Chinese Academy of Sciences , Beijing , China,", "institution_ids": ["https://openalex.org/I4210090176"]}]}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5020470345", "display_name": "Jinda Lu", "orcid": null, "observed_orcids": []}, "institutions": [{"id": "https://openalex.org/I4210090176", "display_name": "Institute of Computing Technology", "ror": "https://ror.org/0090r4d87", "country_code": "CN", "type": "facility", "lineage": ["https://openalex.org/I19820366", "https://openalex.org/I4210090176"]}], "countries": ["CN"], "is_corresponding": false, "raw_author_name": "Lu, Jinda", "raw_affiliation_strings": ["State Key Lab of Processers , Institute of Computing Technology , Chinese Academy of Sciences , Beijing , China,"], "raw_orcid": null, "affiliations": [{"raw_affiliation_string": "State Key Lab of Processers , Institute of Computing Technology , Chinese Academy of Sciences , Beijing , China,", "institution_ids": ["https://openalex.org/I4210090176"]}]}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5102902954", "display_name": "Zheming Yang", "orcid": "https://orcid.org/0000-0001-9957-5792", "observed_orcids": ["https://orcid.org/0000-0001-9957-5792"]}, "institutions": [{"id": "https://openalex.org/I4210090176", "display_name": "Institute of Computing Technology", "ror": "https://ror.org/0090r4d87", "country_code": "CN", "type": "facility", "lineage": ["https://openalex.org/I19820366", "https://openalex.org/I4210090176"]}], "countries": ["CN"], "is_corresponding": false, "raw_author_name": "Yang, Zheming", "raw_affiliation_strings": ["State Key Lab of Processers , Institute of Computing Technology , Chinese Academy of Sciences , Beijing , China,"], "raw_orcid": null, "affiliations": [{"raw_affiliation_string": "State Key Lab of Processers , Institute of Computing Technology , Chinese Academy of Sciences , Beijing , China,", "institution_ids": ["https://openalex.org/I4210090176"]}]}, {"author_position": "last", "author": {"id": null, "display_name": "Wen, Tian", "orcid": null, "observed_orcids": []}, "institutions": [{"id": "https://openalex.org/I4210090176", "display_name": "Institute of Computing Technology", "ror": "https://ror.org/0090r4d87", "country_code": "CN", "type": "facility", "lineage": ["https://openalex.org/I19820366", "https://openalex.org/I4210090176"]}], "countries": ["CN"], "is_corresponding": false, "raw_author_name": "Wen, Tian", "raw_affiliation_strings": ["State Key Lab of Processers , Institute of Computing Technology , Chinese Academy of Sciences , Beijing , China,"], "raw_orcid": null, "affiliations": [{"raw_affiliation_string": "State Key Lab of Processers , Institute of Computing Technology , Chinese Academy of Sciences , Beijing , China,", "institution_ids": ["https://openalex.org/I4210090176"]}]}], "abstract_inverted_index": {"Heterogeneous": [0], "device-edge-cloud": [1], "computing": [2], "infrastructures": [3], "have": [4], "become": [5], "widely": [6], "adopted": [7], "in": [8, 133, 153, 202], "telecommunication": [9], "operators": [10], "and": [11, 38, 76, 114, 160, 188, 206, 209], "Wide": [12], "Area": [13], "Networks": [14], "(WANs),": [15], "offering": [16], "multi-tier": [17, 44, 97, 134], "computational": [18], "support": [19], "for": [20, 130], "emerging": [21], "intelligent": [22, 166], "services.": [23], "With": [24], "the": [25, 90, 211], "rapid": [26], "proliferation": [27], "of": [28, 184], "Large": [29], "Language": [30], "Model": [31], "(LLM)": [32], "services,": [33], "efficiently": [34], "coordinating": [35], "inference": [36, 92], "tasks": [37], "reducing": [39], "communication": [40, 79, 207, 212], "overhead": [41, 80], "within": [42], "these": [43, 120], "network": [45, 112], "architectures": [46], "becomes": [47], "a": [48, 125, 138, 173], "critical": [49], "deployment": [50, 60, 71], "challenge.": [51], "Existing": [52], "LLM": [53, 131], "serving": [54, 132], "paradigms": [55], "exhibit": [56], "significant": [57], "limitations:": [58], "on-device": [59], "supports": [61], "only": [62], "lightweight": [63], "LLMs": [64, 156], "due": [65], "to": [66, 96, 110, 218], "hardware": [67], "constraints,": [68], "while": [69], "cloud-centric": [70], "suffers": [72], "from": [73], "resource": [74], "congestion": [75], "considerable": [77], "prompt": [78], "caused": [81], "by": [82, 214], "frequent": [83], "service": [84, 204], "requests": [85], "during": [86], "peak": [87], "periods.": [88], "Although": [89], "model-cascading-based": [91], "strategy": [93, 177], "adapts": [94], "better": [95], "networks,": [98], "its": [99], "reliance": [100], "on": [101, 149, 194], "fine-grained,": [102], "manually": [103], "adjusted": [104], "thresholds": [105], "makes": [106], "it": [107], "less": [108], "responsive": [109], "dynamic": [111, 175], "conditions": [113], "varying": [115], "task": [116, 151, 167], "complexities.": [117], "To": [118, 163], "address": [119], "challenges,": [121], "we": [122], "propose": [123], "RecServe,": [124], "recursive": [126], "offloading": [127, 146, 176, 190], "framework": [128], "tailored": [129], "networks.": [135], "RecServe": [136, 171, 199], "integrates": [137], "task-specific": [139], "hierarchical": [140], "confidence": [141, 186], "evaluation": [142], "mechanism": [143], "that": [144, 198], "guides": [145], "decisions": [147], "based": [148], "inferred": [150], "complexity": [152], "progressively": [154], "scaled": [155], "across": [157, 169], "device,": [158], "edge,": [159], "cloud": [161], "tiers.": [162], "further": [164], "enable": [165], "routing": [168], "tiers,": [170], "employs": [172], "sliding-window-based": [174], "with": [178], "quantile": [179], "interpolation,": [180], "enabling": [181], "real-time": [182], "tracking": [183], "historical": [185], "distributions": [187], "adaptive": [189], "threshold": [191], "adjustments.": [192], "Experiments": [193], "eight": [195], "datasets": [196], "demonstrate": [197], "outperforms": [200], "CasServe": [201], "both": [203], "quality": [205], "efficiency,": [208], "reduces": [210], "burden": [213], "over": [215], "50\\%": [216], "compared": [217], "centralized": [219], "cloud-based": [220], "serving.": [221]}, "primary_topic": {"id": "https://openalex.org/T10273", "display_name": "IoT and Edge/Fog Computing", "score": 0.9970999956130981, "subfield": {"id": "https://openalex.org/subfields/1705", "display_name": "Computer Networks and Communications"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, "topics": [{"id": "https://openalex.org/T10273", "display_name": "IoT and Edge/Fog Computing", "score": 0.9970999956130981, "subfield": {"id": "https://openalex.org/subfields/1705", "display_name": "Computer Networks and Communications"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, {"id": "https://openalex.org/T10714", "display_name": "Software-Defined Networks and 5G", "score": 0.00039999998989515007, "subfield": {"id": "https://openalex.org/subfields/1705", "display_name": "Computer Networks and Communications"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, {"id": "https://openalex.org/T11478", "display_name": "Caching and Content Delivery", "score": 0.00019999999494757503, "subfield": {"id": "https://openalex.org/subfields/1705", "display_name": "Computer Networks and Communications"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}], "primary_location": {"id": "pmh:oai:arXiv.org:2505.16502", "is_oa": true, "landing_page_url": "http://arxiv.org/abs/2505.16502", "pdf_url": "https://arxiv.org/pdf/2505.16502", "source": {"id": "https://openalex.org/S4306400194", "display_name": "arXiv (Cornell University)", "issn_l": "2331-8422", "issn": ["2331-8422"], "is_oa": true, "is_in_doaj": false, "is_core": false, "listed_in": [], "host_organization": "https://openalex.org/I205783295", "host_organization_name": "Cornell University", "host_organization_lineage": ["https://openalex.org/I205783295"], "host_organization_lineage_names": [], "type": "repository"}, "license": "cc-by", "license_id": "https://openalex.org/licenses/cc-by", "version": "submittedVersion", "is_accepted": false, "is_published": false, "raw_source_name": null, "raw_type": "text"}, "cited_by_count": 0, "is_retracted": false}, {"id": "https://openalex.org/W4417086754", "doi": "https://doi.org/10.48550/arxiv.2512.04515", "title": "EgoLCD: Egocentric Video Generation with Long Context Diffusion", "publication_year": 2025, "publication_date": "2025-12-04", "type": "preprint", "language": null, "authorships": [{"author_position": "first", "author": {"id": null, "display_name": "Zhang, Liuzhou", "orcid": null, "observed_orcids": []}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Zhang, Liuzhou", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5068572222", "display_name": "Jiarui Ye", "orcid": "https://orcid.org/0009-0008-0517-3732", "observed_orcids": ["https://orcid.org/0009-0008-0517-3732"]}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Ye, Jiarui", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5007652343", "display_name": "Yuanlei Wang", "orcid": null, "observed_orcids": []}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Wang, Yuanlei", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": null, "display_name": "Zhong, Ming", "orcid": null, "observed_orcids": []}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Zhong, Ming", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": null, "display_name": "Cao, Mingju", "orcid": null, "observed_orcids": []}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Cao, Mingju", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": null, "display_name": "Xia, Wanke", "orcid": null, "observed_orcids": []}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Xia, Wanke", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": null, "display_name": "Zeng, Bowen", "orcid": null, "observed_orcids": []}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Zeng, Bowen", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5100358736", "display_name": "Zeyu Zhang", "orcid": "https://orcid.org/0000-0002-7157-6272", "observed_orcids": ["https://orcid.org/0000-0002-7157-6272"]}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Zhang, Zeyu", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "last", "author": {"id": "https://openalex.org/A5102376669", "display_name": "Hao Tang", "orcid": null, "observed_orcids": []}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Tang, Hao", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}], "abstract_inverted_index": {"Generating": [0], "long,": [1], "coherent": [2], "egocentric": [3, 44], "videos": [4], "is": [5], "difficult,": [6], "as": [7, 53], "hand-object": [8], "interactions": [9], "and": [10, 27, 58, 92, 116, 123], "procedural": [11], "tasks": [12], "require": [13], "reliable": [14], "long-term": [15], "memory.": [16], "Existing": [17], "autoregressive": [18], "models": [19, 132], "suffer": [20], "from": [21], "content": [22], "drift,": [23], "where": [24], "object": [25], "identity": [26], "scene": [28], "semantics": [29], "degrade": [30], "over": [31], "time.": [32], "To": [33], "address": [34], "this": [35], "challenge,": [36], "we": [37], "introduce": [38], "EgoLCD,": [39], "an": [40, 74], "end-to-end": [41], "framework": [42], "for": [43, 69, 81, 133], "long-context": [45], "video": [46, 51], "generation": [47], "that": [48, 107], "treats": [49], "long": [50], "synthesis": [52], "a": [54, 64, 125], "problem": [55], "of": [56], "efficient": [57], "stable": [59, 70], "memory": [60, 90], "management.": [61], "EgoLCD": [62, 108], "combines": [63], "Long-Term": [65], "Sparse": [66], "KV": [67], "Cache": [68], "global": [71], "context": [72], "with": [73], "attention-based": [75], "short-term": [76], "memory,": [77], "extended": [78], "by": [79], "LoRA": [80], "local": [82], "adaptation.": [83], "A": [84], "Memory": [85], "Regulation": [86], "Loss": [87], "enforces": [88], "consistent": [89], "usage,": [91], "Structured": [93], "Narrative": [94], "Prompting": [95], "provides": [96], "explicit": [97], "temporal": [98, 117], "guidance.": [99], "Extensive": [100], "experiments": [101], "on": [102], "the": [103], "EgoVid-5M": [104], "benchmark": [105], "demonstrate": [106], "achieves": [109], "state-of-the-art": [110], "performance": [111], "in": [112], "both": [113], "perceptual": [114], "quality": [115], "consistency,": [118], "effectively": [119], "mitigating": [120], "generative": [121], "forgetting": [122], "representing": [124], "significant": [126], "step": [127], "toward": [128], "building": [129], "scalable": [130], "world": [131], "embodied": [134], "AI.": [135], "Code:": [136], "https://github.com/AIGeeksGroup/EgoLCD.": [137], "Website:": [138], "https://aigeeksgroup.github.io/EgoLCD.": [139]}, "primary_topic": {"id": "https://openalex.org/T10775", "display_name": "Generative Adversarial Networks and Image Synthesis", "score": 0.9951000213623047, "subfield": {"id": "https://openalex.org/subfields/1707", "display_name": "Computer Vision and Pattern Recognition"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, "topics": [{"id": "https://openalex.org/T10775", "display_name": "Generative Adversarial Networks and Image Synthesis", "score": 0.9951000213623047, "subfield": {"id": "https://openalex.org/subfields/1707", "display_name": "Computer Vision and Pattern Recognition"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, {"id": "https://openalex.org/T11714", "display_name": "Multimodal Machine Learning Applications", "score": 0.0044999998062849045, "subfield": {"id": "https://openalex.org/subfields/1707", "display_name": "Computer Vision and Pattern Recognition"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, {"id": "https://openalex.org/T11605", "display_name": "Visual Attention and Saliency Detection", "score": 0.0, "subfield": {"id": "https://openalex.org/subfields/1707", "display_name": "Computer Vision and Pattern Recognition"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}], "primary_location": {"id": "pmh:oai:arXiv.org:2512.04515", "is_oa": true, "landing_page_url": "http://arxiv.org/abs/2512.04515", "pdf_url": "https://arxiv.org/pdf/2512.04515", "source": {"id": "https://openalex.org/S4306400194", "display_name": "arXiv (Cornell University)", "issn_l": "2331-8422", "issn": ["2331-8422"], "is_oa": true, "is_in_doaj": false, "is_core": false, "listed_in": [], "host_organization": "https://openalex.org/I205783295", "host_organization_name": "Cornell University", "host_organization_lineage": ["https://openalex.org/I205783295"], "host_organization_lineage_names": [], "type": "repository"}, "license": "cc-by-nc-sa", "license_id": "https://openalex.org/licenses/cc-by-nc-sa", "version": "submittedVersion", "is_accepted": false, "is_published": false, "raw_source_name": null, "raw_type": "text"}, "cited_by_count": 0, "is_retracted": false}, {"id": "https://openalex.org/W7104407851", "doi": "https://doi.org/10.5281/zenodo.17547016", "title": "AI Sleep Constructs: Implementing Chronological Intelligence in Stateful Systems", "publication_year": 2025, "publication_date": "2025-11-07", "type": "article", "language": "en", "authorships": [{"author_position": "first", "author": {"id": null, "display_name": "Smith, Tionne", "orcid": "https://orcid.org/0009-0008-8627-6150", "observed_orcids": ["https://orcid.org/0009-0008-8627-6150"]}, "institutions": [], "countries": [], "is_corresponding": true, "raw_author_name": "Smith, Tionne", "raw_affiliation_strings": ["Antiparty Press"], "raw_orcid": "https://orcid.org/0009-0008-8627-6150", "affiliations": [{"raw_affiliation_string": "Antiparty Press", "institution_ids": []}]}], "abstract_inverted_index": {"AI": [0, 123, 130], "Sleep": [1], "Constructs": [2], "is": [3], "a": [4, 89], "production-ready": [5], "Python": [6], "framework": [7, 83], "that": [8, 30], "brings": [9], "temporal": [10], "awareness": [11], "and": [12, 26, 39, 52, 64], "self-maintenance": [13], "to": [14, 34, 99], "artificial": [15], "intelligence": [16], "models.": [17], "Inspired": [18], "by": [19], "biological": [20], "sleep,": [21, 131], "it": [22, 112], "introduces": [23], "engineered": [24], "light": [25], "deep": [27, 56], "sleep": [28, 45, 57], "cycles": [29], "enable": [31], "language": [32], "models": [33, 96], "optimize": [35], "performance,": [36], "consolidate": [37], "memory,": [38], "self-heal": [40], "between": [41], "inference": [42, 102], "sessions.": [43], "Light": [44], "modes": [46, 58], "perform": [47], "gradient": [48], "clipping,": [49], "cache": [50], "management,": [51], "adaptive": [53], "learning,": [54], "while": [55], "execute": [59], "semantic": [60], "consolidation,": [61], "layer": [62], "recalibration,": [63], "diagnostic": [65], "repair.": [66], "Contextual": [67], "triggers": [68], "automatically": [69], "govern": [70], "these": [71], "transitions": [72], "based": [73], "on": [74, 119], "model": [75], "drift,": [76], "performance": [77], "anomalies,": [78], "or": [79], "workload": [80], "scheduling.": [81], "The": [82], "also": [84], "includes": [85], "Cache-to-Cache": [86], "(C2C)": [87], "fusion,": [88], "novel": [90], "technique": [91], "allowing": [92], "KV-cache": [93], "transfer": [94], "across": [95], "for": [97, 117], "up": [98], "10\u00d7": [100], "faster": [101], "with": [103, 108], "minimal": [104], "quality": [105], "loss.": [106], "Integrated": [107], "Hugging": [109], "Face": [110], "APIs,": [111], "provides": [113], "an": [114], "extensible": [115], "foundation": [116], "research": [118], "stateful,": [120], "chronologically": [121], "intelligent": [122], "architectures.": [124], "Repository": [125], "URL": [126], "(Git):": [127], "https://github.com/electricwolfemarshmallowhypertext/ai-sleep/": [128], "Keywords:": [129], "Chronological": [132], "Intelligence,": [133], "Offline": [134], "Optimization,": [135], "Adaptive": [136], "Learning,": [137], "C2C": [138], "Fusion,": [139], "Stateful": [140], "AI,": [141], "Antiparty": [142], "Inc.": [143]}, "primary_topic": {"id": "https://openalex.org/T10028", "display_name": "Topic Modeling", "score": 0.25220000743865967, "subfield": {"id": "https://openalex.org/subfields/1702", "display_name": "Artificial Intelligence"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, "topics": [{"id": "https://openalex.org/T10028", "display_name": "Topic Modeling", "score": 0.25220000743865967, "subfield": {"id": "https://openalex.org/subfields/1702", "display_name": "Artificial Intelligence"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, {"id": "https://openalex.org/T13904", "display_name": "Artificial Intelligence Applications", "score": 0.1429000049829483, "subfield": {"id": "https://openalex.org/subfields/1702", "display_name": "Artificial Intelligence"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, {"id": "https://openalex.org/T12761", "display_name": "Data Stream Mining Techniques", "score": 0.11640000343322754, "subfield": {"id": "https://openalex.org/subfields/1702", "display_name": "Artificial Intelligence"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}], "primary_location": {"id": "doi:10.5281/zenodo.17547016", "is_oa": true, "landing_page_url": "https://doi.org/10.5281/zenodo.17547016", "pdf_url": null, "source": {"id": "https://openalex.org/S4306400562", "display_name": "Zenodo (CERN European Organization for Nuclear Research)", "issn_l": null, "issn": null, "is_oa": true, "is_in_doaj": false, "is_core": false, "listed_in": [], "host_organization": "https://openalex.org/I67311998", "host_organization_name": "European Organization for Nuclear Research", "host_organization_lineage": ["https://openalex.org/I67311998"], "host_organization_lineage_names": [], "type": "repository"}, "license": null, "license_id": null, "version": null, "is_accepted": false, "is_published": false, "raw_source_name": null, "raw_type": "article-journal"}, "cited_by_count": 0, "is_retracted": false}, {"id": "https://openalex.org/W4417295988", "doi": "https://doi.org/10.48550/arxiv.2512.10547", "title": "Unlocking the Address Book: Dissecting the Sparse Semantic Structure of LLM Key-Value Caches via Sparse Autoencoders", "publication_year": 2025, "publication_date": "2025-12-11", "type": "preprint", "language": null, "authorships": [{"author_position": "first", "author": {"id": "https://openalex.org/A5084704002", "display_name": "Qingsen Ma", "orcid": "https://orcid.org/0000-0001-8059-0458", "observed_orcids": ["https://orcid.org/0000-0001-8059-0458"]}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Ma, Qingsen", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": null, "display_name": "Wang, Dianyun", "orcid": null, "observed_orcids": []}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Wang, Dianyun", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5046612778", "display_name": "Jiaming Lyu", "orcid": "https://orcid.org/0009-0003-8868-7248", "observed_orcids": ["https://orcid.org/0009-0003-8868-7248"]}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Lyu, Jiaming", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": null, "display_name": "Wang, Yaoye", "orcid": null, "observed_orcids": []}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Wang, Yaoye", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5081265309", "display_name": "Ning Li", "orcid": "https://orcid.org/0000-0002-4534-9009", "observed_orcids": ["https://orcid.org/0000-0002-4534-9009", "https://orcid.org/0009-0005-1124-1672"]}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Ning, Lechen", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": null, "display_name": "Zhu, Sujie", "orcid": null, "observed_orcids": []}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Zhu, Sujie", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5100517463", "display_name": "Zhenbo Xu", "orcid": null, "observed_orcids": []}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Xu, Zhenbo", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5077408167", "display_name": "Liuyu Xiang", "orcid": "https://orcid.org/0000-0001-8486-6255", "observed_orcids": ["https://orcid.org/0000-0001-8486-6255"]}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Xiang, Liuyu", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5055511404", "display_name": "Huining Li", "orcid": "https://orcid.org/0000-0003-3865-3587", "observed_orcids": ["https://orcid.org/0000-0003-3865-3587"]}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Li, Huining", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5019711261", "display_name": "Huijia Wu", "orcid": null, "observed_orcids": []}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Wu, Huijia", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "last", "author": {"id": "https://openalex.org/A5108370761", "display_name": "Zhaofeng He", "orcid": "https://orcid.org/0000-0003-3043-2122", "observed_orcids": ["https://orcid.org/0000-0003-3043-2122"]}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "He, Zhaofeng", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}], "abstract_inverted_index": {"The": [0], "Key-Value": [1], "(KV)": [2], "cache": [3, 42], "is": [4, 16], "the": [5, 40, 58, 108, 136, 141], "primary": [6], "memory": [7], "bottleneck": [8], "in": [9], "long-context": [10], "Large": [11], "Language": [12], "Models,": [13], "yet": [14], "it": [15], "typically": [17], "treated": [18], "as": [19, 76], "an": [20], "opaque": [21], "numerical": [22], "tensor.": [23], "In": [24], "this": [25, 98], "work,": [26], "we": [27, 100], "propose": [28], "\\textbf{STA-Attention},": [29], "a": [30, 68, 82, 93, 102], "framework": [31], "that": [32, 105, 125], "utilizes": [33], "Top-K": [34, 52], "Sparse": [35], "Autoencoders": [36], "(SAEs)": [37], "to": [38, 135], "decompose": [39], "KV": [41], "into": [43], "interpretable": [44], "``semantic": [45], "atoms.''": [46], "Unlike": [47], "standard": [48], "$L_1$-regularized": [49], "SAEs,": [50], "our": [51, 126], "approach": [53], "eliminates": [54], "shrinkage": [55], "bias,": [56], "preserving": [57], "precise": [59], "dot-product": [60], "geometry": [61], "required": [62], "for": [63], "attention.": [64], "Our": [65], "analysis": [66], "uncovers": [67], "fundamental": [69], "\\textbf{Key-Value": [70], "Asymmetry}:": [71], "while": [72, 113], "Key": [73], "vectors": [74, 87], "serve": [75], "highly": [77], "sparse": [78], "routers": [79], "dominated": [80], "by": [81], "``Semantic": [83], "Elbow,''": [84], "deep": [85], "Value": [86], "carry": [88], "dense": [89], "content": [90], "payloads": [91], "requiring": [92], "larger": [94], "budget.": [95], "Based": [96], "on": [97, 118], "structure,": [99], "introduce": [101], "Dual-Budget": [103], "Strategy": [104], "selectively": [106], "preserves": [107], "most": [109], "informative": [110], "semantic": [111, 127], "components": [112], "filtering": [114], "representational": [115], "noise.": [116], "Experiments": [117], "Yi-6B,": [119], "Mistral-7B,": [120], "Qwen2.5-32B,": [121], "and": [122, 131, 146], "others": [123], "show": [124], "reconstructions": [128], "maintain": [129], "perplexity": [130], "zero-shot": [132], "performance": [133], "comparable": [134], "original": [137], "models,": [138], "effectively": [139], "bridging": [140], "gap": [142], "between": [143], "mechanistic": [144], "interpretability": [145], "faithful": [147], "attention": [148], "modeling.": [149]}, "primary_topic": {"id": "https://openalex.org/T12026", "display_name": "Explainable Artificial Intelligence (XAI)", "score": 0.9589999914169312, "subfield": {"id": "https://openalex.org/subfields/1702", "display_name": "Artificial Intelligence"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, "topics": [{"id": "https://openalex.org/T12026", "display_name": "Explainable Artificial Intelligence (XAI)", "score": 0.9589999914169312, "subfield": {"id": "https://openalex.org/subfields/1702", "display_name": "Artificial Intelligence"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, {"id": "https://openalex.org/T10028", "display_name": "Topic Modeling", "score": 0.03539999946951866, "subfield": {"id": "https://openalex.org/subfields/1702", "display_name": "Artificial Intelligence"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, {"id": "https://openalex.org/T10036", "display_name": "Advanced Neural Network Applications", "score": 0.0012000000569969416, "subfield": {"id": "https://openalex.org/subfields/1707", "display_name": "Computer Vision and Pattern Recognition"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}], "primary_location": {"id": "pmh:oai:arXiv.org:2512.10547", "is_oa": true, "landing_page_url": "http://arxiv.org/abs/2512.10547", "pdf_url": "https://arxiv.org/pdf/2512.10547", "source": {"id": "https://openalex.org/S4306400194", "display_name": "arXiv (Cornell University)", "issn_l": "2331-8422", "issn": ["2331-8422"], "is_oa": true, "is_in_doaj": false, "is_core": false, "listed_in": [], "host_organization": "https://openalex.org/I205783295", "host_organization_name": "Cornell University", "host_organization_lineage": ["https://openalex.org/I205783295"], "host_organization_lineage_names": [], "type": "repository"}, "license": "cc-by-nc-sa", "license_id": "https://openalex.org/licenses/cc-by-nc-sa", "version": "submittedVersion", "is_accepted": false, "is_published": false, "raw_source_name": null, "raw_type": "text"}, "cited_by_count": 0, "is_retracted": false}, {"id": "https://openalex.org/W4415238667", "doi": "https://doi.org/10.48550/arxiv.2508.14239", "title": "A Distributed Learned Hash Table", "publication_year": 2025, "publication_date": "2025-08-19", "type": "preprint", "language": "en", "authorships": [{"author_position": "first", "author": {"id": "https://openalex.org/A5053382292", "display_name": "Shengze Wang", "orcid": "https://orcid.org/0000-0001-9496-158X", "observed_orcids": ["https://orcid.org/0000-0001-9496-158X"]}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Wang, Shengze", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5008288843", "display_name": "Yi Liu", "orcid": "https://orcid.org/0000-0002-8839-9508", "observed_orcids": ["https://orcid.org/0000-0002-8839-9508"]}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Liu, Yi", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": null, "display_name": "Zhang, Xiaoxue", "orcid": null, "observed_orcids": []}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Zhang, Xiaoxue", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5113952097", "display_name": "Liting Hu", "orcid": "https://orcid.org/0000-0001-8512-5176", "observed_orcids": ["https://orcid.org/0000-0001-8512-5176", "https://orcid.org/0009-0007-7222-5507"]}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Hu, Liting", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "last", "author": {"id": "https://openalex.org/A5100737734", "display_name": "Chen Qian", "orcid": "https://orcid.org/0000-0002-6882-9590", "observed_orcids": ["https://orcid.org/0000-0002-6882-9590"]}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Qian, Chen", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}], "abstract_inverted_index": {"Distributed": [0], "Hash": [1], "Tables": [2], "(DHTs)": [3], "are": [4, 43], "pivotal": [5], "in": [6, 38, 129, 141, 150, 183], "numerous": [7], "high-impact": [8], "key-value": [9, 185], "applications": [10, 46], "built": [11], "on": [12], "distributed": [13, 51, 95, 152, 184], "networked": [14], "systems,": [15, 153], "offering": [16], "a": [17, 66, 83, 94, 175], "decentralized": [18], "architecture": [19], "that": [20, 136], "avoids": [21], "single": [22], "points": [23], "of": [24, 102], "failure": [25], "and": [26, 57, 90, 113, 120, 133, 157, 169], "improves": [27], "data": [28, 92, 181], "availability.": [29], "Despite": [30], "their": [31], "widespread": [32], "utility,": [33], "DHTs": [34], "face": [35], "substantial": [36], "challenges": [37], "handling": [39], "range": [40, 78, 110, 147], "queries,": [41], "which": [42], "crucial": [44], "for": [45, 179], "such": [47], "as": [48], "LLM": [49], "serving,": [50], "storage,": [52], "databases,": [53], "content": [54], "delivery": [55], "networks,": [56], "blockchains.": [58], "To": [59], "address": [60], "this": [61], "limitation,": [62], "we": [63], "present": [64], "LEAD,": [65], "novel": [67], "system": [68, 96, 142, 172], "incorporating": [69], "learned": [70], "models": [71], "within": [72], "DHT": [73], "structures": [74], "to": [75, 88, 108, 122, 145, 162], "significantly": [76], "optimize": [77], "query": [79, 111, 148, 155], "performance.": [80], "LEAD": [81, 104, 137, 165], "utilizes": [82], "recursive": [84], "machine": [85], "learning": [86], "model": [87], "map": [89], "retrieve": [91], "across": [93], "while": [97, 116], "preserving": [98], "the": [99, 106], "inherent": [100], "order": [101], "data.": [103], "includes": [105], "designs": [107], "minimize": [109], "latency": [112, 156], "message": [114, 158], "cost": [115, 159], "maintaining": [117], "high": [118], "scalability": [119, 168], "resilience": [121], "network": [123], "churn.": [124], "Our": [125], "comprehensive": [126], "evaluations,": [127], "conducted": [128], "both": [130], "testbed": [131], "implementation": [132], "simulations,": [134], "demonstrate": [135], "achieves": [138], "tremendous": [139], "advantages": [140], "efficiency": [143], "compared": [144], "existing": [146], "methods": [149], "large-scale": [151], "reducing": [154], "by": [160], "80%": [161], "90%+.": [163], "Furthermore,": [164], "exhibits": [166], "remarkable": [167], "robustness": [170], "against": [171], "churn,": [173], "providing": [174], "robust,": [176], "scalable": [177], "solution": [178], "efficient": [180], "retrieval": [182], "systems.": [186]}, "primary_topic": {"id": "https://openalex.org/T10742", "display_name": "Peer-to-Peer Network Technologies", "score": 0.9945999979972839, "subfield": {"id": "https://openalex.org/subfields/1705", "display_name": "Computer Networks and Communications"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, "topics": [{"id": "https://openalex.org/T10742", "display_name": "Peer-to-Peer Network Technologies", "score": 0.9945999979972839, "subfield": {"id": "https://openalex.org/subfields/1705", "display_name": "Computer Networks and Communications"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, {"id": "https://openalex.org/T10317", "display_name": "Advanced Database Systems and Queries", "score": 0.0032999999821186066, "subfield": {"id": "https://openalex.org/subfields/1705", "display_name": "Computer Networks and Communications"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, {"id": "https://openalex.org/T10101", "display_name": "Cloud Computing and Resource Management", "score": 0.00039999998989515007, "subfield": {"id": "https://openalex.org/subfields/1710", "display_name": "Information Systems"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}], "primary_location": {"id": "pmh:oai:arXiv.org:2508.14239", "is_oa": true, "landing_page_url": "http://arxiv.org/abs/2508.14239", "pdf_url": "https://arxiv.org/pdf/2508.14239", "source": {"id": "https://openalex.org/S4306400194", "display_name": "arXiv (Cornell University)", "issn_l": "2331-8422", "issn": ["2331-8422"], "is_oa": true, "is_in_doaj": false, "is_core": false, "listed_in": [], "host_organization": "https://openalex.org/I205783295", "host_organization_name": "Cornell University", "host_organization_lineage": ["https://openalex.org/I205783295"], "host_organization_lineage_names": [], "type": "repository"}, "license": null, "license_id": null, "version": "submittedVersion", "is_accepted": false, "is_published": false, "raw_source_name": null, "raw_type": "text"}, "cited_by_count": 0, "is_retracted": false}, {"id": "https://openalex.org/W4416019938", "doi": "https://doi.org/10.48550/arxiv.2511.03475", "title": "ContextPilot: Fast Long-Context Inference via Context Reuse", "publication_year": 2025, "publication_date": "2025-11-05", "type": "preprint", "language": "en", "authorships": [{"author_position": "first", "author": {"id": "https://openalex.org/A5049301642", "display_name": "Yu Jiang", "orcid": "https://orcid.org/0000-0001-9461-633X", "observed_orcids": ["https://orcid.org/0000-0001-9461-633X"]}, "institutions": [{"id": "https://openalex.org/I98677209", "display_name": "University of Edinburgh", "ror": "https://ror.org/01nrxwf90", "country_code": "GB", "type": "education", "lineage": ["https://openalex.org/I98677209"]}], "countries": ["GB"], "is_corresponding": false, "raw_author_name": "Jiang, Yinsicheng", "raw_affiliation_strings": ["University of Edinburgh , United Kingdom"], "raw_orcid": null, "affiliations": [{"raw_affiliation_string": "University of Edinburgh , United Kingdom", "institution_ids": ["https://openalex.org/I98677209"]}]}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5111780349", "display_name": "Yeqi Huang", "orcid": "https://orcid.org/0009-0004-3526-912X", "observed_orcids": ["https://orcid.org/0009-0004-3526-912X"]}, "institutions": [{"id": "https://openalex.org/I98677209", "display_name": "University of Edinburgh", "ror": "https://ror.org/01nrxwf90", "country_code": "GB", "type": "education", "lineage": ["https://openalex.org/I98677209"]}], "countries": ["GB"], "is_corresponding": false, "raw_author_name": "Huang, Yeqi", "raw_affiliation_strings": ["University of Edinburgh , United Kingdom"], "raw_orcid": null, "affiliations": [{"raw_affiliation_string": "University of Edinburgh , United Kingdom", "institution_ids": ["https://openalex.org/I98677209"]}]}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5082631308", "display_name": "Liang Cheng", "orcid": "https://orcid.org/0000-0001-7427-2939", "observed_orcids": ["https://orcid.org/0000-0001-7427-2939"]}, "institutions": [{"id": "https://openalex.org/I98677209", "display_name": "University of Edinburgh", "ror": "https://ror.org/01nrxwf90", "country_code": "GB", "type": "education", "lineage": ["https://openalex.org/I98677209"]}], "countries": ["GB"], "is_corresponding": false, "raw_author_name": "Cheng, Liang", "raw_affiliation_strings": ["University of Edinburgh , United Kingdom"], "raw_orcid": null, "affiliations": [{"raw_affiliation_string": "University of Edinburgh , United Kingdom", "institution_ids": ["https://openalex.org/I98677209"]}]}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5015874725", "display_name": "Cheng Dan Deng", "orcid": "https://orcid.org/0000-0003-2620-3247", "observed_orcids": ["https://orcid.org/0000-0003-2620-3247"]}, "institutions": [{"id": "https://openalex.org/I98677209", "display_name": "University of Edinburgh", "ror": "https://ror.org/01nrxwf90", "country_code": "GB", "type": "education", "lineage": ["https://openalex.org/I98677209"]}], "countries": ["GB"], "is_corresponding": false, "raw_author_name": "Deng, Cheng", "raw_affiliation_strings": ["University of Edinburgh , United Kingdom"], "raw_orcid": null, "affiliations": [{"raw_affiliation_string": "University of Edinburgh , United Kingdom", "institution_ids": ["https://openalex.org/I98677209"]}]}, {"author_position": "middle", "author": {"id": null, "display_name": "Sun, Xuan", "orcid": null, "observed_orcids": []}, "institutions": [{"id": "https://openalex.org/I98677209", "display_name": "University of Edinburgh", "ror": "https://ror.org/01nrxwf90", "country_code": "GB", "type": "education", "lineage": ["https://openalex.org/I98677209"]}], "countries": ["GB"], "is_corresponding": false, "raw_author_name": "Sun, Xuan", "raw_affiliation_strings": ["University of Edinburgh , United Kingdom"], "raw_orcid": null, "affiliations": [{"raw_affiliation_string": "University of Edinburgh , United Kingdom", "institution_ids": ["https://openalex.org/I98677209"]}]}, {"author_position": "last", "author": {"id": "https://openalex.org/A5034207820", "display_name": "Luo Mai", "orcid": "https://orcid.org/0000-0002-3594-1092", "observed_orcids": ["https://orcid.org/0000-0002-3594-1092"]}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Mai, Luo", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}], "abstract_inverted_index": {"AI": [0], "applications": [1], "increasingly": [2], "depend": [3], "on": [4], "long-context": [5, 84], "inference,": [6], "where": [7], "LLMs": [8], "consume": [9], "substantial": [10], "context": [11, 76, 89, 94, 107, 125, 172], "to": [12, 91, 112, 160, 163], "support": [13], "stronger": [14], "reasoning.": [15], "Common": [16], "examples": [17], "include": [18], "retrieval-augmented": [19], "generation,": [20], "agent": [21], "memory": [22], "layers,": [23], "and": [24, 102, 109], "multi-agent": [25], "orchestration.": [26], "As": [27], "input": [28], "contexts": [29], "get": [30], "longer,": [31], "prefill": [32, 40, 73, 156], "latency": [33, 157], "becomes": [34], "the": [35, 60], "main": [36], "bottleneck.": [37], "Yet": [38], "today's": [39], "acceleration": [41], "techniques": [42, 111], "face": [43], "a": [44, 69, 79, 88, 136, 140], "trade-off:": [45], "they": [46], "either": [47], "preserve": [48, 117], "reasoning": [49, 64, 118, 168, 178], "quality": [50, 119, 129], "but": [51], "deliver": [52], "little": [53], "KV-cache": [54, 114], "reuse,": [55, 121], "or": [56], "improve": [57, 177], "reuse": [58, 77], "at": [59], "cost": [61], "of": [62], "degraded": [63], "quality.": [65, 169, 179], "We": [66], "present": [67], "ContextPilot,": [68], "system": [70], "that": [71, 127, 143, 152], "accelerates": [72], "by": [74, 158], "introducing": [75], "as": [78], "new": [80], "mechanism": [81], "for": [82], "faster": [83], "inference.": [85], "ContextPilot": [86, 132, 153, 180], "introduces": [87, 123], "index": [90], "identify": [92], "overlapping": [93], "blocks": [95], "across": [96, 100], "LLM": [97, 155], "interactions": [98], "(e.g.,": [99], "users": [101], "turns).": [103], "It": [104], "further": [105], "proposes": [106], "ordering": [108], "de-duplication": [110], "maximize": [113], "reuse.": [115], "To": [116], "under": [120], "it": [122, 174], "succinct": [124], "annotations": [126], "prevent": [128], "degradation.": [130], "Finally,": [131], "is": [133, 181], "built": [134], "around": [135], "modular": [137], "architecture": [138], "with": [139, 145], "clean": [141], "interface": [142], "integrates": [144], "existing": [146], "inference": [147], "engines.": [148], "Extensive": [149], "evaluation": [150], "shows": [151], "reduces": [154], "up": [159], "$3\\times{}$": [161], "compared": [162], "state-of-the-art": [164], "methods": [165], "while": [166], "preserving": [167], "At": [170], "longer": [171], "lengths,": [173], "can": [175], "even": [176], "open-sourced": [182], "at:": [183], "https://github.com/EfficientContext/ContextPilot.": [184]}, "primary_topic": {"id": "https://openalex.org/T10028", "display_name": "Topic Modeling", "score": 0.43380001187324524, "subfield": {"id": "https://openalex.org/subfields/1702", "display_name": "Artificial Intelligence"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, "topics": [{"id": "https://openalex.org/T10028", "display_name": "Topic Modeling", "score": 0.43380001187324524, "subfield": {"id": "https://openalex.org/subfields/1702", "display_name": "Artificial Intelligence"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, {"id": "https://openalex.org/T10181", "display_name": "Natural Language Processing Techniques", "score": 0.27320000529289246, "subfield": {"id": "https://openalex.org/subfields/1702", "display_name": "Artificial Intelligence"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, {"id": "https://openalex.org/T10054", "display_name": "Parallel Computing and Optimization Techniques", "score": 0.08049999922513962, "subfield": {"id": "https://openalex.org/subfields/1708", "display_name": "Hardware and Architecture"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}], "primary_location": {"id": "pmh:oai:arXiv.org:2511.03475", "is_oa": true, "landing_page_url": "https://arxiv.org/abs/2511.03475", "pdf_url": "https://arxiv.org/pdf/2511.03475", "source": {"id": "https://openalex.org/S4306400194", "display_name": "arXiv (Cornell University)", "issn_l": "2331-8422", "issn": ["2331-8422"], "is_oa": true, "is_in_doaj": false, "is_core": false, "listed_in": [], "host_organization": "https://openalex.org/I205783295", "host_organization_name": "Cornell University", "host_organization_lineage": ["https://openalex.org/I205783295"], "host_organization_lineage_names": [], "type": "repository"}, "license": "cc-by-sa", "license_id": "https://openalex.org/licenses/cc-by-sa", "version": "submittedVersion", "is_accepted": false, "is_published": false, "raw_source_name": null, "raw_type": "text"}, "cited_by_count": 0, "is_retracted": false}, {"id": "https://openalex.org/W4416887488", "doi": "https://doi.org/10.48550/arxiv.2507.00605", "title": "Quantize-Sample-and-Verify: LLM Acceleration via Adaptive Edge-Cloud Speculative Decoding", "publication_year": 2025, "publication_date": "2025-07-01", "type": "preprint", "language": "en", "authorships": [{"author_position": "first", "author": {"id": "https://openalex.org/A5101487987", "display_name": "Guangyi Zhang", "orcid": "https://orcid.org/0000-0002-1252-7489", "observed_orcids": ["https://orcid.org/0000-0002-1252-7489", "https://orcid.org/0000-0002-4476-5633"]}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Zhang, Guangyi", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5034376590", "display_name": "Yunlong Cai", "orcid": "https://orcid.org/0000-0001-9418-1700", "observed_orcids": ["https://orcid.org/0000-0001-9418-1700"]}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Cai, Yunlong", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5079187892", "display_name": "Guanding Yu", "orcid": "https://orcid.org/0000-0003-1354-2557", "observed_orcids": ["https://orcid.org/0000-0003-1354-2557", "https://orcid.org/0000-0001-7296-1490"]}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Yu, Guanding", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5071289803", "display_name": "Petar Popovski", "orcid": "https://orcid.org/0000-0001-6195-4797", "observed_orcids": ["https://orcid.org/0000-0001-6195-4797"]}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Popovski, Petar", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "last", "author": {"id": "https://openalex.org/A5017736224", "display_name": "Osvaldo Simeone", "orcid": "https://orcid.org/0000-0001-9898-3209", "observed_orcids": ["https://orcid.org/0000-0001-9898-3209"]}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Simeone, Osvaldo", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}], "abstract_inverted_index": {"In": [0, 52], "edge-cloud": [1, 97, 148], "speculative": [2], "decoding": [3, 144], "(SD),": [4], "edge": [5, 39], "devices": [6], "equipped": [7], "with": [8], "small": [9], "language": [10, 21], "models": [11, 22], "(SLMs)": [12], "generate": [13], "draft": [14, 121], "tokens": [15, 76], "that": [16, 62, 73, 82, 99, 113, 137], "are": [17], "verified": [18, 75], "by": [19, 88, 117], "large": [20], "(LLMs)": [23], "in": [24, 30, 126, 146], "the": [25, 34, 46, 65, 69, 74, 78, 89, 120, 138], "cloud.": [26], "A": [27], "key": [28], "bottleneck": [29], "such": [31], "systems": [32], "is": [33], "limited": [35], "communication": [36, 103], "bandwidth": [37], "between": [38], "and": [40, 123, 132], "cloud,": [41], "which": [42], "necessitates": [43], "quantization": [44, 124], "of": [45, 68, 80], "information": [47], "transmitted": [48], "about": [49], "generated": [50, 86], "tokens.": [51], "this": [53, 106], "work,": [54], "we": [55, 108], "introduce": [56], "a": [57, 93], "novel": [58], "quantize-sample": [59], "(Q-S)": [60], "strategy": [61], "provably": [63], "preserves": [64], "output": [66], "distribution": [67, 79], "cloud-based": [70], "model,": [71, 107], "ensuring": [72], "match": [77], "those": [81], "would": [83], "have": [84], "been": [85], "directly": [87], "LLM.": [90], "We": [91], "develop": [92], "throughput": [94, 116], "model": [95], "for": [96, 102], "SD": [98], "explicitly": [100], "accounts": [101], "latency.": [104], "Leveraging": [105], "propose": [109], "an": [110], "adaptive": [111], "mechanism": [112], "optimizes": [114], "token": [115], "dynamically": [118], "adjusting": [119], "length": [122], "precision": [125], "response": [127], "to": [128], "both": [129], "semantic": [130], "uncertainty": [131], "channel": [133], "conditions.": [134], "Simulations": [135], "demonstrate": [136], "proposed": [139], "Q-S": [140], "approach": [141], "significantly": [142], "improves": [143], "efficiency": [145], "realistic": [147], "deployment": [149], "scenarios.": [150]}, "primary_topic": {"id": "https://openalex.org/T10273", "display_name": "IoT and Edge/Fog Computing", "score": 0.9106000065803528, "subfield": {"id": "https://openalex.org/subfields/1705", "display_name": "Computer Networks and Communications"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, "topics": [{"id": "https://openalex.org/T10273", "display_name": "IoT and Edge/Fog Computing", "score": 0.9106000065803528, "subfield": {"id": "https://openalex.org/subfields/1705", "display_name": "Computer Networks and Communications"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, {"id": "https://openalex.org/T10181", "display_name": "Natural Language Processing Techniques", "score": 0.02459999918937683, "subfield": {"id": "https://openalex.org/subfields/1702", "display_name": "Artificial Intelligence"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, {"id": "https://openalex.org/T10028", "display_name": "Topic Modeling", "score": 0.008299999870359898, "subfield": {"id": "https://openalex.org/subfields/1702", "display_name": "Artificial Intelligence"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}], "primary_location": {"id": "pmh:oai:arXiv.org:2507.00605", "is_oa": true, "landing_page_url": "http://arxiv.org/abs/2507.00605", "pdf_url": "https://arxiv.org/pdf/2507.00605", "source": {"id": "https://openalex.org/S4306400194", "display_name": "arXiv (Cornell University)", "issn_l": "2331-8422", "issn": ["2331-8422"], "is_oa": true, "is_in_doaj": false, "is_core": false, "listed_in": [], "host_organization": "https://openalex.org/I205783295", "host_organization_name": "Cornell University", "host_organization_lineage": ["https://openalex.org/I205783295"], "host_organization_lineage_names": [], "type": "repository"}, "license": null, "license_id": null, "version": "submittedVersion", "is_accepted": false, "is_published": false, "raw_source_name": null, "raw_type": "text"}, "cited_by_count": 0, "is_retracted": false}, {"id": "https://openalex.org/W4414815928", "doi": "https://doi.org/10.48550/arxiv.2510.01336", "title": "HiSpec: Hierarchical Speculative Decoding for LLMs", "publication_year": 2025, "publication_date": "2025-10-01", "type": "preprint", "language": "en", "authorships": [{"author_position": "first", "author": {"id": null, "display_name": "Kumar, Avinash", "orcid": null, "observed_orcids": []}, "institutions": [{"id": "https://openalex.org/I86519309", "display_name": "The University of Texas at Austin", "ror": "https://ror.org/00hj54h04", "country_code": "US", "type": "education", "lineage": ["https://openalex.org/I86519309"]}], "countries": ["US"], "is_corresponding": false, "raw_author_name": "Kumar, Avinash", "raw_affiliation_strings": ["Department of Electrical and Computer Engineering , The University of Texas at Austin"], "raw_orcid": null, "affiliations": [{"raw_affiliation_string": "Department of Electrical and Computer Engineering , The University of Texas at Austin", "institution_ids": ["https://openalex.org/I86519309"]}]}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5110619770", "display_name": "Sujay Sanghavi", "orcid": null, "observed_orcids": []}, "institutions": [{"id": "https://openalex.org/I86519309", "display_name": "The University of Texas at Austin", "ror": "https://ror.org/00hj54h04", "country_code": "US", "type": "education", "lineage": ["https://openalex.org/I86519309"]}], "countries": ["US"], "is_corresponding": false, "raw_author_name": "Sanghavi, Sujay", "raw_affiliation_strings": ["Department of Electrical and Computer Engineering , The University of Texas at Austin"], "raw_orcid": null, "affiliations": [{"raw_affiliation_string": "Department of Electrical and Computer Engineering , The University of Texas at Austin", "institution_ids": ["https://openalex.org/I86519309"]}]}, {"author_position": "last", "author": {"id": "https://openalex.org/A5039719490", "display_name": "Poulami Das", "orcid": "https://orcid.org/0000-0002-5811-6108", "observed_orcids": ["https://orcid.org/0000-0002-5811-6108"]}, "institutions": [{"id": "https://openalex.org/I86519309", "display_name": "The University of Texas at Austin", "ror": "https://ror.org/00hj54h04", "country_code": "US", "type": "education", "lineage": ["https://openalex.org/I86519309"]}], "countries": ["US"], "is_corresponding": false, "raw_author_name": "Das, Poulami", "raw_affiliation_strings": ["Department of Electrical and Computer Engineering , The University of Texas at Austin"], "raw_orcid": null, "affiliations": [{"raw_affiliation_string": "Department of Electrical and Computer Engineering , The University of Texas at Austin", "institution_ids": ["https://openalex.org/I86519309"]}]}], "abstract_inverted_index": {"Speculative": [0], "decoding": [1, 104], "accelerates": [2], "LLM": [3], "inference": [4], "by": [5, 57, 88, 121, 190, 211, 216], "using": [6, 200], "a": [7, 15, 34, 39, 99, 160], "smaller": [8], "draft": [9, 60, 187], "model": [10, 18, 36], "to": [11, 79, 118, 165, 218, 221], "speculate": [12], "tokens": [13, 61, 117, 188], "that": [14, 105, 130, 162, 207], "larger": [16], "target": [17, 41, 178, 196], "verifies.": [19], "Verification": [20], "is": [21, 27], "often": [22], "the": [23, 72, 76, 81, 173, 186, 191, 195, 222], "bottleneck": [24], "(e.g.": [25], "verification": [26, 53, 55, 83, 145], "$4\\times$": [28], "slower": [29], "than": [30], "token": [31], "generation": [32], "when": [33], "3B": [35], "speculates": [37], "for": [38, 101, 110, 143], "70B": [40], "model),": [42], "but": [43, 63], "most": [44], "prior": [45], "works": [46], "focus": [47], "only": [48], "on": [49, 90, 213], "accelerating": [50], "drafting.": [51], "$\\textit{``Intermediate\"}$": [52], "reduces": [54], "time": [56], "discarding": [58], "inaccurate": [59], "early,": [62], "existing": [64], "methods": [65], "incur": [66], "substantial": [67], "training": [68], "overheads": [69], "in": [70], "incorporating": [71], "intermediate": [73, 82, 112, 144, 175, 192], "verifier,": [74, 176], "increase": [75], "memory": [77, 151], "footprint": [78], "orchestrate": [80], "step,": [84], "and": [85, 125, 150, 169, 177, 204, 215], "compromise": [86], "accuracy": [87], "relying": [89], "approximate": [91], "heuristics.": [92], "We": [93], "propose": [94], "$\\underline{\\textit{Hi}}\\textit{erarchical": [95], "}\\underline{\\textit{Spec}}\\textit{ulative": [96], "Decoding": [97], "(HiSpec)}$,": [98], "framework": [100], "high-throughput": [102], "speculative": [103], "exploits": [106], "$\\textit{early-exit": [107], "(EE)": [108], "models}$": [109], "low-overhead": [111], "verification.": [113], "EE": [114], "models": [115, 205], "allow": [116], "exit": [119], "early": [120], "skipping": [122], "layer": [123], "traversal": [124], "are": [126], "explicitly": [127], "trained": [128], "so": [129], "hidden": [131, 170], "states": [132, 171], "at": [133], "selected": [134], "layers": [135], "can": [136], "be": [137], "interpreted,": [138], "making": [139], "them": [140], "uniquely": [141], "suited": [142], "without": [146, 226], "drastically": [147], "increasing": [148], "compute": [149], "overheads.": [152], "To": [153, 180], "improve": [154], "resource-efficiency": [155], "even": [156], "further,": [157], "we": [158], "design": [159], "methodology": [161], "enables": [163], "HiSpec": [164, 183, 208], "re-use": [166], "key-value": [167], "caches": [168], "between": [172], "draft,": [174], "models.": [179], "maintain": [181], "accuracy,": [182], "periodically": [184], "validates": [185], "accepted": [189], "verifier": [193], "against": [194], "model.": [197], "Our": [198], "evaluations": [199], "various": [201], "representative": [202], "benchmarks": [203], "show": [206], "improves": [209], "throughput": [210], "1.28$\\times$": [212], "average": [214], "up": [217], "2.01$\\times$": [219], "compared": [220], "baseline": [223], "single-layer": [224], "speculation": [225], "compromising": [227], "accuracy.": [228]}, "primary_topic": {"id": "https://openalex.org/T10181", "display_name": "Natural Language Processing Techniques", "score": 0.7182999849319458, "subfield": {"id": "https://openalex.org/subfields/1702", "display_name": "Artificial Intelligence"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, "topics": [{"id": "https://openalex.org/T10181", "display_name": "Natural Language Processing Techniques", "score": 0.7182999849319458, "subfield": {"id": "https://openalex.org/subfields/1702", "display_name": "Artificial Intelligence"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, {"id": "https://openalex.org/T10028", "display_name": "Topic Modeling", "score": 0.1776999980211258, "subfield": {"id": "https://openalex.org/subfields/1702", "display_name": "Artificial Intelligence"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, {"id": "https://openalex.org/T10036", "display_name": "Advanced Neural Network Applications", "score": 0.040300000458955765, "subfield": {"id": "https://openalex.org/subfields/1707", "display_name": "Computer Vision and Pattern Recognition"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}], "primary_location": {"id": "pmh:oai:arXiv.org:2510.01336", "is_oa": true, "landing_page_url": "https://arxiv.org/abs/2510.01336", "pdf_url": "https://arxiv.org/pdf/2510.01336", "source": {"id": "https://openalex.org/S4306400194", "display_name": "arXiv (Cornell University)", "issn_l": "2331-8422", "issn": ["2331-8422"], "is_oa": true, "is_in_doaj": false, "is_core": false, "listed_in": [], "host_organization": "https://openalex.org/I205783295", "host_organization_name": "Cornell University", "host_organization_lineage": ["https://openalex.org/I205783295"], "host_organization_lineage_names": [], "type": "repository"}, "license": "cc-by", "license_id": "https://openalex.org/licenses/cc-by", "version": "submittedVersion", "is_accepted": false, "is_published": false, "raw_source_name": null, "raw_type": "text"}, "cited_by_count": 0, "is_retracted": false}, {"id": "https://openalex.org/W7104401102", "doi": "https://doi.org/10.5281/zenodo.17547015", "title": "AI Sleep Constructs: Implementing Chronological Intelligence in Stateful Systems", "publication_year": 2025, "publication_date": "2025-11-07", "type": "article", "language": "en", "authorships": [{"author_position": "first", "author": {"id": null, "display_name": "Smith, Tionne", "orcid": "https://orcid.org/0009-0008-8627-6150", "observed_orcids": ["https://orcid.org/0009-0008-8627-6150"]}, "institutions": [], "countries": [], "is_corresponding": true, "raw_author_name": "Smith, Tionne", "raw_affiliation_strings": ["Antiparty Press"], "raw_orcid": "https://orcid.org/0009-0008-8627-6150", "affiliations": [{"raw_affiliation_string": "Antiparty Press", "institution_ids": []}]}], "abstract_inverted_index": {"AI": [0, 123, 130], "Sleep": [1], "Constructs": [2], "is": [3], "a": [4, 89], "production-ready": [5], "Python": [6], "framework": [7, 83], "that": [8, 30], "brings": [9], "temporal": [10], "awareness": [11], "and": [12, 26, 39, 52, 64], "self-maintenance": [13], "to": [14, 34, 99], "artificial": [15], "intelligence": [16], "models.": [17], "Inspired": [18], "by": [19], "biological": [20], "sleep,": [21, 131], "it": [22, 112], "introduces": [23], "engineered": [24], "light": [25], "deep": [27, 56], "sleep": [28, 45, 57], "cycles": [29], "enable": [31], "language": [32], "models": [33, 96], "optimize": [35], "performance,": [36], "consolidate": [37], "memory,": [38], "self-heal": [40], "between": [41], "inference": [42, 102], "sessions.": [43], "Light": [44], "modes": [46, 58], "perform": [47], "gradient": [48], "clipping,": [49], "cache": [50], "management,": [51], "adaptive": [53], "learning,": [54], "while": [55], "execute": [59], "semantic": [60], "consolidation,": [61], "layer": [62], "recalibration,": [63], "diagnostic": [65], "repair.": [66], "Contextual": [67], "triggers": [68], "automatically": [69], "govern": [70], "these": [71], "transitions": [72], "based": [73], "on": [74, 119], "model": [75], "drift,": [76], "performance": [77], "anomalies,": [78], "or": [79], "workload": [80], "scheduling.": [81], "The": [82], "also": [84], "includes": [85], "Cache-to-Cache": [86], "(C2C)": [87], "fusion,": [88], "novel": [90], "technique": [91], "allowing": [92], "KV-cache": [93], "transfer": [94], "across": [95], "for": [97, 117], "up": [98], "10\u00d7": [100], "faster": [101], "with": [103, 108], "minimal": [104], "quality": [105], "loss.": [106], "Integrated": [107], "Hugging": [109], "Face": [110], "APIs,": [111], "provides": [113], "an": [114], "extensible": [115], "foundation": [116], "research": [118], "stateful,": [120], "chronologically": [121], "intelligent": [122], "architectures.": [124], "Repository": [125], "URL": [126], "(Git):": [127], "https://github.com/electricwolfemarshmallowhypertext/ai-sleep/": [128], "Keywords:": [129], "Chronological": [132], "Intelligence,": [133], "Offline": [134], "Optimization,": [135], "Adaptive": [136], "Learning,": [137], "C2C": [138], "Fusion,": [139], "Stateful": [140], "AI,": [141], "Antiparty": [142], "Inc.": [143]}, "primary_topic": {"id": "https://openalex.org/T10028", "display_name": "Topic Modeling", "score": 0.25220000743865967, "subfield": {"id": "https://openalex.org/subfields/1702", "display_name": "Artificial Intelligence"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, "topics": [{"id": "https://openalex.org/T10028", "display_name": "Topic Modeling", "score": 0.25220000743865967, "subfield": {"id": "https://openalex.org/subfields/1702", "display_name": "Artificial Intelligence"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, {"id": "https://openalex.org/T13904", "display_name": "Artificial Intelligence Applications", "score": 0.1429000049829483, "subfield": {"id": "https://openalex.org/subfields/1702", "display_name": "Artificial Intelligence"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, {"id": "https://openalex.org/T12761", "display_name": "Data Stream Mining Techniques", "score": 0.11640000343322754, "subfield": {"id": "https://openalex.org/subfields/1702", "display_name": "Artificial Intelligence"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}], "primary_location": {"id": "doi:10.5281/zenodo.17547015", "is_oa": true, "landing_page_url": "https://doi.org/10.5281/zenodo.17547015", "pdf_url": null, "source": {"id": "https://openalex.org/S4306400562", "display_name": "Zenodo (CERN European Organization for Nuclear Research)", "issn_l": null, "issn": null, "is_oa": true, "is_in_doaj": false, "is_core": false, "listed_in": [], "host_organization": "https://openalex.org/I67311998", "host_organization_name": "European Organization for Nuclear Research", "host_organization_lineage": ["https://openalex.org/I67311998"], "host_organization_lineage_names": [], "type": "repository"}, "license": null, "license_id": null, "version": null, "is_accepted": false, "is_published": false, "raw_source_name": null, "raw_type": "article-journal"}, "cited_by_count": 0, "is_retracted": false}, {"id": "https://openalex.org/W4408947155", "doi": "https://doi.org/10.1109/tmc.2025.3555298", "title": "Security Enhanced Computation Offloading for Collaborative Inference at Semantic-Communication-Empowered Edge", "publication_year": 2025, "publication_date": "2025-03-28", "type": "article", "language": "en", "authorships": [{"author_position": "first", "author": {"id": "https://openalex.org/A5080067958", "display_name": "Jincheng Peng", "orcid": "https://orcid.org/0009-0007-2262-6580", "observed_orcids": ["https://orcid.org/0009-0007-2262-6580"]}, "institutions": [{"id": "https://openalex.org/I4800084", "display_name": "Southwest Jiaotong University", "ror": "https://ror.org/00hn7w693", "country_code": "CN", "type": "education", "lineage": ["https://openalex.org/I4800084"]}], "countries": ["CN"], "is_corresponding": false, "raw_author_name": "Jincheng Peng", "raw_affiliation_strings": ["School of Computing and Artificial Intelligence, Southwest Jiaotong University, Chengdu, China"], "raw_orcid": "https://orcid.org/0009-0007-2262-6580", "affiliations": [{"raw_affiliation_string": "School of Computing and Artificial Intelligence, Southwest Jiaotong University, Chengdu, China", "institution_ids": ["https://openalex.org/I4800084"]}]}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5007865788", "display_name": "Huanlai Xing", "orcid": "https://orcid.org/0000-0002-6345-7265", "observed_orcids": ["https://orcid.org/0000-0002-6345-7265"]}, "institutions": [{"id": "https://openalex.org/I4800084", "display_name": "Southwest Jiaotong University", "ror": "https://ror.org/00hn7w693", "country_code": "CN", "type": "education", "lineage": ["https://openalex.org/I4800084"]}], "countries": ["CN"], "is_corresponding": false, "raw_author_name": "Huanlai Xing", "raw_affiliation_strings": ["School of Computing and Artificial Intelligence, Southwest Jiaotong University, Chengdu, China"], "raw_orcid": "https://orcid.org/0000-0002-6345-7265", "affiliations": [{"raw_affiliation_string": "School of Computing and Artificial Intelligence, Southwest Jiaotong University, Chengdu, China", "institution_ids": ["https://openalex.org/I4800084"]}]}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5049896980", "display_name": "Xiangyi Chen", "orcid": "https://orcid.org/0000-0001-5549-5336", "observed_orcids": ["https://orcid.org/0000-0001-5549-5336"]}, "institutions": [{"id": "https://openalex.org/I4800084", "display_name": "Southwest Jiaotong University", "ror": "https://ror.org/00hn7w693", "country_code": "CN", "type": "education", "lineage": ["https://openalex.org/I4800084"]}], "countries": ["CN"], "is_corresponding": false, "raw_author_name": "Xiangyi Chen", "raw_affiliation_strings": ["School of Computing and Artificial Intelligence, Southwest Jiaotong University, Chengdu, China"], "raw_orcid": "https://orcid.org/0000-0001-5549-5336", "affiliations": [{"raw_affiliation_string": "School of Computing and Artificial Intelligence, Southwest Jiaotong University, Chengdu, China", "institution_ids": ["https://openalex.org/I4800084"]}]}, {"author_position": "middle", "author": {"id": null, "display_name": "Yang Li", "orcid": "https://orcid.org/0000-0003-3720-7729", "observed_orcids": ["https://orcid.org/0000-0003-3720-7729"]}, "institutions": [{"id": "https://openalex.org/I4800084", "display_name": "Southwest Jiaotong University", "ror": "https://ror.org/00hn7w693", "country_code": "CN", "type": "education", "lineage": ["https://openalex.org/I4800084"]}], "countries": ["CN"], "is_corresponding": false, "raw_author_name": "Yang Li", "raw_affiliation_strings": ["School of Computing and Artificial Intelligence, Southwest Jiaotong University, Chengdu, China"], "raw_orcid": "https://orcid.org/0000-0003-3720-7729", "affiliations": [{"raw_affiliation_string": "School of Computing and Artificial Intelligence, Southwest Jiaotong University, Chengdu, China", "institution_ids": ["https://openalex.org/I4800084"]}]}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5084299004", "display_name": "Yunhe Cui", "orcid": "https://orcid.org/0000-0002-0880-675X", "observed_orcids": ["https://orcid.org/0000-0002-0880-675X"]}, "institutions": [{"id": "https://openalex.org/I178232147", "display_name": "Guizhou University", "ror": "https://ror.org/02wmsc916", "country_code": "CN", "type": "education", "lineage": ["https://openalex.org/I178232147"]}], "countries": ["CN"], "is_corresponding": false, "raw_author_name": "Yunhe Cui", "raw_affiliation_strings": ["School of Computer Science and Technology, Guizhou University, Guiyang, China"], "raw_orcid": "https://orcid.org/0000-0002-0880-675X", "affiliations": [{"raw_affiliation_string": "School of Computer Science and Technology, Guizhou University, Guiyang, China", "institution_ids": ["https://openalex.org/I178232147"]}]}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5079356223", "display_name": "Danyang Zheng", "orcid": "https://orcid.org/0000-0002-3031-7856", "observed_orcids": ["https://orcid.org/0000-0002-3031-7856"]}, "institutions": [{"id": "https://openalex.org/I4800084", "display_name": "Southwest Jiaotong University", "ror": "https://ror.org/00hn7w693", "country_code": "CN", "type": "education", "lineage": ["https://openalex.org/I4800084"]}], "countries": ["CN"], "is_corresponding": false, "raw_author_name": "Danyang Zheng", "raw_affiliation_strings": ["School of Computing and Artificial Intelligence, Southwest Jiaotong University, Chengdu, China"], "raw_orcid": "https://orcid.org/0000-0002-3031-7856", "affiliations": [{"raw_affiliation_string": "School of Computing and Artificial Intelligence, Southwest Jiaotong University, Chengdu, China", "institution_ids": ["https://openalex.org/I4800084"]}]}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5063529604", "display_name": "Laha Ale", "orcid": "https://orcid.org/0000-0002-4070-5289", "observed_orcids": ["https://orcid.org/0000-0002-4070-5289"]}, "institutions": [{"id": "https://openalex.org/I4800084", "display_name": "Southwest Jiaotong University", "ror": "https://ror.org/00hn7w693", "country_code": "CN", "type": "education", "lineage": ["https://openalex.org/I4800084"]}], "countries": ["CN"], "is_corresponding": false, "raw_author_name": "Laha Ale", "raw_affiliation_strings": ["School of Computing and Artificial Intelligence, Southwest Jiaotong University, Chengdu, China"], "raw_orcid": "https://orcid.org/0000-0002-4070-5289", "affiliations": [{"raw_affiliation_string": "School of Computing and Artificial Intelligence, Southwest Jiaotong University, Chengdu, China", "institution_ids": ["https://openalex.org/I4800084"]}]}, {"author_position": "last", "author": {"id": null, "display_name": "Li Feng", "orcid": "https://orcid.org/0009-0008-2212-9768", "observed_orcids": ["https://orcid.org/0009-0008-2212-9768"]}, "institutions": [{"id": "https://openalex.org/I4800084", "display_name": "Southwest Jiaotong University", "ror": "https://ror.org/00hn7w693", "country_code": "CN", "type": "education", "lineage": ["https://openalex.org/I4800084"]}], "countries": ["CN"], "is_corresponding": false, "raw_author_name": "Li Feng", "raw_affiliation_strings": ["School of Computing and Artificial Intelligence, Southwest Jiaotong University, Chengdu, China"], "raw_orcid": "https://orcid.org/0009-0008-2212-9768", "affiliations": [{"raw_affiliation_string": "School of Computing and Artificial Intelligence, Southwest Jiaotong University, Chengdu, China", "institution_ids": ["https://openalex.org/I4800084"]}]}], "abstract_inverted_index": {"Semantic": [0], "communication": [1], "(SC)": [2], "has": [3], "emerged": [4], "as": [5], "a": [6, 74, 81, 92, 138, 151, 167], "promising": [7], "paradigm": [8], "for": [9, 76], "upcoming": [10], "intelligent": [11, 19], "applications,": [12], "enabling": [13, 87], "mobile": [14, 88], "devices": [15, 89], "to": [16, 54, 69, 90, 100, 131, 145, 160, 194, 204], "collaboratively": [17], "execute": [18], "tasks": [20], "with": [21, 185, 215], "edge": [22, 84, 101], "servers": [23], "through": [24], "computation": [25], "offloading.": [26], "However,": [27], "few": [28], "studies": [29], "have": [30], "addressed": [31], "the": [32, 98, 109, 133, 174, 181, 191, 196, 206, 209], "problem": [33], "of": [34, 137, 208], "collaborative": [35, 41, 110], "inference": [36, 42, 111], "in": [37, 48], "SC": [38, 49], "networks.": [39], "Traditional": [40], "mechanisms": [43], "may": [44], "suffer": [45], "performance": [46], "decline": [47], "systems": [50], "and": [51, 73, 96, 119, 163, 178], "are": [52], "vulnerable": [53], "eavesdroppers.": [55], "To": [56], "address": [57, 132], "these": [58, 186], "issues,": [59], "first,": [60], "we": [61, 79, 107], "present": [62], "an": [63], "encryptor": [64], "that": [65, 142, 156, 172], "encrypts": [66], "semantic": [67, 94, 139, 182], "information": [68], "avoid": [70], "privacy": [71, 120], "leakage": [72], "decryptor": [75], "restoration.": [77], "Besides,": [78], "propose": [80], "novel": [82], "SC-empowered": [83], "computing": [85], "framework": [86], "deploy": [91], "partial": [93], "encoder": [95], "offload": [97], "rest": [99], "servers.": [102], "Based": [103], "on": [104, 126, 147], "this": [105], "framework,": [106], "formulate": [108], "optimization": [112], "problem,": [113, 134], "jointly": [114], "optimizing": [115], "delay,": [116], "energy": [117], "consumption,": [118], "leakage.": [121], "DNNPart": [122, 188], "is": [123], "devised": [124], "based": [125], "deep": [127], "deterministic": [128], "policy": [129], "gradient": [130], "which": [135], "consists": [136], "attention": [140], "mechanism": [141], "enables": [143], "it": [144, 158, 214], "focus": [146], "important": [148], "state": [149], "variables,": [150], "hybrid": [152], "action": [153, 165], "representation": [154], "method": [155, 211], "makes": [157], "adapt": [159], "mixed": [161], "discrete": [162], "continuous": [164], "spaces,": [166], "dynamic": [168], "model": [169], "splitting": [170], "algorithm": [171], "locates": [173], "optimal": [175, 197], "partition": [176], "layer": [177], "adaptively": [179], "splits": [180], "coders.": [183], "Integrated": [184], "components,": [187], "iteratively": [189], "optimizes": [190], "offloading": [192, 198], "strategy": [193], "find": [195], "strategy.": [199], "Extensive": [200], "simulations": [201], "were": [202], "conducted": [203], "verify": [205], "effectiveness": [207], "proposed": [210], "by": [212], "comparing": [213], "baseline": [216], "mechanisms.": [217]}, "primary_topic": {"id": "https://openalex.org/T10273", "display_name": "IoT and Edge/Fog Computing", "score": 0.9990000128746033, "subfield": {"id": "https://openalex.org/subfields/1705", "display_name": "Computer Networks and Communications"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, "topics": [{"id": "https://openalex.org/T10273", "display_name": "IoT and Edge/Fog Computing", "score": 0.9990000128746033, "subfield": {"id": "https://openalex.org/subfields/1705", "display_name": "Computer Networks and Communications"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, {"id": "https://openalex.org/T10764", "display_name": "Privacy-Preserving Technologies in Data", "score": 0.00019999999494757503, "subfield": {"id": "https://openalex.org/subfields/1702", "display_name": "Artificial Intelligence"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, {"id": "https://openalex.org/T10036", "display_name": "Advanced Neural Network Applications", "score": 0.0, "subfield": {"id": "https://openalex.org/subfields/1707", "display_name": "Computer Vision and Pattern Recognition"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}], "primary_location": {"id": "doi:10.1109/tmc.2025.3555298", "is_oa": false, "landing_page_url": "https://doi.org/10.1109/tmc.2025.3555298", "pdf_url": null, "source": {"id": "https://openalex.org/S69141925", "display_name": "IEEE Transactions on Mobile Computing", "issn_l": "1536-1233", "issn": ["1536-1233", "1558-0660", "2161-9875"], "is_oa": false, "is_in_doaj": false, "is_core": true, "listed_in": ["cwts-core", "jufo-3", "norway-2"], "host_organization": "https://openalex.org/P4310320439", "host_organization_name": "IEEE Computer Society", "host_organization_lineage": ["https://openalex.org/P4310320439", "https://openalex.org/P4310319808"], "host_organization_lineage_names": ["IEEE Computer Society", "Institute of Electrical and Electronics Engineers"], "type": "journal"}, "license": null, "license_id": null, "version": "publishedVersion", "is_accepted": true, "is_published": true, "raw_source_name": "IEEE Transactions on Mobile Computing", "raw_type": "journal-article"}, "cited_by_count": 11, "is_retracted": false}, {"id": "https://openalex.org/W4417302337", "doi": "https://doi.org/10.48550/arxiv.2505.12942", "title": "A3 : an Analytical Low-Rank Approximation Framework for Attention", "publication_year": 2025, "publication_date": "2025-05-19", "type": "preprint", "language": "en", "authorships": [{"author_position": "first", "author": {"id": null, "display_name": "Wong, Jeffrey T. H.", "orcid": null, "observed_orcids": []}, "institutions": [{"id": "https://openalex.org/I47508984", "display_name": "Imperial College London", "ror": "https://ror.org/041kmwe10", "country_code": "GB", "type": "education", "lineage": ["https://openalex.org/I47508984"]}], "countries": ["GB"], "is_corresponding": false, "raw_author_name": "Wong, Jeffrey T. H.", "raw_affiliation_strings": ["Department of Electrical and Electronic Engineering Imperial College London", "Elec- tronic Engineering , Imperial College London"], "raw_orcid": null, "affiliations": [{"raw_affiliation_string": "Department of Electrical and Electronic Engineering Imperial College London", "institution_ids": ["https://openalex.org/I47508984"]}, {"raw_affiliation_string": "Elec- tronic Engineering , Imperial College London", "institution_ids": ["https://openalex.org/I47508984"]}]}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5100769098", "display_name": "Cheng Zhang", "orcid": "https://orcid.org/0000-0001-8282-9901", "observed_orcids": ["https://orcid.org/0000-0001-8282-9901"]}, "institutions": [{"id": "https://openalex.org/I47508984", "display_name": "Imperial College London", "ror": "https://ror.org/041kmwe10", "country_code": "GB", "type": "education", "lineage": ["https://openalex.org/I47508984"]}], "countries": ["GB"], "is_corresponding": false, "raw_author_name": "Zhang, Cheng", "raw_affiliation_strings": ["Department of Electrical and Electronic Engineering Imperial College London", "Elec- tronic Engineering , Imperial College London"], "raw_orcid": null, "affiliations": [{"raw_affiliation_string": "Department of Electrical and Electronic Engineering Imperial College London", "institution_ids": ["https://openalex.org/I47508984"]}, {"raw_affiliation_string": "Elec- tronic Engineering , Imperial College London", "institution_ids": ["https://openalex.org/I47508984"]}]}, {"author_position": "middle", "author": {"id": null, "display_name": "Cao, Xinye", "orcid": null, "observed_orcids": []}, "institutions": [{"id": "https://openalex.org/I47508984", "display_name": "Imperial College London", "ror": "https://ror.org/041kmwe10", "country_code": "GB", "type": "education", "lineage": ["https://openalex.org/I47508984"]}], "countries": ["GB"], "is_corresponding": false, "raw_author_name": "Cao, Xinye", "raw_affiliation_strings": ["Department of Electrical and Electronic Engineering Imperial College London", "Elec- tronic Engineering , Imperial College London"], "raw_orcid": null, "affiliations": [{"raw_affiliation_string": "Department of Electrical and Electronic Engineering Imperial College London", "institution_ids": ["https://openalex.org/I47508984"]}, {"raw_affiliation_string": "Elec- tronic Engineering , Imperial College London", "institution_ids": ["https://openalex.org/I47508984"]}]}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5109819205", "display_name": "Pedro Gimenes", "orcid": "https://orcid.org/0009-0003-9207-2684", "observed_orcids": ["https://orcid.org/0009-0003-9207-2684"]}, "institutions": [{"id": "https://openalex.org/I47508984", "display_name": "Imperial College London", "ror": "https://ror.org/041kmwe10", "country_code": "GB", "type": "education", "lineage": ["https://openalex.org/I47508984"]}], "countries": ["GB"], "is_corresponding": false, "raw_author_name": "Gimenes, Pedro", "raw_affiliation_strings": ["Department of Electrical and Electronic Engineering Imperial College London", "Elec- tronic Engineering , Imperial College London"], "raw_orcid": null, "affiliations": [{"raw_affiliation_string": "Department of Electrical and Electronic Engineering Imperial College London", "institution_ids": ["https://openalex.org/I47508984"]}, {"raw_affiliation_string": "Elec- tronic Engineering , Imperial College London", "institution_ids": ["https://openalex.org/I47508984"]}]}, {"author_position": "middle", "author": {"id": null, "display_name": "Bouganis, Christos-Savvas", "orcid": null, "observed_orcids": []}, "institutions": [{"id": "https://openalex.org/I47508984", "display_name": "Imperial College London", "ror": "https://ror.org/041kmwe10", "country_code": "GB", "type": "education", "lineage": ["https://openalex.org/I47508984"]}], "countries": ["GB"], "is_corresponding": false, "raw_author_name": "Bouganis, Christos-Savvas", "raw_affiliation_strings": ["Elec- tronic Engineering , Imperial College London"], "raw_orcid": null, "affiliations": [{"raw_affiliation_string": "Elec- tronic Engineering , Imperial College London", "institution_ids": ["https://openalex.org/I47508984"]}]}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5029829952", "display_name": "George Anthony Constantinides", "orcid": "https://orcid.org/0000-0002-0201-310X", "observed_orcids": ["https://orcid.org/0000-0002-0201-310X"]}, "institutions": [{"id": "https://openalex.org/I47508984", "display_name": "Imperial College London", "ror": "https://ror.org/041kmwe10", "country_code": "GB", "type": "education", "lineage": ["https://openalex.org/I47508984"]}], "countries": ["GB"], "is_corresponding": false, "raw_author_name": "Constantinides, George A.", "raw_affiliation_strings": ["Department of Electrical and Electronic Engineering Imperial College London", "Elec- tronic Engineering , Imperial College London"], "raw_orcid": null, "affiliations": [{"raw_affiliation_string": "Department of Electrical and Electronic Engineering Imperial College London", "institution_ids": ["https://openalex.org/I47508984"]}, {"raw_affiliation_string": "Elec- tronic Engineering , Imperial College London", "institution_ids": ["https://openalex.org/I47508984"]}]}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5001082417", "display_name": "Wayne Luk", "orcid": null, "observed_orcids": []}, "institutions": [{"id": "https://openalex.org/I47508984", "display_name": "Imperial College London", "ror": "https://ror.org/041kmwe10", "country_code": "GB", "type": "education", "lineage": ["https://openalex.org/I47508984"]}], "countries": ["GB"], "is_corresponding": false, "raw_author_name": "Luk, Wayne", "raw_affiliation_strings": ["Department of Electrical and Electronic Engineering Imperial College London", "Elec- tronic Engineering , Imperial College London"], "raw_orcid": null, "affiliations": [{"raw_affiliation_string": "Department of Electrical and Electronic Engineering Imperial College London", "institution_ids": ["https://openalex.org/I47508984"]}, {"raw_affiliation_string": "Elec- tronic Engineering , Imperial College London", "institution_ids": ["https://openalex.org/I47508984"]}]}, {"author_position": "last", "author": {"id": "https://openalex.org/A5048483915", "display_name": "Yiren Zhao", "orcid": "https://orcid.org/0009-0005-2486-1491", "observed_orcids": ["https://orcid.org/0009-0005-2486-1491", "https://orcid.org/0000-0002-2373-2672"]}, "institutions": [{"id": "https://openalex.org/I47508984", "display_name": "Imperial College London", "ror": "https://ror.org/041kmwe10", "country_code": "GB", "type": "education", "lineage": ["https://openalex.org/I47508984"]}], "countries": ["GB"], "is_corresponding": false, "raw_author_name": "Zhao, Yiren", "raw_affiliation_strings": ["Department of Electrical and Electronic Engineering Imperial College London", "Elec- tronic Engineering , Imperial College London"], "raw_orcid": null, "affiliations": [{"raw_affiliation_string": "Department of Electrical and Electronic Engineering Imperial College London", "institution_ids": ["https://openalex.org/I47508984"]}, {"raw_affiliation_string": "Elec- tronic Engineering , Imperial College London", "institution_ids": ["https://openalex.org/I47508984"]}]}], "abstract_inverted_index": {"Large": [0], "language": [1], "models": [2], "have": [3, 26], "demonstrated": [4], "remarkable": [5], "performance;": [6], "however,": [7], "their": [8], "massive": [9], "parameter": [10], "counts": [11], "make": [12], "deployment": [13], "highly": [14], "expensive.": [15], "Low-rank": [16], "approximation": [17, 105], "offers": [18], "a": [19, 53, 102, 109, 186], "promising": [20], "compression": [21, 71], "solution,": [22], "yet": [23], "existing": [24], "approaches": [25], "two": [27, 58], "main": [28], "limitations:": [29], "(1)": [30], "They": [31], "focus": [32], "on": [33, 190], "minimizing": [34, 135], "the": [35, 44, 83, 127, 136, 172, 193], "output": [36], "error": [37], "of": [38, 47, 188, 204], "individual": [39], "linear": [40], "layers,": [41], "without": [42, 151], "considering": [43], "architectural": [45], "characteristics": [46], "Transformers,": [48], "and": [49, 75, 77, 88, 119, 121, 149, 178, 214, 221], "(2)": [50], "they": [51], "decompose": [52], "large": [54], "weight": [55], "matrix": [56], "into": [57, 112], "small": [59, 93], "low-rank": [60, 104, 181], "matrices.": [61, 94], "Consequently,": [62], "these": [63, 97], "methods": [64], "often": [65], "fall": [66], "short": [67], "compared": [68, 166], "to": [69, 167], "other": [70], "techniques": [72], "like": [73], "pruning": [74], "quantization,": [76, 212], "introduce": [78], "runtime": [79, 154], "overhead": [80], "such": [81], "as": [82], "extra": [84], "GEMM": [85], "kernel": [86], "launches": [87], "memory": [89], "operations": [90], "for": [91], "decomposed": [92], "To": [95], "address": [96], "limitations,": [98], "we": [99, 159], "propose": [100], "$A^3$,": [101], "post-training": [103], "framework.": [106], "$A^3$": [107, 162, 205], "splits": [108], "Transformer": [110], "layer": [111], "three": [113], "functional": [114, 138], "components,": [115], "namely": [116], "$\\texttt{QK}$,": [117], "$\\texttt{OV}$,": [118], "$\\texttt{MLP}$": [120], "provides": [122], "analytical": [123], "solutions": [124], "that": [125, 161], "reduces": [126, 143], "hidden": [128], "dimension": [129], "size": [130], "inside": [131], "each": [132], "component": [133], "while": [134], "component's": [137], "loss.": [139], "This": [140], "approach": [141], "directly": [142], "model": [144], "sizes,": [145, 148], "KV": [146, 207], "cache": [147, 208], "FLOPs": [150], "introducing": [152], "any": [153], "overheads.": [155], "Through": [156], "extensive": [157], "experiments,": [158], "show": [160, 201], "maintains": [163], "superior": [164], "performance": [165], "SoTAs.": [168], "For": [169], "example,": [170], "under": [171], "same": [173], "reduction": [174], "budget": [175], "in": [176, 206], "computation": [177], "memory,": [179], "our": [180, 219], "approximated": [182], "LLaMA": [183], "3.1-70B": [184], "achieves": [185], "perplexity": [187], "4.69": [189], "WikiText-2,": [191], "outperforming": [192], "previous": [194], "SoTA's": [195], "7.87": [196], "by": [197], "3.18.": [198], "We": [199, 217], "also": [200], "versatile": [202], "applications": [203], "compression,": [209], "integration": [210], "with": [211], "fine-tuning": [213], "mixed-rank": [215], "assignments.": [216], "open-sourced": [218], "framework": [220], "code": [222], "at": [223], "https://github.com/DeepWok/a3.": [224]}, "primary_topic": {"id": "https://openalex.org/T10036", "display_name": "Advanced Neural Network Applications", "score": 0.8389000296592712, "subfield": {"id": "https://openalex.org/subfields/1707", "display_name": "Computer Vision and Pattern Recognition"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, "topics": [{"id": "https://openalex.org/T10036", "display_name": "Advanced Neural Network Applications", "score": 0.8389000296592712, "subfield": {"id": "https://openalex.org/subfields/1707", "display_name": "Computer Vision and Pattern Recognition"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, {"id": "https://openalex.org/T10028", "display_name": "Topic Modeling", "score": 0.13130000233650208, "subfield": {"id": "https://openalex.org/subfields/1702", "display_name": "Artificial Intelligence"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, {"id": "https://openalex.org/T10181", "display_name": "Natural Language Processing Techniques", "score": 0.006800000090152025, "subfield": {"id": "https://openalex.org/subfields/1702", "display_name": "Artificial Intelligence"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}], "primary_location": {"id": "pmh:oai:arXiv.org:2505.12942", "is_oa": true, "landing_page_url": "https://arxiv.org/abs/2505.12942", "pdf_url": "https://arxiv.org/pdf/2505.12942", "source": {"id": "https://openalex.org/S4306400194", "display_name": "arXiv (Cornell University)", "issn_l": "2331-8422", "issn": ["2331-8422"], "is_oa": true, "is_in_doaj": false, "is_core": false, "listed_in": [], "host_organization": "https://openalex.org/I205783295", "host_organization_name": "Cornell University", "host_organization_lineage": ["https://openalex.org/I205783295"], "host_organization_lineage_names": [], "type": "repository"}, "license": "cc-by", "license_id": "https://openalex.org/licenses/cc-by", "version": "submittedVersion", "is_accepted": false, "is_published": false, "raw_source_name": null, "raw_type": "text"}, "cited_by_count": 0, "is_retracted": false}, {"id": "https://openalex.org/W7104830444", "doi": "https://doi.org/10.48550/arxiv.2511.05534", "title": "FlowMM: Cross-Modal Information Flow Guided KV Cache Merging for Efficient Multimodal Context Inference", "publication_year": 2025, "publication_date": "2025-10-29", "type": "preprint", "language": null, "authorships": [{"author_position": "first", "author": {"id": null, "display_name": "Li, Kunxi", "orcid": null, "observed_orcids": []}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Li, Kunxi", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": null, "display_name": "Xiong, Yufan", "orcid": null, "observed_orcids": []}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Xiong, Yufan", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": null, "display_name": "Jiang, Zhonghua", "orcid": null, "observed_orcids": []}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Jiang, Zhonghua", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": null, "display_name": "Zhou, Yiyun", "orcid": null, "observed_orcids": []}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Zhou, Yiyun", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": null, "display_name": "Wang, Zhaode", "orcid": null, "observed_orcids": []}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Wang, Zhaode", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": null, "display_name": "Lv, Chengfei", "orcid": null, "observed_orcids": []}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Lv, Chengfei", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "last", "author": {"id": null, "display_name": "Zhang, Shengyu", "orcid": null, "observed_orcids": []}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Zhang, Shengyu", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}], "abstract_inverted_index": {"Traditional": [0], "KV": [1, 27, 68, 122], "cache": [2, 69, 123], "eviction": [3, 30], "strategies,": [4, 81], "which": [5], "discard": [6], "less": [7], "critical": [8], "KV-pairs": [9], "based": [10, 35], "on": [11, 36], "attention": [12], "scores,": [13], "often": [14], "degrade": [15], "generation": [16], "quality,": [17], "causing": [18], "context": [19], "loss": [20], "or": [21], "hallucinations.": [22], "Recent": [23], "efforts": [24], "shift": [25], "toward": [26], "merging,": [28], "merging": [29, 80, 105], "tokens": [31, 34, 46, 107], "with": [32], "retention": [33], "similarity.": [37], "However,": [38], "in": [39, 50], "multimodal": [40, 67], "scenarios,": [41], "distributional": [42], "biases": [43, 49], "across": [44, 114], "modality": [45], "and": [47, 102, 129], "attentional": [48], "cross-modal": [51, 64, 73], "interactions": [52], "limit": [53], "its": [54], "effectiveness.": [55], "This": [56], "work": [57], "introduces": [58], "FlowMM,": [59], "an": [60], "adaptive": [61], "framework": [62], "for": [63], "information": [65, 74], "flow-guided": [66], "merging.": [70], "FlowMM": [71, 120], "leverages": [72], "flow": [75], "to": [76, 127], "dynamically": [77], "apply": [78], "layer-specific": [79], "capturing": [82], "modality-specific": [83], "patterns": [84], "while": [85, 108, 134], "preserving": [86], "contextual": [87], "integrity.": [88], "Furthermore,": [89], "we": [90], "introduce": [91], "a": [92], "sensitivity-adaptive": [93], "token": [94, 100], "matching": [95], "mechanism": [96], "that": [97, 119], "jointly": [98], "evaluates": [99], "similarity": [101], "task-critical": [103], "sensitivity,": [104], "low-risk": [106], "safeguarding": [109], "high-sensitivity": [110], "ones.": [111], "Extensive": [112], "experiments": [113], "diverse": [115], "leading": [116], "MLLMs": [117], "show": [118], "reduces": [121], "memory": [124], "by": [125, 132], "80%": [126], "95%": [128], "decoding": [130], "latency": [131], "1.3-1.8x,": [133], "maintaining": [135], "competitive": [136], "task": [137], "performance.": [138]}, "primary_topic": {"id": "https://openalex.org/T11714", "display_name": "Multimodal Machine Learning Applications", "score": 0.9800000190734863, "subfield": {"id": "https://openalex.org/subfields/1707", "display_name": "Computer Vision and Pattern Recognition"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, "topics": [{"id": "https://openalex.org/T11714", "display_name": "Multimodal Machine Learning Applications", "score": 0.9800000190734863, "subfield": {"id": "https://openalex.org/subfields/1707", "display_name": "Computer Vision and Pattern Recognition"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, {"id": "https://openalex.org/T10181", "display_name": "Natural Language Processing Techniques", "score": 0.007699999958276749, "subfield": {"id": "https://openalex.org/subfields/1702", "display_name": "Artificial Intelligence"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, {"id": "https://openalex.org/T10028", "display_name": "Topic Modeling", "score": 0.005900000222027302, "subfield": {"id": "https://openalex.org/subfields/1702", "display_name": "Artificial Intelligence"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}], "primary_location": {"id": "pmh:oai:arXiv.org:2511.05534", "is_oa": true, "landing_page_url": "http://arxiv.org/abs/2511.05534", "pdf_url": "https://arxiv.org/pdf/2511.05534", "source": {"id": "https://openalex.org/S4306400194", "display_name": "arXiv (Cornell University)", "issn_l": "2331-8422", "issn": ["2331-8422"], "is_oa": true, "is_in_doaj": false, "is_core": false, "listed_in": [], "host_organization": "https://openalex.org/I205783295", "host_organization_name": "Cornell University", "host_organization_lineage": ["https://openalex.org/I205783295"], "host_organization_lineage_names": [], "type": "repository"}, "license": null, "license_id": null, "version": "submittedVersion", "is_accepted": false, "is_published": false, "raw_source_name": null, "raw_type": "text"}, "cited_by_count": 0, "is_retracted": false}, {"id": "https://openalex.org/W4414536423", "doi": null, "title": "Datarus-R1: An Adaptive Multi-Step Reasoning LLM for Automated Data Analysis", "publication_year": 2025, "publication_date": "2025-08-18", "type": "preprint", "language": "en", "authorships": [{"author_position": "first", "author": {"id": "https://openalex.org/A5115756659", "display_name": "Ayoub Ben Chaliah", "orcid": null, "observed_orcids": []}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Chaliah, Ayoub Ben", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "last", "author": {"id": "https://openalex.org/A5115756660", "display_name": "Hela Dellagi", "orcid": null, "observed_orcids": []}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Dellagi, Hela", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}], "abstract_inverted_index": {"We": [0], "present": [1], "Datarus-R1-14B,": [2], "a": [3, 16, 51, 69, 82, 86, 92, 107], "14": [4], "B-parameter": [5], "open-weights": [6], "language": [7], "model": [8, 162], "fine-tuned": [9], "from": [10, 130], "Qwen": [11], "2.5-14B-Instruct": [12], "to": [13, 133, 170, 212, 238], "act": [14], "as": [15, 234], "virtual": [17], "data": [18, 72], "analyst": [19], "and": [20, 45, 60, 102, 105, 121, 140, 185, 204, 224, 245], "graduate-level": [21], "problem": [22], "solver.": [23], "Datarus": [24, 152, 191, 219], "is": [25, 153], "trained": [26], "not": [27], "on": [28, 33, 242], "isolated": [29], "question-answer": [30], "pairs": [31], "but": [32], "full": [34], "analytical": [35], "trajectories": [36], "including": [37], "reasoning": [38, 156, 231], "steps,": [39], "code": [40], "execution,": [41], "error": [42], "traces,": [43], "self-corrections,": [44], "final": [46], "conclusions,": [47], "all": [48], "captured": [49], "in": [50, 151, 174], "ReAct-style": [52], "notebook": [53, 79], "format": [54, 138], "spanning": [55], "finance,": [56], "medicine,": [57], "numerical": [58], "analysis,": [59], "other": [61], "quantitative": [62], "domains.": [63], "Our": [64], "training": [65], "pipeline": [66], "combines": [67], "(i)": [68], "trajectory-centric": [70], "synthetic": [71], "generator": [73], "that": [74, 97, 142, 166], "yielded": [75], "144": [76], "000": [77], "tagged": [78], "episodes,": [80], "(ii)": [81], "dual-reward": [83], "framework": [84], "blending": [85], "lightweight": [87], "tag-based": [88], "structural": [89, 131], "signal": [90], "with": [91], "Hierarchical": [93], "Reward": [94], "Model": [95], "(HRM)": [96], "scores": [98], "both": [99], "single-step": [100], "soundness": [101], "end-to-end": [103], "coherence,": [104], "(iii)": [106], "memory-optimized": [108], "implementation": [109], "of": [110, 229], "Group": [111], "Relative": [112], "Policy": [113], "Optimization": [114], "(GRPO)": [115], "featuring": [116], "KV-cache": [117], "reuse,": [118], "sequential": [119], "generation,": [120], "reference-model": [122], "sharding.": [123], "A": [124, 147], "cosine": [125], "curriculum": [126], "smoothly": [127], "shifts": [128], "emphasis": [129], "fidelity": [132], "semantic": [134], "depth,": [135], "reducing": [136], "the": [137, 161, 207, 227], "collapse": [139], "verbosity": [141], "often": [143], "plague": [144], "RL-aligned": [145], "LLMs.": [146], "central": [148], "design": [149], "choice": [150], "it": [154, 177, 196], "dual": [155], "interface.": [157], "In": [158], "agentic": [159], "mode": [160, 176], "produces": [163], "ReAct-tagged": [164], "steps": [165], "invoke": [167], "Python": [168], "tools": [169], "execute": [171], "real": [172], "code;": [173], "reflection": [175], "outputs": [178], "compact": [179], "Chain-of-Thought": [180], "(CoT)": [181], "traces": [182], "delimited": [183], "by": [184], "tags.": [186], "On": [187], "demanding": [188], "postgraduate-level": [189], "problems,": [190], "exhibits": [192], "an": [193], "\"AHA-moment\"": [194], "pattern:": [195], "sketches": [197], "hypotheses,": [198], "revises": [199], "them": [200], "once": [201], "or": [202], "twice,": [203], "converges": [205], "avoiding": [206], "circular,": [208], "token-inflating": [209], "loops": [210], "common": [211], "contemporary": [213], "systems.": [214], "Across": [215], "standard": [216], "public": [217], "benchmarks": [218], "surpasses": [220], "similar": [221], "size": [222], "models": [223, 232], "even": [225], "reaches": [226], "level": [228], "larger": [230], "such": [233], "QwQ-32B": [235], "achieving": [236], "up": [237], "30%": [239], "higher": [240], "accuracy": [241], "AIME": [243], "2024/2025": [244], "LiveCodeBench": [246], "while": [247], "emitting": [248], "18-49%": [249], "fewer": [250], "tokens": [251], "per": [252], "solution.": [253]}, "primary_topic": {"id": "https://openalex.org/T10028", "display_name": "Topic Modeling", "score": 0.7354000210762024, "subfield": {"id": "https://openalex.org/subfields/1702", "display_name": "Artificial Intelligence"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, "topics": [{"id": "https://openalex.org/T10028", "display_name": "Topic Modeling", "score": 0.7354000210762024, "subfield": {"id": "https://openalex.org/subfields/1702", "display_name": "Artificial Intelligence"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, {"id": "https://openalex.org/T11714", "display_name": "Multimodal Machine Learning Applications", "score": 0.04309999942779541, "subfield": {"id": "https://openalex.org/subfields/1707", "display_name": "Computer Vision and Pattern Recognition"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, {"id": "https://openalex.org/T13904", "display_name": "Artificial Intelligence Applications", "score": 0.040699999779462814, "subfield": {"id": "https://openalex.org/subfields/1702", "display_name": "Artificial Intelligence"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}], "primary_location": {"id": "pmh:oai:arXiv.org:2508.13382", "is_oa": true, "landing_page_url": "http://arxiv.org/abs/2508.13382", "pdf_url": "https://arxiv.org/pdf/2508.13382", "source": {"id": "https://openalex.org/S4306400194", "display_name": "arXiv (Cornell University)", "issn_l": "2331-8422", "issn": ["2331-8422"], "is_oa": true, "is_in_doaj": false, "is_core": false, "listed_in": [], "host_organization": "https://openalex.org/I205783295", "host_organization_name": "Cornell University", "host_organization_lineage": ["https://openalex.org/I205783295"], "host_organization_lineage_names": [], "type": "repository"}, "license": null, "license_id": null, "version": "submittedVersion", "is_accepted": false, "is_published": false, "raw_source_name": null, "raw_type": "text"}, "cited_by_count": 0, "is_retracted": false}, {"id": "https://openalex.org/W4414682088", "doi": "https://doi.org/10.48550/arxiv.2506.18952", "title": "LLMs on a Budget? Say HOLA", "publication_year": 2025, "publication_date": "2025-06-23", "type": "preprint", "language": "en", "authorships": [{"author_position": "first", "author": {"id": "https://openalex.org/A5114105249", "display_name": "Zohaib Hasan Siddiqui", "orcid": null, "observed_orcids": []}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Siddiqui, Zohaib Hasan", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5012051562", "display_name": "Jiechao Gao", "orcid": "https://orcid.org/0000-0003-0628-1416", "observed_orcids": ["https://orcid.org/0000-0003-0628-1416"]}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Gao, Jiechao", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5119042124", "display_name": "Ebad Shabbir", "orcid": "https://orcid.org/0009-0003-0062-9159", "observed_orcids": ["https://orcid.org/0009-0003-0062-9159"]}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Shabbir, Ebad", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5093258760", "display_name": "Mohammad Anas Azeez", "orcid": "https://orcid.org/0009-0000-3344-0166", "observed_orcids": ["https://orcid.org/0009-0000-3344-0166"]}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Azeez, Mohammad Anas", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": null, "display_name": "Ali, Rafiq", "orcid": null, "observed_orcids": []}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Ali, Rafiq", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5076626481", "display_name": "Gautam Siddharth Kashyap", "orcid": "https://orcid.org/0000-0003-2140-9617", "observed_orcids": ["https://orcid.org/0000-0003-2140-9617"]}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Kashyap, Gautam Siddharth", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "last", "author": {"id": "https://openalex.org/A5077006200", "display_name": "Usman Naseem", "orcid": "https://orcid.org/0000-0003-0191-7171", "observed_orcids": ["https://orcid.org/0000-0003-0191-7171"]}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Naseem, Usman", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}], "abstract_inverted_index": {"Running": [0], "Large": [1], "Language": [2], "Models": [3], "(LLMs)": [4], "on": [5, 47, 81, 100, 104, 111], "edge": [6, 112], "devices": [7, 113], "is": [8], "constrained": [9], "by": [10], "high": [11], "compute": [12], "and": [13, 27, 36, 44, 92, 106, 109, 119], "memory": [14, 110], "demands": [15], "posing": [16], "a": [17], "barrier": [18], "for": [19, 58, 69], "real-time": [20], "applications": [21], "in": [22], "sectors": [23], "like": [24, 114], "healthcare,": [25], "education,": [26], "embedded": [28], "systems.": [29], "Current": [30], "solutions": [31], "such": [32], "as": [33], "quantization,": [34, 93], "pruning,": [35], "retrieval-augmented": [37], "generation": [38], "(RAG)": [39], "offer": [40], "only": [41], "partial": [42], "optimizations": [43], "often": [45], "compromise": [46], "speed": [48], "or": [49], "accuracy.": [50], "We": [51], "introduce": [52], "HOLA,": [53], "an": [54], "end-to-end": [55], "optimization": [56], "framework": [57], "efficient": [59], "LLM": [60], "deployment.": [61], "Internally,": [62], "it": [63], "leverages": [64], "Hierarchical": [65], "Speculative": [66], "Decoding": [67], "(HSD)": [68], "faster": [70], "inference": [71], "without": [72], "quality": [73], "loss.": [74], "Externally,": [75], "AdaComp-RAG": [76], "adjusts": [77], "retrieval": [78], "complexity": [79], "based": [80], "context": [82], "needs.": [83], "Together": [84], "with": [85], "LoBi,": [86], "which": [87], "blends": [88], "structured": [89], "pruning": [90], "(LoRA)": [91], "HOLA": [94], "delivers": [95], "significant": [96], "gains:": [97], "17.6%": [98], "EMA": [99], "GSM8K,": [101], "10.5%": [102], "MCA": [103], "ARC,": [105], "reduced": [107], "latency": [108], "Jetson": [115], "Nano--proving": [116], "both": [117], "scalable": [118], "production-ready.": [120]}, "primary_topic": {"id": "https://openalex.org/T10036", "display_name": "Advanced Neural Network Applications", "score": 0.7368000149726868, "subfield": {"id": "https://openalex.org/subfields/1707", "display_name": "Computer Vision and Pattern Recognition"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, "topics": [{"id": "https://openalex.org/T10036", "display_name": "Advanced Neural Network Applications", "score": 0.7368000149726868, "subfield": {"id": "https://openalex.org/subfields/1707", "display_name": "Computer Vision and Pattern Recognition"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, {"id": "https://openalex.org/T10028", "display_name": "Topic Modeling", "score": 0.12720000743865967, "subfield": {"id": "https://openalex.org/subfields/1702", "display_name": "Artificial Intelligence"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, {"id": "https://openalex.org/T10181", "display_name": "Natural Language Processing Techniques", "score": 0.047600001096725464, "subfield": {"id": "https://openalex.org/subfields/1702", "display_name": "Artificial Intelligence"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}], "primary_location": {"id": "pmh:oai:arXiv.org:2506.18952", "is_oa": true, "landing_page_url": "http://arxiv.org/abs/2506.18952", "pdf_url": "https://arxiv.org/pdf/2506.18952", "source": {"id": "https://openalex.org/S4306400194", "display_name": "arXiv (Cornell University)", "issn_l": "2331-8422", "issn": ["2331-8422"], "is_oa": true, "is_in_doaj": false, "is_core": false, "listed_in": [], "host_organization": "https://openalex.org/I205783295", "host_organization_name": "Cornell University", "host_organization_lineage": ["https://openalex.org/I205783295"], "host_organization_lineage_names": [], "type": "repository"}, "license": null, "license_id": null, "version": "submittedVersion", "is_accepted": false, "is_published": false, "raw_source_name": null, "raw_type": "text"}, "cited_by_count": 0, "is_retracted": false}, {"id": "https://openalex.org/W4414581508", "doi": "https://doi.org/10.48550/arxiv.2505.18610", "title": "PM-KVQ: Progressive Mixed-precision KV Cache Quantization for Long-CoT LLMs", "publication_year": 2025, "publication_date": "2025-05-24", "type": "preprint", "language": "en", "authorships": [{"author_position": "first", "author": {"id": "https://openalex.org/A5040810891", "display_name": "Tengxuan Liu", "orcid": null, "observed_orcids": []}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Liu, Tengxuan", "raw_affiliation_strings": ["Tsinghua University , 2 Infinigence-AI ,"], "raw_orcid": null, "affiliations": [{"raw_affiliation_string": "Tsinghua University , 2 Infinigence-AI ,", "institution_ids": []}]}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5101827286", "display_name": "Shiyao Li", "orcid": "https://orcid.org/0000-0002-7258-0564", "observed_orcids": ["https://orcid.org/0000-0002-7258-0564"]}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Li, Shiyao", "raw_affiliation_strings": ["Tsinghua University , 2 Infinigence-AI ,"], "raw_orcid": null, "affiliations": [{"raw_affiliation_string": "Tsinghua University , 2 Infinigence-AI ,", "institution_ids": []}]}, {"author_position": "middle", "author": {"id": null, "display_name": "Yang, Jiayi", "orcid": null, "observed_orcids": []}, "institutions": [{"id": "https://openalex.org/I78577930", "display_name": "Columbia University", "ror": "https://ror.org/00hj8s172", "country_code": "US", "type": "education", "lineage": ["https://openalex.org/I78577930"]}], "countries": ["US"], "is_corresponding": false, "raw_author_name": "Yang, Jiayi", "raw_affiliation_strings": ["Columbia University ,"], "raw_orcid": null, "affiliations": [{"raw_affiliation_string": "Columbia University ,", "institution_ids": ["https://openalex.org/I78577930"]}]}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5039134666", "display_name": "Tianchen Zhao", "orcid": "https://orcid.org/0000-0002-2071-7514", "observed_orcids": ["https://orcid.org/0000-0002-2071-7514", "https://orcid.org/0000-0003-2911-7205", "https://orcid.org/0009-0003-6001-5669"]}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Zhao, Tianchen", "raw_affiliation_strings": ["Tsinghua University , 2 Infinigence-AI ,"], "raw_orcid": null, "affiliations": [{"raw_affiliation_string": "Tsinghua University , 2 Infinigence-AI ,", "institution_ids": []}]}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5064086456", "display_name": "Feng Zhou", "orcid": "https://orcid.org/0000-0002-2906-8127", "observed_orcids": ["https://orcid.org/0000-0002-2906-8127", "https://orcid.org/0009-0006-5301-7019"]}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Zhou, Feng", "raw_affiliation_strings": ["OPPO AI Center , Beijing , China,"], "raw_orcid": null, "affiliations": [{"raw_affiliation_string": "OPPO AI Center , Beijing , China,", "institution_ids": []}]}, {"author_position": "middle", "author": {"id": null, "display_name": "Song, Xiaohui", "orcid": null, "observed_orcids": []}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Song, Xiaohui", "raw_affiliation_strings": ["OPPO AI Center , Beijing , China,"], "raw_orcid": null, "affiliations": [{"raw_affiliation_string": "OPPO AI Center , Beijing , China,", "institution_ids": []}]}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5102805465", "display_name": "Guohao Dai", "orcid": "https://orcid.org/0000-0002-8464-0130", "observed_orcids": ["https://orcid.org/0000-0002-8464-0130", "https://orcid.org/0000-0003-0849-3252"]}, "institutions": [{"id": "https://openalex.org/I183067930", "display_name": "Shanghai Jiao Tong University", "ror": "https://ror.org/0220qvk04", "country_code": "CN", "type": "education", "lineage": ["https://openalex.org/I183067930"]}], "countries": ["CN"], "is_corresponding": false, "raw_author_name": "Dai, Guohao", "raw_affiliation_strings": ["Shanghai Jiaotong University"], "raw_orcid": null, "affiliations": [{"raw_affiliation_string": "Shanghai Jiaotong University", "institution_ids": ["https://openalex.org/I183067930"]}]}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5049910854", "display_name": "Shengen Yan", "orcid": "https://orcid.org/0009-0005-3858-7972", "observed_orcids": ["https://orcid.org/0009-0005-3858-7972"]}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Yan, Shengen", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5103867707", "display_name": "Huazhong Yang", "orcid": "https://orcid.org/0000-0003-2421-353X", "observed_orcids": ["https://orcid.org/0000-0003-2421-353X"]}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Yang, Huazhong", "raw_affiliation_strings": ["Tsinghua University , 2 Infinigence-AI ,"], "raw_orcid": null, "affiliations": [{"raw_affiliation_string": "Tsinghua University , 2 Infinigence-AI ,", "institution_ids": []}]}, {"author_position": "last", "author": {"id": "https://openalex.org/A5101879008", "display_name": "Yu Wang", "orcid": "https://orcid.org/0000-0003-0212-6830", "observed_orcids": ["https://orcid.org/0000-0003-0212-6830", "https://orcid.org/0000-0003-3511-0288"]}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Wang, Yu", "raw_affiliation_strings": ["Tsinghua University , 2 Infinigence-AI ,"], "raw_orcid": null, "affiliations": [{"raw_affiliation_string": "Tsinghua University , 2 Infinigence-AI ,", "institution_ids": []}]}], "abstract_inverted_index": {"Recently,": [0], "significant": [1, 64], "progress": [2], "has": [3, 40, 48], "been": [4, 49], "made": [5], "in": [6, 52, 128, 133, 152, 174], "developing": [7], "reasoning-capable": [8], "Large": [9, 74], "Language": [10], "Models": [11], "(LLMs)": [12], "through": [13], "long": [14], "Chain-of-Thought": [15], "(CoT)": [16], "techniques.": [17], "However,": [18, 55], "this": [19], "long-CoT": [20, 61, 145, 231], "reasoning": [21, 237], "process": [22], "imposes": [23], "substantial": [24], "memory": [25, 34, 181, 250], "overhead": [26], "due": [27, 67], "to": [28, 60, 68, 80, 97, 106, 119, 147, 166, 183, 188, 219, 242], "the": [29, 69, 89, 111, 122, 129, 149, 169, 196, 221, 248, 257], "large": [30, 98], "Key-Value": [31], "(KV)": [32], "Cache": [33, 38, 91, 141, 173], "overhead.": [35], "Post-training": [36], "KV": [37, 90, 140, 172], "quantization": [39, 100, 164], "emerged": [41], "as": [42], "a": [43, 162, 185, 204], "promising": [44], "compression": [45], "technique": [46], "and": [47, 85, 252], "extensively": [50], "studied": [51], "short-context": [53, 114], "scenarios.": [54], "directly": [56, 87], "applying": [57], "existing": [58], "methods": [59, 78], "LLMs": [62, 146, 232], "causes": [63], "performance": [65, 134, 239], "degradation": [66], "following": [70], "two": [71, 153], "reasons:": [72], "(1)": [73, 155], "cumulative": [75, 99, 158], "error:": [76], "Existing": [77], "fail": [79], "adequately": [81], "leverage": [82], "available": [83], "memory,": [84], "they": [86], "quantize": [88], "during": [92, 116], "each": [93, 175], "decoding": [94], "step,": [95], "leading": [96], "error.": [101], "(2)": [102, 193], "Short-context": [103], "calibration:": [104], "Due": [105], "Rotary": [107], "Positional": [108], "Embedding": [109], "(RoPE),": [110], "use": [112], "of": [113, 124, 171, 224], "data": [115, 215, 222], "calibration": [117, 197, 206, 214], "fails": [118], "account": [120], "for": [121, 144], "distribution": [123, 223], "less": [125], "frequent": [126], "channels": [127], "Key": [130], "Cache,": [131], "resulting": [132], "loss.": [135], "We": [136], "propose": [137, 179, 203], "Progressive": [138], "Mixed-Precision": [139], "Quantization": [142], "(PM-KVQ)": [143], "address": [148], "above": [150], "issues": [151], "folds:": [154], "To": [156, 194], "reduce": [157], "error,": [159], "we": [160, 178, 202], "design": [161], "progressive": [163], "strategy": [165, 207], "gradually": [167], "lower": [168], "bit-width": [170, 187], "block.": [176], "Then,": [177], "block-wise": [180], "allocation": [182], "assign": [184], "higher": [186], "more": [189], "sensitive": [190], "transformer": [191], "blocks.": [192], "increase": [195], "length": [198], "without": [199], "additional": [200], "overhead,": [201], "new": [205], "with": [208, 216], "positional": [209, 217], "interpolation": [210, 218], "that": [211, 234], "leverages": [212], "short": [213], "approximate": [220], "long-context": [225], "data.": [226], "Extensive": [227], "experiments": [228], "on": [229], "7B-70B": [230], "show": [233], "PM-KVQ": [235], "improves": [236], "benchmark": [238], "by": [240], "up": [241], "8%": [243], "over": [244, 256], "SOTA": [245], "baselines": [246], "under": [247], "same": [249], "budget": [251], "achieves": [253], "2.73-5.18x": [254], "throughput": [255], "original": [258], "16-bit": [259], "LLMs.": [260]}, "primary_topic": {"id": "https://openalex.org/T10036", "display_name": "Advanced Neural Network Applications", "score": 0.8003000020980835, "subfield": {"id": "https://openalex.org/subfields/1707", "display_name": "Computer Vision and Pattern Recognition"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, "topics": [{"id": "https://openalex.org/T10036", "display_name": "Advanced Neural Network Applications", "score": 0.8003000020980835, "subfield": {"id": "https://openalex.org/subfields/1707", "display_name": "Computer Vision and Pattern Recognition"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, {"id": "https://openalex.org/T10028", "display_name": "Topic Modeling", "score": 0.05260000005364418, "subfield": {"id": "https://openalex.org/subfields/1702", "display_name": "Artificial Intelligence"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, {"id": "https://openalex.org/T10181", "display_name": "Natural Language Processing Techniques", "score": 0.020999999716877937, "subfield": {"id": "https://openalex.org/subfields/1702", "display_name": "Artificial Intelligence"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}], "primary_location": {"id": "pmh:oai:arXiv.org:2505.18610", "is_oa": true, "landing_page_url": "https://arxiv.org/abs/2505.18610", "pdf_url": "https://arxiv.org/pdf/2505.18610", "source": {"id": "https://openalex.org/S4306400194", "display_name": "arXiv (Cornell University)", "issn_l": "2331-8422", "issn": ["2331-8422"], "is_oa": true, "is_in_doaj": false, "is_core": false, "listed_in": [], "host_organization": "https://openalex.org/I205783295", "host_organization_name": "Cornell University", "host_organization_lineage": ["https://openalex.org/I205783295"], "host_organization_lineage_names": [], "type": "repository"}, "license": null, "license_id": null, "version": "submittedVersion", "is_accepted": false, "is_published": false, "raw_source_name": null, "raw_type": "text"}, "cited_by_count": 0, "is_retracted": false}, {"id": "https://openalex.org/W4416052729", "doi": "https://doi.org/10.48550/arxiv.2504.11320", "title": "Optimizing LLM Inference: Fluid-Guided Online Scheduling with Memory Constraints", "publication_year": 2025, "publication_date": "2025-04-15", "type": "preprint", "language": "en", "authorships": [{"author_position": "first", "author": {"id": "https://openalex.org/A5093973906", "display_name": "Ruicheng Ao", "orcid": "https://orcid.org/0009-0009-6217-9695", "observed_orcids": ["https://orcid.org/0009-0009-6217-9695"]}, "institutions": [{"id": "https://openalex.org/I63966007", "display_name": "Massachusetts Institute of Technology", "ror": "https://ror.org/042nb2s44", "country_code": "US", "type": "education", "lineage": ["https://openalex.org/I63966007"]}], "countries": ["US"], "is_corresponding": false, "raw_author_name": "Ao, Ruicheng", "raw_affiliation_strings": ["Institute for Data, Systems, and Society , Massachusetts Institute of Technology , Cambridge , MA 02139", "Massachusetts Institute of Technology"], "raw_orcid": null, "affiliations": [{"raw_affiliation_string": "Institute for Data, Systems, and Society , Massachusetts Institute of Technology , Cambridge , MA 02139", "institution_ids": ["https://openalex.org/I63966007"]}, {"raw_affiliation_string": "Massachusetts Institute of Technology", "institution_ids": ["https://openalex.org/I63966007"]}]}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5101862787", "display_name": "Gan Luo", "orcid": "https://orcid.org/0000-0002-3644-5752", "observed_orcids": ["https://orcid.org/0000-0002-3644-5752", "https://orcid.org/0000-0001-6570-8673", "https://orcid.org/0000-0002-9588-7008"]}, "institutions": [{"id": "https://openalex.org/I20231570", "display_name": "Peking University", "ror": "https://ror.org/02v51f717", "country_code": "CN", "type": "education", "lineage": ["https://openalex.org/I20231570"]}], "countries": ["CN"], "is_corresponding": false, "raw_author_name": "Luo, Gan", "raw_affiliation_strings": ["Peking University", "School of Mathematical Sciences , Peking University , Beijing 100871"], "raw_orcid": null, "affiliations": [{"raw_affiliation_string": "Peking University", "institution_ids": ["https://openalex.org/I20231570"]}, {"raw_affiliation_string": "School of Mathematical Sciences , Peking University , Beijing 100871", "institution_ids": ["https://openalex.org/I20231570"]}]}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5112431388", "display_name": "David Simchi\u2010Levi", "orcid": "https://orcid.org/0000-0002-4650-1519", "observed_orcids": ["https://orcid.org/0000-0002-4650-1519"]}, "institutions": [{"id": "https://openalex.org/I63966007", "display_name": "Massachusetts Institute of Technology", "ror": "https://ror.org/042nb2s44", "country_code": "US", "type": "education", "lineage": ["https://openalex.org/I63966007"]}], "countries": ["US"], "is_corresponding": false, "raw_author_name": "Simchi-Levi, David", "raw_affiliation_strings": ["Institute for Data, Systems, and Society , Massachusetts Institute of Technology , Cambridge , MA 02139", "Massachusetts Institute of Technology"], "raw_orcid": null, "affiliations": [{"raw_affiliation_string": "Institute for Data, Systems, and Society , Massachusetts Institute of Technology , Cambridge , MA 02139", "institution_ids": ["https://openalex.org/I63966007"]}, {"raw_affiliation_string": "Massachusetts Institute of Technology", "institution_ids": ["https://openalex.org/I63966007"]}]}, {"author_position": "last", "author": {"id": "https://openalex.org/A5071705536", "display_name": "Xinshang Wang", "orcid": "https://orcid.org/0000-0003-4683-7167", "observed_orcids": ["https://orcid.org/0000-0003-4683-7167"]}, "institutions": [{"id": "https://openalex.org/I45928872", "display_name": "Alibaba Group (China)", "ror": "https://ror.org/00k642b80", "country_code": "CN", "type": "company", "lineage": ["https://openalex.org/I45928872"]}], "countries": ["CN"], "is_corresponding": false, "raw_author_name": "Wang, Xinshang", "raw_affiliation_strings": ["Alibaba Group"], "raw_orcid": null, "affiliations": [{"raw_affiliation_string": "Alibaba Group", "institution_ids": ["https://openalex.org/I45928872"]}]}], "abstract_inverted_index": {"Large": [0], "language": [1], "models": [2], "now": [3], "serve": [4], "millions": [5], "of": [6, 148], "users": [7], "daily,": [8], "with": [9, 63, 170], "providers": [10], "incurring": [11], "costs": [12], "exceeding": [13], "$700,000": [14], "per": [15], "day.": [16], "Each": [17], "request": [18], "requires": [19], "token-by-token": [20], "inference,": [21], "making": [22], "GPU": [23], "scheduling": [24, 61], "central": [25], "to": [26, 117, 151, 188], "latency,": [27], "capacity,": [28], "and": [29, 44, 50, 70, 86, 110, 193, 199], "cost.": [30], "The": [31], "difficulty": [32], "is": [33], "endogenous": [34, 64], "memory": [35, 65, 84, 139], "growth:": [36], "generated": [37], "tokens": [38], "expand": [39], "the": [40, 91, 115, 132, 137, 176, 178, 181], "Key-Value": [41], "(KV)": [42], "cache,": [43], "overflow": [45], "can": [46], "evict": [47], "in-progress": [48], "requests": [49, 124], "waste": [51], "prior": [52], "computation.": [53], "We": [54, 74], "formulate": [55], "inference": [56], "as": [57], "a": [58, 76, 102], "multi-stage": [59], "online": [60], "problem": [62], "growth,": [66], "linear": [67], "iteration": [68], "times,": [69], "GPU-resident": [71], "KV-cache": [72], "constraints.": [73], "introduce": [75], "fluid": [77, 92, 133], "model": [78], "that": [79], "characterizes": [80], "equilibrium": [81], "batch": [82], "composition,": [83], "requirement,": [85], "stability": [87], "region.": [88], "Guided": [89], "by": [90, 121], "model,": [93], "we": [94], "design": [95], "WAIT": [96, 142], "(Waiting": [97], "for": [98, 106, 164], "Accumulated": [99], "Inference": [100], "Threshold),": [101], "threshold-based": [103], "admission": [104], "rule": [105, 116], "known": [107], "output": [108, 119, 158], "lengths,": [109], "Nested": [111, 141], "WAIT,": [112], "which": [113], "extends": [114], "unknown": [118, 157], "lengths": [120], "regulating": [122], "how": [123], "advance": [125], "across": [126], "decode-stage": [127], "segments.": [128], "Both": [129], "algorithms": [130, 192], "approximate": [131], "benchmark": [134], "asymptotically": [135], "under": [136, 156], "stated": [138], "conditions.": [140], "uses": [143], "an": [144, 167], "additional": [145], "safety": [146], "buffer": [147], "moderate": [149], "scale": [150], "hedge": [152], "against": [153], "memory-overflow-induced": [154], "evictions": [155], "lengths.": [159], "In": [160], "Vidur": [161], "simulations": [162], "configured": [163], "Llama-2-7B": [165], "on": [166], "A100": [168], "GPU,": [169], "supplemental": [171], "real-GPU": [172], "validation": [173], "reported": [174], "in": [175, 197], "appendix,": [177], "policies": [179], "enlarge": [180], "empirically": [182], "observed": [183], "stable": [184], "operating": [185], "range": [186], "relative": [187], "widely": [189], "used": [190], "baseline": [191], "reduce": [194], "latency": [195], "especially": [196], "near-overloaded": [198], "overloaded": [200], "regimes.": [201]}, "primary_topic": {"id": "https://openalex.org/T10054", "display_name": "Parallel Computing and Optimization Techniques", "score": 0.3375999927520752, "subfield": {"id": "https://openalex.org/subfields/1708", "display_name": "Hardware and Architecture"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, "topics": [{"id": "https://openalex.org/T10054", "display_name": "Parallel Computing and Optimization Techniques", "score": 0.3375999927520752, "subfield": {"id": "https://openalex.org/subfields/1708", "display_name": "Hardware and Architecture"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, {"id": "https://openalex.org/T10101", "display_name": "Cloud Computing and Resource Management", "score": 0.15219999849796295, "subfield": {"id": "https://openalex.org/subfields/1710", "display_name": "Information Systems"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, {"id": "https://openalex.org/T10974", "display_name": "Advanced Queuing Theory Analysis", "score": 0.11089999973773956, "subfield": {"id": "https://openalex.org/subfields/1404", "display_name": "Management Information Systems"}, "field": {"id": "https://openalex.org/fields/14", "display_name": "Business, Management and Accounting"}, "domain": {"id": "https://openalex.org/domains/2", "display_name": "Social Sciences"}}], "primary_location": {"id": "pmh:oai:arXiv.org:2504.11320", "is_oa": true, "landing_page_url": "https://arxiv.org/abs/2504.11320", "pdf_url": "https://arxiv.org/pdf/2504.11320", "source": {"id": "https://openalex.org/S4306400194", "display_name": "arXiv (Cornell University)", "issn_l": "2331-8422", "issn": ["2331-8422"], "is_oa": true, "is_in_doaj": false, "is_core": false, "listed_in": [], "host_organization": "https://openalex.org/I205783295", "host_organization_name": "Cornell University", "host_organization_lineage": ["https://openalex.org/I205783295"], "host_organization_lineage_names": [], "type": "repository"}, "license": "cc-by-nc-sa", "license_id": "https://openalex.org/licenses/cc-by-nc-sa", "version": "submittedVersion", "is_accepted": false, "is_published": false, "raw_source_name": null, "raw_type": "text"}, "cited_by_count": 0, "is_retracted": false}, {"id": "https://openalex.org/W4417531298", "doi": "https://doi.org/10.48550/arxiv.2512.15773", "title": "TS-DP: Reinforcement Speculative Decoding For Temporal Adaptive Diffusion Policy Acceleration", "publication_year": 2025, "publication_date": "2025-12-13", "type": "preprint", "language": null, "authorships": [{"author_position": "first", "author": {"id": null, "display_name": "Li, Ye", "orcid": null, "observed_orcids": []}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Li, Ye", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": null, "display_name": "Feng, Jiahe", "orcid": null, "observed_orcids": []}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Feng, Jiahe", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": null, "display_name": "Meng, Yuan", "orcid": null, "observed_orcids": []}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Meng, Yuan", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": null, "display_name": "Ji, Kangye", "orcid": null, "observed_orcids": []}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Ji, Kangye", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": null, "display_name": "Tang, Chen", "orcid": null, "observed_orcids": []}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Tang, Chen", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": null, "display_name": "Wen, Xinwan", "orcid": null, "observed_orcids": []}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Wen, Xinwan", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": null, "display_name": "Xia, Shutao", "orcid": null, "observed_orcids": []}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Xia, Shutao", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": null, "display_name": "Wang, Zhi", "orcid": null, "observed_orcids": []}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Wang, Zhi", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "last", "author": {"id": null, "display_name": "Zhu, Wenwu", "orcid": null, "observed_orcids": []}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Zhu, Wenwu", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}], "abstract_inverted_index": {"Diffusion": [0, 113], "Policy": [1, 114], "(DP)": [2], "excels": [3], "in": [4, 87, 102], "embodied": [5, 26, 48, 88, 179], "control": [6, 207], "but": [7], "suffers": [8], "from": [9], "high": [10], "inference": [11, 190, 198], "latency": [12], "and": [13, 31, 36, 56, 90, 94, 149, 203], "computational": [14], "cost": [15, 82], "due": [16], "to": [17, 44, 67, 73, 92, 129, 144, 161, 169, 186], "multiple": [18], "iterative": [19], "denoising": [20, 78, 153], "steps.": [21], "The": [22], "temporal": [23, 126], "complexity": [24], "of": [25, 200], "tasks": [27], "demands": [28], "a": [29, 54, 141], "dynamic": [30, 47, 131], "adaptable": [32], "computation": [33, 97], "mode.": [34], "Static": [35], "lossy": [37], "acceleration": [38], "methods,": [39], "such": [40, 46, 103], "as": [41], "quantization,": [42], "fail": [43], "handle": [45, 130], "tasks,": [49], "while": [50, 172], "speculative": [51, 121, 167], "decoding": [52, 122], "offers": [53], "lossless": [55], "adaptive": [57], "yet": [58], "underexplored": [59], "alternative": [60], "for": [61, 123], "DP.": [62], "However,": [63], "it": [64], "is": [65], "non-trivial": [66], "address": [68], "the": [69, 75, 116, 146], "following": [70], "challenges:": [71], "how": [72, 91], "match": [74], "base": [76, 147], "model's": [77], "quality": [79], "at": [80], "lower": [81], "under": [83], "time-varying": [84, 162], "task": [85, 100, 134, 163], "difficulty": [86, 101, 135, 164], "settings,": [89], "dynamically": [93], "interactively": [95], "adjust": [96], "based": [98], "on": [99], "environments.": [104], "In": [105], "this": [106], "paper,": [107], "we": [108, 139], "propose": [109], "Temporal-aware": [110], "Reinforcement-based": [111], "Speculative": [112], "(TS-DP),": [115], "first": [117], "framework": [118], "that": [119, 182], "enables": [120], "DP": [124], "with": [125, 191], "adaptivity.": [127], "First,": [128], "environments": [132, 180], "where": [133], "varies": [136], "over": [137, 192], "time,": [138], "distill": [140], "Transformer-based": [142], "drafter": [143], "imitate": [145], "model": [148], "replace": [150], "its": [151], "costly": [152], "calls.": [154], "Second,": [155], "an": [156, 197], "RL-based": [157], "scheduler": [158], "further": [159], "adapts": [160], "by": [165], "adjusting": [166], "parameters": [168], "maintain": [170], "accuracy": [171], "improving": [173], "efficiency.": [174], "Extensive": [175], "experiments": [176], "across": [177], "diverse": [178], "demonstrate": [181], "TS-DP": [183], "achieves": [184], "up": [185], "4.17": [187], "times": [188], "faster": [189], "94%": [193], "accepted": [194], "drafts,": [195], "reaching": [196], "frequency": [199], "25": [201], "Hz": [202], "enabling": [204], "real-time": [205], "diffusion-based": [206], "without": [208], "performance": [209], "degradation.": [210]}, "primary_topic": {"id": "https://openalex.org/T10462", "display_name": "Reinforcement Learning in Robotics", "score": 0.8500999808311462, "subfield": {"id": "https://openalex.org/subfields/1702", "display_name": "Artificial Intelligence"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, "topics": [{"id": "https://openalex.org/T10462", "display_name": "Reinforcement Learning in Robotics", "score": 0.8500999808311462, "subfield": {"id": "https://openalex.org/subfields/1702", "display_name": "Artificial Intelligence"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, {"id": "https://openalex.org/T10653", "display_name": "Robot Manipulation and Learning", "score": 0.03530000150203705, "subfield": {"id": "https://openalex.org/subfields/2207", "display_name": "Control and Systems Engineering"}, "field": {"id": "https://openalex.org/fields/22", "display_name": "Engineering"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, {"id": "https://openalex.org/T10775", "display_name": "Generative Adversarial Networks and Image Synthesis", "score": 0.026200000196695328, "subfield": {"id": "https://openalex.org/subfields/1707", "display_name": "Computer Vision and Pattern Recognition"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}], "primary_location": {"id": "pmh:oai:arXiv.org:2512.15773", "is_oa": true, "landing_page_url": "http://arxiv.org/abs/2512.15773", "pdf_url": "https://arxiv.org/pdf/2512.15773", "source": {"id": "https://openalex.org/S4306400194", "display_name": "arXiv (Cornell University)", "issn_l": "2331-8422", "issn": ["2331-8422"], "is_oa": true, "is_in_doaj": false, "is_core": false, "listed_in": [], "host_organization": "https://openalex.org/I205783295", "host_organization_name": "Cornell University", "host_organization_lineage": ["https://openalex.org/I205783295"], "host_organization_lineage_names": [], "type": "repository"}, "license": null, "license_id": null, "version": "submittedVersion", "is_accepted": false, "is_published": false, "raw_source_name": null, "raw_type": "text"}, "cited_by_count": 0, "is_retracted": false}, {"id": "https://openalex.org/W7116895983", "doi": "https://doi.org/10.48550/arxiv.2512.17452", "title": "KV Admission: Learning What to Write for Efficient Long-Context LLM Inference", "publication_year": 2025, "publication_date": "2025-12-19", "type": "preprint", "language": null, "authorships": [{"author_position": "first", "author": {"id": "https://openalex.org/A5121023040", "display_name": "Yen-Chieh Huang", "orcid": null, "observed_orcids": []}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Huang, Yen-Chieh", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5036230837", "display_name": "Pi-Cheng Hsiu", "orcid": "https://orcid.org/0000-0001-8035-4033", "observed_orcids": ["https://orcid.org/0000-0001-8035-4033"]}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Hsiu, Pi-Cheng", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5101550831", "display_name": "Rui Fang", "orcid": "https://orcid.org/0000-0003-3676-1795", "observed_orcids": ["https://orcid.org/0000-0003-3676-1795"]}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Fang, Rui", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "last", "author": {"id": "https://openalex.org/A5146777586", "display_name": "Ming-Syan Chen", "orcid": null, "observed_orcids": []}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Chen, Ming-Syan", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}], "abstract_inverted_index": {"Long-context": [0], "LLM": [1], "inference": [2], "is": [3, 109, 120], "bottlenecked": [4], "by": [5], "the": [6, 27], "quadratic": [7], "attention": [8], "complexity": [9], "and": [10, 50, 94, 98, 112], "linear": [11], "Key-Value": [12], "(KV)": [13], "cache": [14, 70, 83], "growth.": [15], "Prior": [16], "approaches": [17], "mitigate": [18], "this": [19, 34], "via": [20, 56], "post-hoc": [21], "selection": [22], "or": [23], "eviction": [24], "but": [25], "overlook": [26], "root": [28], "inefficiency:": [29], "indiscriminate": [30], "token": [31, 67], "admission.": [32], "In": [33], "paper,": [35], "we": [36], "formalize": [37], "KV": [38, 47, 54, 58], "management": [39], "as": [40], "a": [41, 60, 80, 85, 110], "causal": [42], "system": [43], "of": [44], "three": [45], "primitives:": [46], "Admission,": [48], "Selection,": [49], "Eviction.": [51], "We": [52], "instantiate": [53], "Admission": [55], "Write-Gated": [57], "(WG-KV),": [59], "lightweight": [61], "mechanism": [62], "that": [63, 104], "learns": [64], "to": [65, 78, 107], "predict": [66], "utility": [68], "before": [69], "entry.": [71], "By": [72], "filtering": [73], "out": [74], "redundant": [75], "states": [76], "early": [77], "maintain": [79], "compact": [81], "global": [82], "alongside": [84], "sliding": [86], "local": [87], "cache,": [88], "WG-KV": [89], "significantly": [90], "reduces": [91], "memory": [92], "usage": [93], "accelerates": [95], "both": [96], "prefill": [97], "decode": [99], "phases.": [100], "Our": [101], "results": [102], "demonstrate": [103], "learning": [105], "what": [106], "write": [108], "principled": [111], "practical": [113], "recipe": [114], "for": [115], "efficient": [116], "long-context": [117], "inference.": [118], "Code": [119], "available": [121], "at": [122], "https://github.com/EMCLab-Sinica/WG-KV.": [123]}, "primary_topic": {"id": "https://openalex.org/T10181", "display_name": "Natural Language Processing Techniques", "score": 0.37959998846054077, "subfield": {"id": "https://openalex.org/subfields/1702", "display_name": "Artificial Intelligence"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, "topics": [{"id": "https://openalex.org/T10181", "display_name": "Natural Language Processing Techniques", "score": 0.37959998846054077, "subfield": {"id": "https://openalex.org/subfields/1702", "display_name": "Artificial Intelligence"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, {"id": "https://openalex.org/T10028", "display_name": "Topic Modeling", "score": 0.3140000104904175, "subfield": {"id": "https://openalex.org/subfields/1702", "display_name": "Artificial Intelligence"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, {"id": "https://openalex.org/T10036", "display_name": "Advanced Neural Network Applications", "score": 0.09189999848604202, "subfield": {"id": "https://openalex.org/subfields/1707", "display_name": "Computer Vision and Pattern Recognition"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}], "primary_location": {"id": "pmh:oai:arXiv.org:2512.17452", "is_oa": true, "landing_page_url": "https://arxiv.org/abs/2512.17452", "pdf_url": "https://arxiv.org/pdf/2512.17452", "source": {"id": "https://openalex.org/S4306400194", "display_name": "arXiv (Cornell University)", "issn_l": "2331-8422", "issn": ["2331-8422"], "is_oa": true, "is_in_doaj": false, "is_core": false, "listed_in": [], "host_organization": "https://openalex.org/I205783295", "host_organization_name": "Cornell University", "host_organization_lineage": ["https://openalex.org/I205783295"], "host_organization_lineage_names": [], "type": "repository"}, "license": null, "license_id": null, "version": "submittedVersion", "is_accepted": false, "is_published": false, "raw_source_name": null, "raw_type": "text"}, "cited_by_count": 0, "is_retracted": false}, {"id": "https://openalex.org/W4415275681", "doi": "https://doi.org/10.48550/arxiv.2510.13334", "title": "Taming the Fragility of KV Cache Eviction in LLM Inference", "publication_year": 2025, "publication_date": "2025-10-15", "type": "preprint", "language": null, "authorships": [{"author_position": "first", "author": {"id": "https://openalex.org/A5062436243", "display_name": "Yuan Feng", "orcid": "https://orcid.org/0009-0006-6448-9345", "observed_orcids": ["https://orcid.org/0009-0006-6448-9345", "https://orcid.org/0000-0001-9888-4622"]}, "institutions": [{"id": "https://openalex.org/I126520041", "display_name": "University of Science and Technology of China", "ror": "https://ror.org/04c4dkn09", "country_code": "CN", "type": "education", "lineage": ["https://openalex.org/I126520041", "https://openalex.org/I19820366"]}], "countries": ["CN"], "is_corresponding": false, "raw_author_name": "Feng, Yuan", "raw_affiliation_strings": ["Data Darkness Lab , MIRACLE Center , Suzhou Institute for Advanced Research", "School of Computer Science , University of Science and Technology of China"], "raw_orcid": null, "affiliations": [{"raw_affiliation_string": "Data Darkness Lab , MIRACLE Center , Suzhou Institute for Advanced Research", "institution_ids": []}, {"raw_affiliation_string": "School of Computer Science , University of Science and Technology of China", "institution_ids": ["https://openalex.org/I126520041"]}]}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5120033994", "display_name": "Haoyu Guo", "orcid": null, "observed_orcids": []}, "institutions": [{"id": "https://openalex.org/I91090688", "display_name": "University of Science and Technology Chittagong", "ror": "https://ror.org/00w9tx359", "country_code": "BD", "type": "education", "lineage": ["https://openalex.org/I91090688"]}], "countries": ["BD"], "is_corresponding": false, "raw_author_name": "Guo, Haoyu", "raw_affiliation_strings": ["Data Darkness Lab , MIRACLE Center , Suzhou Institute for Advanced Research", "School of Biomedical Engineering , USTC"], "raw_orcid": null, "affiliations": [{"raw_affiliation_string": "Data Darkness Lab , MIRACLE Center , Suzhou Institute for Advanced Research", "institution_ids": []}, {"raw_affiliation_string": "School of Biomedical Engineering , USTC", "institution_ids": ["https://openalex.org/I91090688"]}]}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5049757790", "display_name": "Junlin Lv", "orcid": null, "observed_orcids": []}, "institutions": [{"id": "https://openalex.org/I126520041", "display_name": "University of Science and Technology of China", "ror": "https://ror.org/04c4dkn09", "country_code": "CN", "type": "education", "lineage": ["https://openalex.org/I126520041", "https://openalex.org/I19820366"]}], "countries": ["CN"], "is_corresponding": false, "raw_author_name": "Lv, JunLin", "raw_affiliation_strings": ["Data Darkness Lab , MIRACLE Center , Suzhou Institute for Advanced Research", "School of Computer Science , University of Science and Technology of China"], "raw_orcid": null, "affiliations": [{"raw_affiliation_string": "Data Darkness Lab , MIRACLE Center , Suzhou Institute for Advanced Research", "institution_ids": []}, {"raw_affiliation_string": "School of Computer Science , University of Science and Technology of China", "institution_ids": ["https://openalex.org/I126520041"]}]}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5001768407", "display_name": "S. Kevin Zhou", "orcid": "https://orcid.org/0009-0009-3863-1100", "observed_orcids": ["https://orcid.org/0009-0009-3863-1100"]}, "institutions": [{"id": "https://openalex.org/I91090688", "display_name": "University of Science and Technology Chittagong", "ror": "https://ror.org/00w9tx359", "country_code": "BD", "type": "education", "lineage": ["https://openalex.org/I91090688"]}], "countries": ["BD"], "is_corresponding": false, "raw_author_name": "Zhou, S. Kevin", "raw_affiliation_strings": ["Data Darkness Lab , MIRACLE Center , Suzhou Institute for Advanced Research", "School of Biomedical Engineering , USTC"], "raw_orcid": null, "affiliations": [{"raw_affiliation_string": "Data Darkness Lab , MIRACLE Center , Suzhou Institute for Advanced Research", "institution_ids": []}, {"raw_affiliation_string": "School of Biomedical Engineering , USTC", "institution_ids": ["https://openalex.org/I91090688"]}]}, {"author_position": "last", "author": {"id": "https://openalex.org/A5037366245", "display_name": "Xike Xie", "orcid": "https://orcid.org/0000-0001-5290-5408", "observed_orcids": ["https://orcid.org/0000-0001-5290-5408"]}, "institutions": [{"id": "https://openalex.org/I91090688", "display_name": "University of Science and Technology Chittagong", "ror": "https://ror.org/00w9tx359", "country_code": "BD", "type": "education", "lineage": ["https://openalex.org/I91090688"]}], "countries": ["BD"], "is_corresponding": false, "raw_author_name": "Xie, Xike", "raw_affiliation_strings": ["Data Darkness Lab , MIRACLE Center , Suzhou Institute for Advanced Research", "School of Biomedical Engineering , USTC"], "raw_orcid": null, "affiliations": [{"raw_affiliation_string": "Data Darkness Lab , MIRACLE Center , Suzhou Institute for Advanced Research", "institution_ids": []}, {"raw_affiliation_string": "School of Biomedical Engineering , USTC", "institution_ids": ["https://openalex.org/I91090688"]}]}], "abstract_inverted_index": {"Large": [0], "language": [1, 6], "models": [2], "have": [3], "revolutionized": [4], "natural": [5], "processing,": [7], "yet": [8, 107], "their": [9], "deployment": [10], "remains": [11, 49], "hampered": [12], "by": [13, 161], "the": [14, 21, 41, 77, 167], "substantial": [15], "memory": [16], "and": [17, 140, 163, 181], "runtime": [18], "overhead": [19], "of": [20, 47], "transformer's": [22], "Key-Value": [23], "cache.": [24], "To": [25, 100], "mitigate": [26], "this,": [27, 102], "recent": [28], "methods": [29, 156], "employ": [30], "a": [31, 44, 73, 105, 112, 134, 171, 183], "scoring-aggregation": [32], "framework": [33], "to": [34, 68, 72], "evict": [35], "unimportant": [36], "cache": [37, 136, 173, 188], "entries,": [38], "based": [39], "on": [40, 60], "stability": [42, 78], "assumption-that": [43], "fixed": [45], "subset": [46], "entries": [48], "consistently": [50], "important": [51], "during": [52], "generation.": [53], "However,": [54], "prior": [55], "work": [56], "has": [57], "largely": [58], "focused": [59], "refining": [61], "importance": [62], "indicators": [63], "for": [64, 186], "scoring,": [65], "while": [66], "defaulting": [67], "mean": [69, 93], "aggregation": [70, 94, 110], "due": [71], "faithful": [74], "trust": [75], "in": [76, 97], "assumption.": [79], "In": [80], "this": [81, 86, 130], "work,": [82], "we": [83, 103, 132], "argue": [84], "that": [85, 116], "underlying": [87, 191], "assumption": [88], "is": [89, 199], "inherently": [90], "fragile,": [91], "making": [92], "highly": [95], "vulnerable": [96], "extreme": [98, 123], "cases.": [99], "counter": [101], "propose": [104, 133], "simple": [106], "elegant": [108], "defensive": [109], "strategy:": [111], "two-step,": [113], "linear-time": [114], "approach": [115], "controls": [117], "worst-case": [118, 194], "risk,": [119], "thereby": [120], "defending": [121], "against": [122, 190], "cases": [124], "with": [125], "negligible": [126], "computational": [127], "overhead.": [128], "Embodying": [129], "strategy,": [131], "novel": [135], "eviction": [137, 189], "method,": [138], "DefensiveKV": [139], "its": [141], "extension,": [142], "Layer-DefensiveKV,": [143], "which": [144], "incorporates": [145], "layer-wise": [146], "budget": [147], "allocation.": [148], "Across": [149], "seven": [150], "task": [151], "domains": [152], "(18": [153], "datasets),": [154], "our": [155], "reduce": [157], "generation": [158], "quality": [159], "loss": [160], "2.3x": [162], "4.3x": [164], "respectively,": [165], "versus": [166], "strongest": [168], "baseline": [169], "under": [170], "20%": [172], "size.": [174], "These": [175], "results": [176], "set": [177], "new": [178], "performance": [179], "benchmarks": [180], "pioneer": [182], "promising": [184], "direction": [185], "optimizing": [187], "fragility": [192], "through": [193], "risk": [195], "management.": [196], "Our": [197], "code": [198], "available": [200], "at": [201], "https://github.com/FFY0/DefensiveKV.": [202]}, "primary_topic": {"id": "https://openalex.org/T10181", "display_name": "Natural Language Processing Techniques", "score": 0.5717999935150146, "subfield": {"id": "https://openalex.org/subfields/1702", "display_name": "Artificial Intelligence"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, "topics": [{"id": "https://openalex.org/T10181", "display_name": "Natural Language Processing Techniques", "score": 0.5717999935150146, "subfield": {"id": "https://openalex.org/subfields/1702", "display_name": "Artificial Intelligence"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, {"id": "https://openalex.org/T10028", "display_name": "Topic Modeling", "score": 0.34279999136924744, "subfield": {"id": "https://openalex.org/subfields/1702", "display_name": "Artificial Intelligence"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, {"id": "https://openalex.org/T10036", "display_name": "Advanced Neural Network Applications", "score": 0.025200000032782555, "subfield": {"id": "https://openalex.org/subfields/1707", "display_name": "Computer Vision and Pattern Recognition"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}], "primary_location": {"id": "pmh:oai:arXiv.org:2510.13334", "is_oa": true, "landing_page_url": "http://arxiv.org/abs/2510.13334", "pdf_url": "https://arxiv.org/pdf/2510.13334", "source": {"id": "https://openalex.org/S4306400194", "display_name": "arXiv (Cornell University)", "issn_l": "2331-8422", "issn": ["2331-8422"], "is_oa": true, "is_in_doaj": false, "is_core": false, "listed_in": [], "host_organization": "https://openalex.org/I205783295", "host_organization_name": "Cornell University", "host_organization_lineage": ["https://openalex.org/I205783295"], "host_organization_lineage_names": [], "type": "repository"}, "license": "cc-by", "license_id": "https://openalex.org/licenses/cc-by", "version": "submittedVersion", "is_accepted": false, "is_published": false, "raw_source_name": null, "raw_type": "text"}, "cited_by_count": 0, "is_retracted": false}, {"id": "https://openalex.org/W7111265756", "doi": "https://doi.org/10.6084/m9.figshare.29951927.v1", "title": "Hyper-Specific Sub-Field Selection & Research Topic Generation: Adaptive Quantization-Aware Neural Network Compilers for FPGA-based Direct Digital Synthesizers (DDS)", "publication_year": 2025, "publication_date": "2025-08-20", "type": "article", "language": null, "authorships": [{"author_position": "first", "author": {"id": null, "display_name": "LIM KYUNGJUN (21983285)", "orcid": null, "observed_orcids": []}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "LIM KYUNGJUN (21983285)", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "last", "author": {"id": null, "display_name": "Freederia AI Researcher (21983315)", "orcid": null, "observed_orcids": []}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Freederia AI Researcher (21983315)", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}], "abstract_inverted_index": {"**Random": [0], "Sub-Field:**": [1], "Adaptive": [2], "Quantization-Aware": [3], "Neural": [4], "Network": [5], "Compilers": [6], "(AQNNCs)": [7], "applied": [8], "to": [9, 47], "Field-Programmable": [10], "Gate": [11], "Arrays": [12], "(FPGAs)": [13], "for": [14], "efficient": [15], "Direct": [16], "Digital": [17], "Synthesizers": [18], "(DDS).": [19], "---": [20], "*This": [21], "document": [22], "is": [23], "a": [24], "part": [25], "of": [26, 35], "the": [27], "Freederia": [28], "Research": [29], "Archive.": [30], "Explore": [31], "our": [32, 42, 51], "complete": [33], "collection": [34], "advanced": [36], "research": [37], "at": [38, 45], "[en.freederia.com](https://en.freederia.com),": [39], "or": [40], "visit": [41], "main": [43], "portal": [44], "[freederia.com](https://freederia.com)": [46], "learn": [48], "more": [49], "about": [50], "mission": [52], "and": [53], "other": [54], "initiatives.*": [55]}, "primary_topic": {"id": "https://openalex.org/T11417", "display_name": "Advancements in PLL and VCO Technologies", "score": 0.37059998512268066, "subfield": {"id": "https://openalex.org/subfields/2208", "display_name": "Electrical and Electronic Engineering"}, "field": {"id": "https://openalex.org/fields/22", "display_name": "Engineering"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, "topics": [{"id": "https://openalex.org/T11417", "display_name": "Advancements in PLL and VCO Technologies", "score": 0.37059998512268066, "subfield": {"id": "https://openalex.org/subfields/2208", "display_name": "Electrical and Electronic Engineering"}, "field": {"id": "https://openalex.org/fields/22", "display_name": "Engineering"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, {"id": "https://openalex.org/T10904", "display_name": "Embedded Systems Design Techniques", "score": 0.25459998846054077, "subfield": {"id": "https://openalex.org/subfields/1708", "display_name": "Hardware and Architecture"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, {"id": "https://openalex.org/T11522", "display_name": "VLSI and FPGA Design Techniques", "score": 0.12409999966621399, "subfield": {"id": "https://openalex.org/subfields/2208", "display_name": "Electrical and Electronic Engineering"}, "field": {"id": "https://openalex.org/fields/22", "display_name": "Engineering"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}], "primary_location": {"id": "pmh:oai:figshare.com:article/29951927", "is_oa": true, "landing_page_url": "https://figshare.com/articles/journal_contribution/Hyper-Specific_Sub-Field_Selection_Research_Topic_Generation_Adaptive_Quantization-Aware_Neural_Network_Compilers_for_FPGA-based_Direct_Digital_Synthesizers_DDS_/29951927", "pdf_url": null, "source": {"id": "https://openalex.org/S4377196282", "display_name": "Figshare", "issn_l": null, "issn": null, "is_oa": false, "is_in_doaj": false, "is_core": false, "listed_in": [], "host_organization": "https://openalex.org/I4210132348", "host_organization_name": "Figshare (United Kingdom)", "host_organization_lineage": ["https://openalex.org/I4210132348"], "host_organization_lineage_names": [], "type": "repository"}, "license": "cc-by", "license_id": "https://openalex.org/licenses/cc-by", "version": "submittedVersion", "is_accepted": false, "is_published": false, "raw_source_name": null, "raw_type": "Text"}, "cited_by_count": 0, "is_retracted": false}, {"id": "https://openalex.org/W4415021917", "doi": "https://doi.org/10.48550/arxiv.2505.14163", "title": "DSMentor: Enhancing Data Science Agents with Curriculum Learning and Online Knowledge Accumulation", "publication_year": 2025, "publication_date": "2025-05-20", "type": "preprint", "language": "en", "authorships": [{"author_position": "first", "author": {"id": "https://openalex.org/A5050061175", "display_name": "He Wang", "orcid": "https://orcid.org/0000-0002-2281-5679", "observed_orcids": ["https://orcid.org/0000-0002-2281-5679"]}, "institutions": [{"id": "https://openalex.org/I74973139", "display_name": "Carnegie Mellon University", "ror": "https://ror.org/05x2bcf33", "country_code": "US", "type": "education", "lineage": ["https://openalex.org/I74973139"]}], "countries": ["US"], "is_corresponding": false, "raw_author_name": "Wang, He", "raw_affiliation_strings": ["Carnegie Mellon University"], "raw_orcid": null, "affiliations": [{"raw_affiliation_string": "Carnegie Mellon University", "institution_ids": ["https://openalex.org/I74973139"]}]}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5023587890", "display_name": "Alexander Hanbo Li", "orcid": null, "observed_orcids": []}, "institutions": [{"id": "https://openalex.org/I74973139", "display_name": "Carnegie Mellon University", "ror": "https://ror.org/05x2bcf33", "country_code": "US", "type": "education", "lineage": ["https://openalex.org/I74973139"]}], "countries": ["US"], "is_corresponding": false, "raw_author_name": "Li, Alexander Hanbo", "raw_affiliation_strings": ["Carnegie Mellon University"], "raw_orcid": null, "affiliations": [{"raw_affiliation_string": "Carnegie Mellon University", "institution_ids": ["https://openalex.org/I74973139"]}]}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5102837904", "display_name": "Yiqun Hu", "orcid": "https://orcid.org/0000-0001-9157-7865", "observed_orcids": ["https://orcid.org/0000-0001-9157-7865"]}, "institutions": [{"id": "https://openalex.org/I74973139", "display_name": "Carnegie Mellon University", "ror": "https://ror.org/05x2bcf33", "country_code": "US", "type": "education", "lineage": ["https://openalex.org/I74973139"]}], "countries": ["US"], "is_corresponding": false, "raw_author_name": "Hu, Yiqun", "raw_affiliation_strings": ["Carnegie Mellon University"], "raw_orcid": null, "affiliations": [{"raw_affiliation_string": "Carnegie Mellon University", "institution_ids": ["https://openalex.org/I74973139"]}]}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5100394043", "display_name": "Sheng Zhang", "orcid": "https://orcid.org/0000-0003-0843-0622", "observed_orcids": ["https://orcid.org/0000-0003-0843-0622"]}, "institutions": [{"id": "https://openalex.org/I74973139", "display_name": "Carnegie Mellon University", "ror": "https://ror.org/05x2bcf33", "country_code": "US", "type": "education", "lineage": ["https://openalex.org/I74973139"]}], "countries": ["US"], "is_corresponding": false, "raw_author_name": "Zhang, Sheng", "raw_affiliation_strings": ["Carnegie Mellon University"], "raw_orcid": null, "affiliations": [{"raw_affiliation_string": "Carnegie Mellon University", "institution_ids": ["https://openalex.org/I74973139"]}]}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5102100621", "display_name": "Hideo Kobayashi", "orcid": null, "observed_orcids": []}, "institutions": [{"id": "https://openalex.org/I74973139", "display_name": "Carnegie Mellon University", "ror": "https://ror.org/05x2bcf33", "country_code": "US", "type": "education", "lineage": ["https://openalex.org/I74973139"]}], "countries": ["US"], "is_corresponding": false, "raw_author_name": "Kobayashi, Hideo", "raw_affiliation_strings": ["Carnegie Mellon University"], "raw_orcid": null, "affiliations": [{"raw_affiliation_string": "Carnegie Mellon University", "institution_ids": ["https://openalex.org/I74973139"]}]}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5040806743", "display_name": "Jiani Zhang", "orcid": "https://orcid.org/0000-0003-0074-6761", "observed_orcids": ["https://orcid.org/0000-0003-0074-6761"]}, "institutions": [{"id": "https://openalex.org/I74973139", "display_name": "Carnegie Mellon University", "ror": "https://ror.org/05x2bcf33", "country_code": "US", "type": "education", "lineage": ["https://openalex.org/I74973139"]}], "countries": ["US"], "is_corresponding": false, "raw_author_name": "Zhang, Jiani", "raw_affiliation_strings": ["Carnegie Mellon University"], "raw_orcid": null, "affiliations": [{"raw_affiliation_string": "Carnegie Mellon University", "institution_ids": ["https://openalex.org/I74973139"]}]}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5091847816", "display_name": "Henry Zhu", "orcid": "https://orcid.org/0009-0008-4846-728X", "observed_orcids": ["https://orcid.org/0009-0008-4846-728X", "https://orcid.org/0000-0002-2242-8038"]}, "institutions": [{"id": "https://openalex.org/I74973139", "display_name": "Carnegie Mellon University", "ror": "https://ror.org/05x2bcf33", "country_code": "US", "type": "education", "lineage": ["https://openalex.org/I74973139"]}], "countries": ["US"], "is_corresponding": false, "raw_author_name": "Zhu, Henry", "raw_affiliation_strings": ["Carnegie Mellon University"], "raw_orcid": null, "affiliations": [{"raw_affiliation_string": "Carnegie Mellon University", "institution_ids": ["https://openalex.org/I74973139"]}]}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5042384129", "display_name": "Chung-Wei Hang", "orcid": null, "observed_orcids": []}, "institutions": [{"id": "https://openalex.org/I74973139", "display_name": "Carnegie Mellon University", "ror": "https://ror.org/05x2bcf33", "country_code": "US", "type": "education", "lineage": ["https://openalex.org/I74973139"]}], "countries": ["US"], "is_corresponding": false, "raw_author_name": "Hang, Chung-Wei", "raw_affiliation_strings": ["Carnegie Mellon University"], "raw_orcid": null, "affiliations": [{"raw_affiliation_string": "Carnegie Mellon University", "institution_ids": ["https://openalex.org/I74973139"]}]}, {"author_position": "last", "author": {"id": "https://openalex.org/A5103996035", "display_name": "Patrick Ng", "orcid": null, "observed_orcids": []}, "institutions": [{"id": "https://openalex.org/I74973139", "display_name": "Carnegie Mellon University", "ror": "https://ror.org/05x2bcf33", "country_code": "US", "type": "education", "lineage": ["https://openalex.org/I74973139"]}], "countries": ["US"], "is_corresponding": false, "raw_author_name": "Ng, Patrick", "raw_affiliation_strings": ["Carnegie Mellon University"], "raw_orcid": null, "affiliations": [{"raw_affiliation_string": "Carnegie Mellon University", "institution_ids": ["https://openalex.org/I74973139"]}]}], "abstract_inverted_index": {"Large": [0], "language": [1], "model": [2], "(LLM)": [3], "agents": [4], "have": [5], "shown": [6], "promising": [7], "performance": [8, 89, 214], "in": [9, 40, 90, 102], "generating": [10], "code": [11], "for": [12, 195, 211], "solving": [13], "complex": [14, 78], "data": [15, 92, 99], "science": [16, 93, 100], "problems.": [17], "Recent": [18], "studies": [19], "primarily": [20], "focus": [21], "on": [22, 136, 155, 176], "enhancing": [23], "in-context": [24], "learning": [25, 64, 120, 205], "through": [26, 133, 215], "improved": [27], "search,": [28], "sampling,": [29], "and": [30, 73, 107, 122, 138, 157, 197, 207], "planning": [31], "techniques,": [32], "while": [33], "overlooking": [34], "the": [35, 38, 81, 118, 148, 171, 177, 189, 203], "importance": [36, 190], "of": [37, 104, 127, 191], "order": [39, 103], "which": [41, 61], "problems": [42, 179], "are": [43], "tackled": [44], "during": [45, 200], "inference.": [46], "In": [47], "this": [48], "work,": [49], "we": [50], "develop": [51], "a": [52, 66, 109], "novel": [53], "inference-time": [54], "optimization": [55], "framework,": [56], "referred": [57], "to": [58, 76, 85, 113, 153, 160, 181], "as": [59, 80], "DSMentor,": [60], "leverages": [62], "curriculum": [63], "--": [65, 84], "strategy": [67], "that": [68, 143], "introduces": [69], "simpler": [70], "task": [71], "first": [72], "progressively": [74], "moves": [75], "more": [77, 124], "ones": [79], "learner": [82], "improves": [83, 147], "enhance": [86], "LLM": [87, 213], "agent": [88], "challenging": [91], "tasks.": [94], "Our": [95, 186], "mentor-guided": [96], "framework": [97], "organizes": [98], "tasks": [101], "increasing": [105], "difficulty": [106], "incorporates": [108], "growing": [110], "long-term": [111], "memory": [112], "retain": [114], "prior": [115], "experiences,": [116], "guiding": [117], "agent's": [119], "progression": [121], "enabling": [123], "effective": [125, 193], "utilization": [126], "accumulated": [128], "knowledge.": [129], "We": [130], "evaluate": [131], "DSMentor": [132, 144, 164], "extensive": [134], "experiments": [135], "DSEval": [137, 156], "QRData": [139, 158], "benchmarks.": [140], "Experiments": [141], "show": [142], "using": [145, 183], "Claude-3.5-Sonnet": [146], "pass": [149, 172], "rate": [150, 173], "by": [151, 174], "up": [152], "5.2%": [154], "compared": [159, 180], "baseline": [161], "agents.": [162], "Furthermore,": [163], "demonstrates": [165], "stronger": [166], "causal": [167], "reasoning": [168], "ability,": [169], "improving": [170, 212], "8.8%": [175], "causality": [178], "GPT-4": [182], "Program-of-Thoughts": [184], "prompts.": [185], "work": [187], "underscores": [188], "developing": [192], "strategies": [194], "accumulating": [196], "utilizing": [198], "knowledge": [199], "inference,": [201], "mirroring": [202], "human": [204], "process": [206], "opening": [208], "new": [209], "avenues": [210], "curriculum-based": [216], "inference": [217], "optimization.": [218]}, "primary_topic": {"id": "https://openalex.org/T10260", "display_name": "Software Engineering Research", "score": 0.2671000063419342, "subfield": {"id": "https://openalex.org/subfields/1710", "display_name": "Information Systems"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, "topics": [{"id": "https://openalex.org/T10260", "display_name": "Software Engineering Research", "score": 0.2671000063419342, "subfield": {"id": "https://openalex.org/subfields/1710", "display_name": "Information Systems"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, {"id": "https://openalex.org/T10028", "display_name": "Topic Modeling", "score": 0.2199999988079071, "subfield": {"id": "https://openalex.org/subfields/1702", "display_name": "Artificial Intelligence"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, {"id": "https://openalex.org/T10906", "display_name": "AI-based Problem Solving and Planning", "score": 0.15360000729560852, "subfield": {"id": "https://openalex.org/subfields/1702", "display_name": "Artificial Intelligence"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}], "primary_location": {"id": "pmh:oai:arXiv.org:2505.14163", "is_oa": true, "landing_page_url": "http://arxiv.org/abs/2505.14163", "pdf_url": "https://arxiv.org/pdf/2505.14163", "source": {"id": "https://openalex.org/S4306400194", "display_name": "arXiv (Cornell University)", "issn_l": "2331-8422", "issn": ["2331-8422"], "is_oa": true, "is_in_doaj": false, "is_core": false, "listed_in": [], "host_organization": "https://openalex.org/I205783295", "host_organization_name": "Cornell University", "host_organization_lineage": ["https://openalex.org/I205783295"], "host_organization_lineage_names": [], "type": "repository"}, "license": null, "license_id": null, "version": "submittedVersion", "is_accepted": false, "is_published": false, "raw_source_name": null, "raw_type": "text"}, "cited_by_count": 0, "is_retracted": false}, {"id": "https://openalex.org/W4416258732", "doi": "https://doi.org/10.48550/arxiv.2508.08256", "title": "FIER: Fine-Grained and Efficient KV Cache Retrieval for Long-context LLM Inference", "publication_year": 2025, "publication_date": "2025-05-28", "type": "preprint", "language": "en", "authorships": [{"author_position": "first", "author": {"id": "https://openalex.org/A5100735740", "display_name": "Dongwei Wang", "orcid": "https://orcid.org/0000-0002-0943-0088", "observed_orcids": ["https://orcid.org/0000-0002-0943-0088"]}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Wang, Dongwei", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5101923344", "display_name": "Zijie Liu", "orcid": "https://orcid.org/0009-0002-4288-4388", "observed_orcids": ["https://orcid.org/0009-0002-4288-4388"]}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Liu, Zijie", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5102991129", "display_name": "Song Wang", "orcid": "https://orcid.org/0000-0001-7986-9806", "observed_orcids": ["https://orcid.org/0000-0001-7986-9806"]}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Wang, Song", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5102904368", "display_name": "Yu\u2010Xin Ren", "orcid": "https://orcid.org/0000-0002-3047-4923", "observed_orcids": ["https://orcid.org/0000-0002-3047-4923"]}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Ren, Yuxin", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5018946131", "display_name": "Jianing Deng", "orcid": "https://orcid.org/0000-0001-7307-6629", "observed_orcids": ["https://orcid.org/0000-0001-7307-6629"]}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Deng, Jianing", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5066534595", "display_name": "Jingtong Hu", "orcid": "https://orcid.org/0000-0003-4029-4034", "observed_orcids": ["https://orcid.org/0000-0003-4029-4034"]}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Hu, Jingtong", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5017041036", "display_name": "Tianlong Chen", "orcid": "https://orcid.org/0009-0009-1397-9266", "observed_orcids": ["https://orcid.org/0009-0009-1397-9266"]}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Chen, Tianlong", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "last", "author": {"id": "https://openalex.org/A5076154259", "display_name": "Huanrui Yang", "orcid": "https://orcid.org/0000-0002-3384-4512", "observed_orcids": ["https://orcid.org/0000-0002-3384-4512"]}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Yang, Huanrui", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}], "abstract_inverted_index": {"The": [0], "Key-Value": [1], "(KV)": [2], "cache": [3, 30, 55, 101, 135], "reading": [4], "latency": [5, 143], "increases": [6], "significantly": [7], "with": [8], "context": [9], "lengths,": [10], "hindering": [11], "the": [12, 68, 111, 134], "efficiency": [13], "of": [14, 28, 113, 133], "long-context": [15, 139], "LLM": [16], "inference.": [17], "To": [18], "address": [19], "this,": [20], "previous": [21], "works": [22], "propose": [23, 94], "retaining": [24], "a": [25, 96], "small": [26], "fraction": [27], "KV": [29, 37, 46, 76, 100, 128], "based": [31], "on": [32], "token": [33], "importance.": [34], "For": [35], "example,": [36], "eviction": [38], "uses": [39, 105], "static": [40], "heuristics": [41], "to": [42, 109, 146], "retain": [43], "tokens,": [44], "while": [45], "retrieval": [47, 77], "dynamically": [48], "selects": [49], "query-relevant": [50], "tokens": [51, 62, 85], "for": [52], "more": [53], "adaptive": [54], "management.": [56], "However,": [57], "we": [58, 93], "observe": [59], "that": [60, 124], "important": [61], "are": [63], "often": [64], "sparsely": [65], "distributed": [66], "across": [67, 137], "long": [69], "context.": [70], "This": [71], "sparsity": [72], "makes": [73], "existing": [74], "page-level": [75], "inaccurate,": [78], "as": [79], "each": [80, 114], "page": [81], "may": [82], "include": [83], "irrelevant": [84], "and": [86, 98, 119], "miss": [87], "critical": [88], "ones.": [89], "In": [90], "this": [91], "work,": [92], "Fier,": [95], "\\underline{Fi}ne-Grained": [97], "\\underline{E}fficient": [99], "\\underline{R}etrieval": [102], "method.": [103], "Fier": [104, 125], "1-bit": [106], "quantized": [107], "keys": [108], "estimate": [110], "importance": [112], "token,": [115], "resulting": [116], "in": [117], "efficient": [118], "precise": [120], "retrieval.": [121], "Experiments": [122], "show": [123], "matches": [126], "full": [127], "performance": [129], "using": [130], "only": [131], "11\\%": [132], "budget": [136], "various": [138], "tasks,": [140], "reducing": [141], "decoding": [142], "by": [144], "1.2$\\times$": [145], "1.5$\\times$.Code": [147], "is": [148], "available": [149], "at": [150], "https://github.com/SimWangArizona/FIER": [151]}, "primary_topic": {"id": "https://openalex.org/T10181", "display_name": "Natural Language Processing Techniques", "score": 0.3434999883174896, "subfield": {"id": "https://openalex.org/subfields/1702", "display_name": "Artificial Intelligence"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, "topics": [{"id": "https://openalex.org/T10181", "display_name": "Natural Language Processing Techniques", "score": 0.3434999883174896, "subfield": {"id": "https://openalex.org/subfields/1702", "display_name": "Artificial Intelligence"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, {"id": "https://openalex.org/T10028", "display_name": "Topic Modeling", "score": 0.34200000762939453, "subfield": {"id": "https://openalex.org/subfields/1702", "display_name": "Artificial Intelligence"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, {"id": "https://openalex.org/T10036", "display_name": "Advanced Neural Network Applications", "score": 0.10279999673366547, "subfield": {"id": "https://openalex.org/subfields/1707", "display_name": "Computer Vision and Pattern Recognition"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}], "primary_location": {"id": "pmh:oai:arXiv.org:2508.08256", "is_oa": true, "landing_page_url": "http://arxiv.org/abs/2508.08256", "pdf_url": "https://arxiv.org/pdf/2508.08256", "source": {"id": "https://openalex.org/S4306400194", "display_name": "arXiv (Cornell University)", "issn_l": "2331-8422", "issn": ["2331-8422"], "is_oa": true, "is_in_doaj": false, "is_core": false, "listed_in": [], "host_organization": "https://openalex.org/I205783295", "host_organization_name": "Cornell University", "host_organization_lineage": ["https://openalex.org/I205783295"], "host_organization_lineage_names": [], "type": "repository"}, "license": null, "license_id": null, "version": "submittedVersion", "is_accepted": false, "is_published": false, "raw_source_name": null, "raw_type": "text"}, "cited_by_count": 0, "is_retracted": false}, {"id": "https://openalex.org/W7106253752", "doi": "https://doi.org/10.48550/arxiv.2511.15927", "title": "DiffuMamba: High-Throughput Diffusion LMs with Mamba Backbone", "publication_year": 2025, "publication_date": "2025-11-19", "type": "preprint", "language": null, "authorships": [{"author_position": "first", "author": {"id": null, "display_name": "Singh, Vaibhav", "orcid": null, "observed_orcids": []}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Singh, Vaibhav", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": null, "display_name": "Ostapenko, Oleksiy", "orcid": null, "observed_orcids": []}, "institutions": [{"id": "https://openalex.org/I4210167039", "display_name": "ServiceNow (United States)", "ror": "https://ror.org/05xr0bc04", "country_code": "US", "type": "company", "lineage": ["https://openalex.org/I4210167039"]}, {"id": "https://openalex.org/I60158472", "display_name": "Concordia University", "ror": "https://ror.org/0420zvk78", "country_code": "CA", "type": "education", "lineage": ["https://openalex.org/I60158472"]}], "countries": ["CA", "US"], "is_corresponding": false, "raw_author_name": "Ostapenko, Oleksiy", "raw_affiliation_strings": ["ServiceNow Research , Montreal , Canada", "ServiceNow Research"], "raw_orcid": null, "affiliations": [{"raw_affiliation_string": "ServiceNow Research , Montreal , Canada", "institution_ids": ["https://openalex.org/I60158472"]}, {"raw_affiliation_string": "ServiceNow Research", "institution_ids": ["https://openalex.org/I4210167039"]}]}, {"author_position": "middle", "author": {"id": null, "display_name": "No\u00ebl, Pierre-Andr\u00e9", "orcid": null, "observed_orcids": []}, "institutions": [{"id": "https://openalex.org/I4210167039", "display_name": "ServiceNow (United States)", "ror": "https://ror.org/05xr0bc04", "country_code": "US", "type": "company", "lineage": ["https://openalex.org/I4210167039"]}], "countries": ["US"], "is_corresponding": false, "raw_author_name": "No\u00ebl, Pierre-Andr\u00e9", "raw_affiliation_strings": ["ServiceNow Research"], "raw_orcid": null, "affiliations": [{"raw_affiliation_string": "ServiceNow Research", "institution_ids": ["https://openalex.org/I4210167039"]}]}, {"author_position": "middle", "author": {"id": null, "display_name": "Belilovsky, Eugene", "orcid": null, "observed_orcids": []}, "institutions": [{"id": "https://openalex.org/I4210164802", "display_name": "Mila - Quebec Artificial Intelligence Institute", "ror": "https://ror.org/05c22rx21", "country_code": "CA", "type": "facility", "lineage": ["https://openalex.org/I4210164802"]}, {"id": "https://openalex.org/I60158472", "display_name": "Concordia University", "ror": "https://ror.org/0420zvk78", "country_code": "CA", "type": "education", "lineage": ["https://openalex.org/I60158472"]}], "countries": ["CA"], "is_corresponding": false, "raw_author_name": "Belilovsky, Eugene", "raw_affiliation_strings": ["Concordia University", "Mila"], "raw_orcid": null, "affiliations": [{"raw_affiliation_string": "Concordia University", "institution_ids": ["https://openalex.org/I60158472"]}, {"raw_affiliation_string": "Mila", "institution_ids": ["https://openalex.org/I4210164802"]}]}, {"author_position": "last", "author": {"id": null, "display_name": "Scholak, Torsten", "orcid": null, "observed_orcids": []}, "institutions": [{"id": "https://openalex.org/I4210167039", "display_name": "ServiceNow (United States)", "ror": "https://ror.org/05xr0bc04", "country_code": "US", "type": "company", "lineage": ["https://openalex.org/I4210167039"]}], "countries": ["US"], "is_corresponding": false, "raw_author_name": "Scholak, Torsten", "raw_affiliation_strings": ["ServiceNow Research"], "raw_orcid": null, "affiliations": [{"raw_affiliation_string": "ServiceNow Research", "institution_ids": ["https://openalex.org/I4210167039"]}]}], "abstract_inverted_index": {"Diffusion": [0], "language": [1, 36], "models": [2, 68], "(DLMs)": [3], "have": [4], "emerged": [5], "as": [6, 116], "a": [7, 33, 40, 55, 92, 135], "promising": [8, 136], "alternative": [9], "to": [10, 24, 64, 78], "autoregressive": [11], "(AR)": [12], "generation,": [13], "yet": [14], "their": [15], "reliance": [16], "on": [17, 39, 86], "Transformer": [18], "backbones": [19], "limits": [20], "inference": [21, 83, 96], "efficiency": [22, 97], "due": [23], "quadratic": [25], "attention": [26], "or": [27], "KV-cache": [28], "overhead.": [29], "We": [30, 89], "introduce": [31], "DiffuMamba,": [32], "masked": [34], "diffusion": [35, 47, 71, 111], "model": [37], "built": [38], "bidirectional": [41], "Mamba": [42, 113], "backbone": [43], "that": [44, 120], "combines": [45], "the": [46, 117, 128], "objective": [48], "with": [49, 58, 105, 112, 123], "linear-time": [50], "sequence": [51, 124], "modeling,": [52], "and": [53, 80, 126], "DiffuMamba-H,": [54], "hybrid": [56], "variant": [57], "interleaved": [59], "attention.": [60], "Across": [61], "scales": [62, 121], "up": [63, 77], "1.3B": [65], "parameters,": [66], "our": [67], "match": [69], "Transformer-based": [70], "in": [72], "downstream": [73], "performance": [74, 130], "while": [75], "achieving": [76], "8.2x": [79], "4.3x": [81], "higher": [82], "throughput,": [84], "respectively,": [85], "long": [87], "sequences.": [88], "further": [90], "present": [91], "systematic": [93], "analysis": [94], "of": [95], "across": [98, 131], "modern": [99], "DLM": [100], "variants": [101], "combining": [102], "asymptotic": [103], "complexity": [104], "empirical": [106], "measurements.": [107], "Notably,": [108], "cache-efficient": [109], "block": [110], "mixers": [114], "emerges": [115], "only": [118], "strategy": [119], "linearly": [122], "length": [125], "achieves": [127], "strongest": [129], "all": [132], "baselines,": [133], "suggesting": [134], "direction": [137], "for": [138], "future": [139], "diffusion-based": [140], "generation": [141], "systems.": [142]}, "primary_topic": {"id": "https://openalex.org/T10028", "display_name": "Topic Modeling", "score": 0.6011000275611877, "subfield": {"id": "https://openalex.org/subfields/1702", "display_name": "Artificial Intelligence"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, "topics": [{"id": "https://openalex.org/T10028", "display_name": "Topic Modeling", "score": 0.6011000275611877, "subfield": {"id": "https://openalex.org/subfields/1702", "display_name": "Artificial Intelligence"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, {"id": "https://openalex.org/T10181", "display_name": "Natural Language Processing Techniques", "score": 0.3529999852180481, "subfield": {"id": "https://openalex.org/subfields/1702", "display_name": "Artificial Intelligence"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, {"id": "https://openalex.org/T10775", "display_name": "Generative Adversarial Networks and Image Synthesis", "score": 0.008500000461935997, "subfield": {"id": "https://openalex.org/subfields/1707", "display_name": "Computer Vision and Pattern Recognition"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}], "primary_location": {"id": "pmh:oai:arXiv.org:2511.15927", "is_oa": true, "landing_page_url": "https://arxiv.org/abs/2511.15927", "pdf_url": "https://arxiv.org/pdf/2511.15927", "source": {"id": "https://openalex.org/S4306400194", "display_name": "arXiv (Cornell University)", "issn_l": "2331-8422", "issn": ["2331-8422"], "is_oa": true, "is_in_doaj": false, "is_core": false, "listed_in": [], "host_organization": "https://openalex.org/I205783295", "host_organization_name": "Cornell University", "host_organization_lineage": ["https://openalex.org/I205783295"], "host_organization_lineage_names": [], "type": "repository"}, "license": "cc-by", "license_id": "https://openalex.org/licenses/cc-by", "version": "submittedVersion", "is_accepted": false, "is_published": false, "raw_source_name": null, "raw_type": "text"}, "cited_by_count": 0, "is_retracted": false}, {"id": "https://openalex.org/W7106504155", "doi": "https://doi.org/10.5281/zenodo.17694277", "title": "Self-Traversed Reasoning: A Zero-Cost Anti-Model Protocol for Transformer Inference Optimization", "publication_year": 2025, "publication_date": "2025-11-24", "type": "preprint", "language": null, "authorships": [{"author_position": "first", "author": {"id": null, "display_name": "Wippersberger, Kevin - Lee", "orcid": null, "observed_orcids": []}, "institutions": [], "countries": [], "is_corresponding": true, "raw_author_name": "Wippersberger, Kevin - Lee", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}], "abstract_inverted_index": {"This": [0, 104], "paper": [1], "introduces": [2], "Self-Traversed": [3], "Reasoning": [4], "(STR),": [5], "a": [6, 52, 124], "zero-cost": [7], "inference": [8], "optimization": [9], "protocol": [10, 105], "for": [11, 129], "transformer": [12], "models.": [13, 132], "STR": [14, 41, 88], "transforms": [15], "standard": [16], "transformers": [17, 107], "into": [18, 123], "self-auditing": [19], "systems": [20], "by": [21], "recovering": [22], "already-computed": [23], "probability": [24], "distributions": [25], "that": [26], "are": [27], "normally": [28], "discarded": [29, 121], "during": [30], "token": [31], "generation.": [32], "Instead": [33], "of": [34, 38, 95], "throwing": [35], "away": [36], "99.9%": [37], "computed": [39], "information,": [40], "logs": [42], "alternative": [43], "paths,": [44], "rejected": [45], "tokens,": [46], "and": [47, 78, 114], "entropy": [48], "measures": [49], "to": [50, 108], "create": [51], "complete": [53], "decision": [54], "map.": [55], "The": [56], "framework": [57], "achieves": [58], "four": [59], "key": [60], "benefits:": [61], "enhanced": [62], "output": [63], "quality": [64, 126], "through": [65], "internal": [66], "validation,": [67], "zero": [68], "additional": [69], "computational": [70], "cost,": [71], "continuous": [72], "self-improvement": [73], "via": [74], "aggregated": [75], "failure": [76], "modes,": [77], "inherent": [79], "uncertainty": [80], "calibration.": [81], "Grounded": [82], "in": [83], "the": [84, 93], "Universal": [85], "Meaning": [86], "Equation,": [87], "defines": [89], "traversal": [90], "meaning": [91], "as": [92], "reconciliation": [94], "primary": [96], "paths": [97], "against": [98], "compressed": [99], "alternatives": [100], "over": [101], "divergence": [102], "entropy.": [103], "enables": [106], "explain": [109], "their": [110], "reasoning,": [111], "quantify": [112], "confidence,": [113], "self-correct": [115], "without": [116], "any": [117], "extra": [118], "GPU": [119], "cycles\u2014turning": [120], "computations": [122], "powerful": [125], "assurance": [127], "mechanism": [128], "large": [130], "language": [131]}, "primary_topic": {"id": "https://openalex.org/T12026", "display_name": "Explainable Artificial Intelligence (XAI)", "score": 0.6352999806404114, "subfield": {"id": "https://openalex.org/subfields/1702", "display_name": "Artificial Intelligence"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, "topics": [{"id": "https://openalex.org/T12026", "display_name": "Explainable Artificial Intelligence (XAI)", "score": 0.6352999806404114, "subfield": {"id": "https://openalex.org/subfields/1702", "display_name": "Artificial Intelligence"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, {"id": "https://openalex.org/T10028", "display_name": "Topic Modeling", "score": 0.267300009727478, "subfield": {"id": "https://openalex.org/subfields/1702", "display_name": "Artificial Intelligence"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, {"id": "https://openalex.org/T10181", "display_name": "Natural Language Processing Techniques", "score": 0.06989999860525131, "subfield": {"id": "https://openalex.org/subfields/1702", "display_name": "Artificial Intelligence"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}], "primary_location": {"id": "doi:10.5281/zenodo.17694277", "is_oa": true, "landing_page_url": "https://doi.org/10.5281/zenodo.17694277", "pdf_url": null, "source": {"id": "https://openalex.org/S4306400562", "display_name": "Zenodo (CERN European Organization for Nuclear Research)", "issn_l": null, "issn": null, "is_oa": true, "is_in_doaj": false, "is_core": false, "listed_in": [], "host_organization": "https://openalex.org/I67311998", "host_organization_name": "European Organization for Nuclear Research", "host_organization_lineage": ["https://openalex.org/I67311998"], "host_organization_lineage_names": [], "type": "repository"}, "license": "cc-by", "license_id": "https://openalex.org/licenses/cc-by", "version": null, "is_accepted": false, "is_published": false, "raw_source_name": null, "raw_type": "Preprint"}, "cited_by_count": 0, "is_retracted": false}, {"id": "https://openalex.org/W7117174144", "doi": null, "title": "MixKVQ: Query-Aware Mixed-Precision KV Cache Quantization for Long-Context Reasoning", "publication_year": 2025, "publication_date": "2025-12-22", "type": "preprint", "language": null, "authorships": [{"author_position": "first", "author": {"id": "https://openalex.org/A5121175017", "display_name": "Tao Zhang", "orcid": null, "observed_orcids": []}, "institutions": [{"id": "https://openalex.org/I90610280", "display_name": "South China University of Technology", "ror": "https://ror.org/0530pts50", "country_code": "CN", "type": "education", "lineage": ["https://openalex.org/I90610280"]}], "countries": ["CN"], "is_corresponding": false, "raw_author_name": "Zhang, Tao", "raw_affiliation_strings": ["South China University of Technology , China,"], "raw_orcid": null, "affiliations": [{"raw_affiliation_string": "South China University of Technology , China,", "institution_ids": ["https://openalex.org/I90610280"]}]}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5101403806", "display_name": "Ziqian Zeng", "orcid": "https://orcid.org/0000-0003-0060-7956", "observed_orcids": ["https://orcid.org/0000-0003-0060-7956"]}, "institutions": [{"id": "https://openalex.org/I90610280", "display_name": "South China University of Technology", "ror": "https://ror.org/0530pts50", "country_code": "CN", "type": "education", "lineage": ["https://openalex.org/I90610280"]}], "countries": ["CN"], "is_corresponding": false, "raw_author_name": "Zeng, Ziqian", "raw_affiliation_strings": ["South China University of Technology , China,"], "raw_orcid": null, "affiliations": [{"raw_affiliation_string": "South China University of Technology , China,", "institution_ids": ["https://openalex.org/I90610280"]}]}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5121131136", "display_name": "Hao Peng", "orcid": "https://orcid.org/0009-0006-5203-5936", "observed_orcids": ["https://orcid.org/0009-0006-5203-5936"]}, "institutions": [{"id": "https://openalex.org/I82880672", "display_name": "Beihang University", "ror": "https://ror.org/00wk2mp56", "country_code": "CN", "type": "education", "lineage": ["https://openalex.org/I82880672"]}], "countries": ["CN"], "is_corresponding": false, "raw_author_name": "Peng, Hao", "raw_affiliation_strings": ["Beihang University , China,"], "raw_orcid": null, "affiliations": [{"raw_affiliation_string": "Beihang University , China,", "institution_ids": ["https://openalex.org/I82880672"]}]}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5121162302", "display_name": "Huiping Zhuang", "orcid": null, "observed_orcids": []}, "institutions": [{"id": "https://openalex.org/I90610280", "display_name": "South China University of Technology", "ror": "https://ror.org/0530pts50", "country_code": "CN", "type": "education", "lineage": ["https://openalex.org/I90610280"]}], "countries": ["CN"], "is_corresponding": false, "raw_author_name": "Zhuang, Huiping", "raw_affiliation_strings": ["South China University of Technology , China,"], "raw_orcid": null, "affiliations": [{"raw_affiliation_string": "South China University of Technology , China,", "institution_ids": ["https://openalex.org/I90610280"]}]}, {"author_position": "last", "author": {"id": "https://openalex.org/A5121167888", "display_name": "Cen Chen", "orcid": null, "observed_orcids": []}, "institutions": [{"id": "https://openalex.org/I90610280", "display_name": "South China University of Technology", "ror": "https://ror.org/0530pts50", "country_code": "CN", "type": "education", "lineage": ["https://openalex.org/I90610280"]}], "countries": ["CN"], "is_corresponding": false, "raw_author_name": "Chen, Cen", "raw_affiliation_strings": ["Pazhou Laboratory , China ,", "South China University of Technology , China,"], "raw_orcid": null, "affiliations": [{"raw_affiliation_string": "Pazhou Laboratory , China ,", "institution_ids": []}, {"raw_affiliation_string": "South China University of Technology , China,", "institution_ids": ["https://openalex.org/I90610280"]}]}], "abstract_inverted_index": {"Long": [0], "Chain-of-Thought": [1], "(CoT)": [2], "reasoning": [3, 51, 140], "has": [4], "significantly": [5, 146], "advanced": [6], "the": [7, 26, 61, 100], "capabilities": [8], "of": [9], "Large": [10], "Language": [11], "Models": [12], "(LLMs),": [13], "but": [14], "this": [15, 104], "progress": [16], "is": [17, 35], "accompanied": [18], "by": [19], "substantial": [20], "memory": [21, 162], "and": [22, 96, 121], "latency": [23], "overhead": [24], "from": [25], "extensive": [27], "Key-Value": [28], "(KV)": [29], "cache.": [30, 136], "Although": [31], "KV": [32, 82], "cache": [33, 83], "quantization": [34, 42, 54, 84, 94, 133], "a": [36, 90, 109, 115, 155, 159], "promising": [37], "compression": [38], "technique,": [39], "existing": [40, 148], "low-bit": [41, 81, 149], "methods": [43], "often": [44], "exhibit": [45], "severe": [46], "performance": [47, 152], "degradation": [48], "on": [49, 103, 138], "complex": [50, 139], "tasks.": [52], "Fixed-precision": [53], "struggles": [55], "to": [56, 69, 99, 119, 154], "handle": [57], "outlier": [58], "channels": [59, 125], "in": [60], "key": [62, 91, 124], "cache,": [63], "while": [64, 130], "current": [65], "mixed-precision": [66], "strategies": [67], "fail": [68], "accurately": [70], "identify": [71, 120], "components": [72], "requiring": [73], "high-precision": [74], "representation.": [75], "We": [76], "find": [77], "that": [78, 113, 126, 143], "an": [79], "effective": [80], "strategy": [85], "must": [86], "consider": [87], "two": [88], "factors:": [89], "channel's": [92], "intrinsic": [93], "difficulty": [95], "its": [97], "relevance": [98], "query.": [101], "Based": [102], "insight,": [105], "we": [106], "propose": [107], "MixKVQ,": [108], "novel": [110], "plug-and-play": [111], "method": [112], "introduces": [114], "lightweight,": [116], "query-aware": [117], "algorithm": [118], "preserve": [122], "critical": [123], "need": [127], "higher": [128], "precision,": [129], "applying": [131], "per-token": [132], "for": [134], "value": [135], "Experiments": [137], "datasets": [141], "demonstrate": [142], "our": [144], "approach": [145], "outperforms": [147], "methods,": [150], "achieving": [151], "comparable": [153], "full-precision": [156], "baseline": [157], "at": [158], "substantially": [160], "reduced": [161], "footprint.": [163]}, "primary_topic": {"id": "https://openalex.org/T10036", "display_name": "Advanced Neural Network Applications", "score": 0.8672000169754028, "subfield": {"id": "https://openalex.org/subfields/1707", "display_name": "Computer Vision and Pattern Recognition"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, "topics": [{"id": "https://openalex.org/T10036", "display_name": "Advanced Neural Network Applications", "score": 0.8672000169754028, "subfield": {"id": "https://openalex.org/subfields/1707", "display_name": "Computer Vision and Pattern Recognition"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, {"id": "https://openalex.org/T10028", "display_name": "Topic Modeling", "score": 0.0478999987244606, "subfield": {"id": "https://openalex.org/subfields/1702", "display_name": "Artificial Intelligence"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, {"id": "https://openalex.org/T10181", "display_name": "Natural Language Processing Techniques", "score": 0.012900000438094139, "subfield": {"id": "https://openalex.org/subfields/1702", "display_name": "Artificial Intelligence"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}], "primary_location": {"id": "pmh:oai:arXiv.org:2512.19206", "is_oa": true, "landing_page_url": "http://arxiv.org/abs/2512.19206", "pdf_url": "https://arxiv.org/pdf/2512.19206", "source": {"id": "https://openalex.org/S4306400194", "display_name": "arXiv (Cornell University)", "issn_l": "2331-8422", "issn": ["2331-8422"], "is_oa": true, "is_in_doaj": false, "is_core": false, "listed_in": [], "host_organization": "https://openalex.org/I205783295", "host_organization_name": "Cornell University", "host_organization_lineage": ["https://openalex.org/I205783295"], "host_organization_lineage_names": [], "type": "repository"}, "license": "cc-by", "license_id": "https://openalex.org/licenses/cc-by", "version": "submittedVersion", "is_accepted": false, "is_published": false, "raw_source_name": null, "raw_type": "text"}, "cited_by_count": 0, "is_retracted": false}, {"id": "https://openalex.org/W7115596576", "doi": "https://doi.org/10.48550/arxiv.2512.11221", "title": "Adaptive Soft Rolling KV Freeze with Entropy-Guided Recovery: Sublinear Memory Growth for Efficient LLM Inference", "publication_year": 2025, "publication_date": "2025-12-12", "type": "preprint", "language": null, "authorships": [{"author_position": "first", "author": {"id": "https://openalex.org/A5130364278", "display_name": "Adilet Metinov", "orcid": "https://orcid.org/0009-0007-6400-4173", "observed_orcids": ["https://orcid.org/0009-0007-6400-4173"]}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Metinov, Adilet", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": null, "display_name": "Kudakeeva, Gulida M.", "orcid": null, "observed_orcids": []}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Kudakeeva, Gulida M.", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": null, "display_name": "Nursultan, Bolotbek uulu", "orcid": null, "observed_orcids": []}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Nursultan, Bolotbek uulu", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "last", "author": {"id": null, "display_name": "Kabaeva, Gulnara D.", "orcid": null, "observed_orcids": []}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Kabaeva, Gulnara D.", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}], "abstract_inverted_index": {"We": [0, 62], "present": [1], "Adaptive": [2], "Soft": [3], "Rolling": [4], "KV": [5, 92], "Freeze": [6], "with": [7, 66, 75], "Entropy-Guided": [8], "Recovery": [9], "(ASR-KF-EGR),": [10], "a": [11, 24, 39, 113], "training-free": [12], "inference-time": [13], "framework": [14, 65], "for": [15, 34, 116], "efficient": [16], "large": [17], "language": [18], "model": [19], "generation.": [20], "Our": [21], "method": [22, 105], "introduces": [23], "reversible": [25], "soft-freeze": [26], "mechanism": [27], "that": [28, 46], "temporarily": [29], "suspends": [30], "key-value": [31], "(KV)": [32], "updates": [33], "low-importance": [35, 77], "tokens": [36, 53], "identified": [37], "within": [38], "sliding": [40], "attention": [41], "window.": [42], "Unlike": [43], "eviction-based": [44], "approaches": [45], "permanently": [47], "discard": [48], "context,": [49], "ASR-KF-EGR": [50], "preserves": [51], "all": [52], "in": [54, 90], "off-GPU": [55], "storage": [56], "and": [57, 99, 111], "restores": [58], "them": [59], "on": [60, 84], "demand.": [61], "extend": [63], "the": [64], "sublinear": [67], "freeze": [68, 71], "scheduling,": [69], "where": [70], "duration": [72], "grows": [73], "sublinearly": [74], "repeated": [76], "detections,": [78], "preventing": [79], "over-aggressive": [80], "compression.": [81], "Preliminary": [82], "experiments": [83], "LLaMA-3": [85], "8B": [86], "demonstrate": [87], "55-67%": [88], "reduction": [89], "active": [91], "cache": [93], "size": [94], "while": [95], "maintaining": [96], "generation": [97], "quality": [98], "passing": [100], "needle-in-haystack": [101], "retrieval": [102], "tests.": [103], "The": [104], "is": [106], "architecture-agnostic,": [107], "requires": [108], "no": [109], "fine-tuning,": [110], "provides": [112], "practical": [114], "solution": [115], "memory-constrained": [117], "deployment": [118], "of": [119], "long-context": [120], "LLMs.": [121]}, "primary_topic": {"id": "https://openalex.org/T10028", "display_name": "Topic Modeling", "score": 0.5059000253677368, "subfield": {"id": "https://openalex.org/subfields/1702", "display_name": "Artificial Intelligence"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, "topics": [{"id": "https://openalex.org/T10028", "display_name": "Topic Modeling", "score": 0.5059000253677368, "subfield": {"id": "https://openalex.org/subfields/1702", "display_name": "Artificial Intelligence"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, {"id": "https://openalex.org/T10181", "display_name": "Natural Language Processing Techniques", "score": 0.3792000114917755, "subfield": {"id": "https://openalex.org/subfields/1702", "display_name": "Artificial Intelligence"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, {"id": "https://openalex.org/T10036", "display_name": "Advanced Neural Network Applications", "score": 0.034699998795986176, "subfield": {"id": "https://openalex.org/subfields/1707", "display_name": "Computer Vision and Pattern Recognition"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}], "primary_location": {"id": "pmh:oai:arXiv.org:2512.11221", "is_oa": true, "landing_page_url": "http://arxiv.org/abs/2512.11221", "pdf_url": "https://arxiv.org/pdf/2512.11221", "source": {"id": "https://openalex.org/S4306400194", "display_name": "arXiv (Cornell University)", "issn_l": "2331-8422", "issn": ["2331-8422"], "is_oa": true, "is_in_doaj": false, "is_core": false, "listed_in": [], "host_organization": "https://openalex.org/I205783295", "host_organization_name": "Cornell University", "host_organization_lineage": ["https://openalex.org/I205783295"], "host_organization_lineage_names": [], "type": "repository"}, "license": null, "license_id": null, "version": "submittedVersion", "is_accepted": false, "is_published": false, "raw_source_name": null, "raw_type": "text"}, "cited_by_count": 0, "is_retracted": false}, {"id": "https://openalex.org/W4414634795", "doi": "https://doi.org/10.48550/arxiv.2504.16083", "title": "MMInference: Accelerating Pre-filling for Long-Context VLMs via Modality-Aware Permutation Sparse Attention", "publication_year": 2025, "publication_date": "2025-04-22", "type": "preprint", "language": "en", "authorships": [{"author_position": "first", "author": {"id": "https://openalex.org/A5100381401", "display_name": "Yucheng Li", "orcid": "https://orcid.org/0000-0003-3473-3176", "observed_orcids": ["https://orcid.org/0000-0003-3473-3176"]}, "institutions": [{"id": "https://openalex.org/I28290843", "display_name": "University of Surrey", "ror": "https://ror.org/00ks66431", "country_code": "GB", "type": "education", "lineage": ["https://openalex.org/I28290843"]}], "countries": ["GB"], "is_corresponding": false, "raw_author_name": "Li, Yucheng", "raw_affiliation_strings": ["University of Surrey"], "raw_orcid": null, "affiliations": [{"raw_affiliation_string": "University of Surrey", "institution_ids": ["https://openalex.org/I28290843"]}]}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5070156365", "display_name": "Huiqiang Jiang", "orcid": "https://orcid.org/0000-0002-1327-4882", "observed_orcids": ["https://orcid.org/0000-0002-1327-4882"]}, "institutions": [{"id": "https://openalex.org/I4210105678", "display_name": "Microsoft (Finland)", "ror": "https://ror.org/01nehjf29", "country_code": "FI", "type": "company", "lineage": ["https://openalex.org/I1290206253", "https://openalex.org/I4210105678"]}], "countries": ["FI"], "is_corresponding": false, "raw_author_name": "Jiang, Huiqiang", "raw_affiliation_strings": ["Microsoft Corporation "], "raw_orcid": null, "affiliations": [{"raw_affiliation_string": "Microsoft Corporation ", "institution_ids": ["https://openalex.org/I4210105678"]}]}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5040430995", "display_name": "Chengruidong Zhang", "orcid": "https://orcid.org/0009-0002-4939-0436", "observed_orcids": ["https://orcid.org/0009-0002-4939-0436"]}, "institutions": [{"id": "https://openalex.org/I4210105678", "display_name": "Microsoft (Finland)", "ror": "https://ror.org/01nehjf29", "country_code": "FI", "type": "company", "lineage": ["https://openalex.org/I1290206253", "https://openalex.org/I4210105678"]}], "countries": ["FI"], "is_corresponding": false, "raw_author_name": "Zhang, Chengruidong", "raw_affiliation_strings": ["Microsoft Corporation "], "raw_orcid": null, "affiliations": [{"raw_affiliation_string": "Microsoft Corporation ", "institution_ids": ["https://openalex.org/I4210105678"]}]}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5024011052", "display_name": "Qianhui Wu", "orcid": "https://orcid.org/0000-0001-9146-0675", "observed_orcids": ["https://orcid.org/0000-0001-9146-0675"]}, "institutions": [{"id": "https://openalex.org/I4210105678", "display_name": "Microsoft (Finland)", "ror": "https://ror.org/01nehjf29", "country_code": "FI", "type": "company", "lineage": ["https://openalex.org/I1290206253", "https://openalex.org/I4210105678"]}], "countries": ["FI"], "is_corresponding": false, "raw_author_name": "Wu, Qianhui", "raw_affiliation_strings": ["Microsoft Corporation "], "raw_orcid": null, "affiliations": [{"raw_affiliation_string": "Microsoft Corporation ", "institution_ids": ["https://openalex.org/I4210105678"]}]}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5100929082", "display_name": "Xufang Luo", "orcid": "https://orcid.org/0000-0002-3405-554X", "observed_orcids": ["https://orcid.org/0000-0002-3405-554X"]}, "institutions": [{"id": "https://openalex.org/I4210105678", "display_name": "Microsoft (Finland)", "ror": "https://ror.org/01nehjf29", "country_code": "FI", "type": "company", "lineage": ["https://openalex.org/I1290206253", "https://openalex.org/I4210105678"]}], "countries": ["FI"], "is_corresponding": false, "raw_author_name": "Luo, Xufang", "raw_affiliation_strings": ["Microsoft Corporation "], "raw_orcid": null, "affiliations": [{"raw_affiliation_string": "Microsoft Corporation ", "institution_ids": ["https://openalex.org/I4210105678"]}]}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5083787197", "display_name": "Surin Ahn", "orcid": "https://orcid.org/0000-0001-5615-4859", "observed_orcids": ["https://orcid.org/0000-0001-5615-4859"]}, "institutions": [{"id": "https://openalex.org/I4210105678", "display_name": "Microsoft (Finland)", "ror": "https://ror.org/01nehjf29", "country_code": "FI", "type": "company", "lineage": ["https://openalex.org/I1290206253", "https://openalex.org/I4210105678"]}], "countries": ["FI"], "is_corresponding": false, "raw_author_name": "Ahn, Surin", "raw_affiliation_strings": ["Microsoft Corporation "], "raw_orcid": null, "affiliations": [{"raw_affiliation_string": "Microsoft Corporation ", "institution_ids": ["https://openalex.org/I4210105678"]}]}, {"author_position": "middle", "author": {"id": null, "display_name": "Abdi, Amir H.", "orcid": null, "observed_orcids": []}, "institutions": [{"id": "https://openalex.org/I4210105678", "display_name": "Microsoft (Finland)", "ror": "https://ror.org/01nehjf29", "country_code": "FI", "type": "company", "lineage": ["https://openalex.org/I1290206253", "https://openalex.org/I4210105678"]}], "countries": ["FI"], "is_corresponding": false, "raw_author_name": "Abdi, Amir H.", "raw_affiliation_strings": ["Microsoft Corporation "], "raw_orcid": null, "affiliations": [{"raw_affiliation_string": "Microsoft Corporation ", "institution_ids": ["https://openalex.org/I4210105678"]}]}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5100440904", "display_name": "Dongsheng Li", "orcid": "https://orcid.org/0000-0001-9962-1373", "observed_orcids": ["https://orcid.org/0000-0001-9962-1373"]}, "institutions": [{"id": "https://openalex.org/I4210105678", "display_name": "Microsoft (Finland)", "ror": "https://ror.org/01nehjf29", "country_code": "FI", "type": "company", "lineage": ["https://openalex.org/I1290206253", "https://openalex.org/I4210105678"]}], "countries": ["FI"], "is_corresponding": false, "raw_author_name": "Li, Dongsheng", "raw_affiliation_strings": ["Microsoft Corporation "], "raw_orcid": null, "affiliations": [{"raw_affiliation_string": "Microsoft Corporation ", "institution_ids": ["https://openalex.org/I4210105678"]}]}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5091035881", "display_name": "Jianfeng Gao", "orcid": "https://orcid.org/0000-0001-5540-7228", "observed_orcids": ["https://orcid.org/0000-0001-5540-7228"]}, "institutions": [{"id": "https://openalex.org/I4210105678", "display_name": "Microsoft (Finland)", "ror": "https://ror.org/01nehjf29", "country_code": "FI", "type": "company", "lineage": ["https://openalex.org/I1290206253", "https://openalex.org/I4210105678"]}], "countries": ["FI"], "is_corresponding": false, "raw_author_name": "Gao, Jianfeng", "raw_affiliation_strings": ["Microsoft Corporation "], "raw_orcid": null, "affiliations": [{"raw_affiliation_string": "Microsoft Corporation ", "institution_ids": ["https://openalex.org/I4210105678"]}]}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5118093092", "display_name": "Yuqing Yang", "orcid": "https://orcid.org/0000-0001-8182-1350", "observed_orcids": ["https://orcid.org/0000-0001-8182-1350"]}, "institutions": [{"id": "https://openalex.org/I4210105678", "display_name": "Microsoft (Finland)", "ror": "https://ror.org/01nehjf29", "country_code": "FI", "type": "company", "lineage": ["https://openalex.org/I1290206253", "https://openalex.org/I4210105678"]}], "countries": ["FI"], "is_corresponding": false, "raw_author_name": "Yang, Yuqing", "raw_affiliation_strings": ["Microsoft Corporation "], "raw_orcid": null, "affiliations": [{"raw_affiliation_string": "Microsoft Corporation ", "institution_ids": ["https://openalex.org/I4210105678"]}]}, {"author_position": "last", "author": {"id": "https://openalex.org/A5032760128", "display_name": "Lili Qiu", "orcid": "https://orcid.org/0000-0002-1590-9749", "observed_orcids": ["https://orcid.org/0000-0002-1590-9749", "https://orcid.org/0009-0003-8131-7439"]}, "institutions": [{"id": "https://openalex.org/I4210105678", "display_name": "Microsoft (Finland)", "ror": "https://ror.org/01nehjf29", "country_code": "FI", "type": "company", "lineage": ["https://openalex.org/I1290206253", "https://openalex.org/I4210105678"]}], "countries": ["FI"], "is_corresponding": false, "raw_author_name": "Qiu, Lili", "raw_affiliation_strings": ["Microsoft Corporation "], "raw_orcid": null, "affiliations": [{"raw_affiliation_string": "Microsoft Corporation ", "institution_ids": ["https://openalex.org/I4210105678"]}]}], "abstract_inverted_index": {"The": [0], "integration": [1], "of": [2, 67], "long-context": [3, 54, 161], "capabilities": [4], "with": [5], "visual": [6], "understanding": [7], "unlocks": [8], "unprecedented": [9], "potential": [10], "for": [11, 53, 112, 131], "Vision": [12], "Language": [13], "Models": [14], "(VLMs).": [15], "However,": [16], "the": [17, 22, 50, 62, 76, 96, 108, 117, 123, 171], "quadratic": [18], "attention": [19, 46], "complexity": [20], "during": [21], "pre-filling": [23, 172], "phase": [24], "remains": [25], "a": [26, 43, 72, 91], "significant": [27], "obstacle": [28], "to": [29, 71, 94, 176], "real-world": [30], "deployment.": [31], "To": [32], "overcome": [33], "this": [34], "limitation,": [35], "we": [36], "introduce": [37, 90], "MMInference": [38, 115, 136, 169], "(Multimodality": [39], "Million": [40], "tokens": [41, 180], "Inference),": [42], "dynamic": [44], "sparse": [45, 74, 84, 110, 118, 133], "method": [47, 93], "that": [48, 61, 168], "accelerates": [49, 170], "prefilling": [51], "stage": [52, 173], "multi-modal": [55, 151], "inputs.": [56], "First,": [57], "our": [58], "analysis": [59], "reveals": [60], "temporal": [63], "and": [64, 100, 157], "spatial": [65], "locality": [66], "video": [68], "input": [69], "leads": [70], "unique": [73, 97], "pattern,": [75], "Grid": [77, 98], "pattern.": [78], "Simultaneously,": [79], "VLMs": [80, 162], "exhibit": [81], "markedly": [82], "different": [83, 87], "distributions": [85], "across": [86], "modalities.": [88], "We": [89, 125], "permutation-based": [92], "leverage": [95], "pattern": [99], "handle": [101], "modality": [102], "boundary": [103], "issues.": [104], "By": [105], "offline": [106], "search": [107], "optimal": [109], "patterns": [111], "each": [113], "head,": [114], "constructs": [116], "distribution": [119], "dynamically": [120], "based": [121], "on": [122, 150], "input.": [124], "also": [126], "provide": [127], "optimized": [128], "GPU": [129], "kernels": [130], "efficient": [132], "computations.": [134], "Notably,": [135], "integrates": [137], "seamlessly": [138], "into": [139], "existing": [140], "VLM": [141], "pipelines": [142], "without": [143], "any": [144], "model": [145], "modifications": [146], "or": [147], "fine-tuning.": [148], "Experiments": [149], "benchmarks-including": [152], "Video": [153], "QA,": [154], "Captioning,": [155], "VisionNIAH,": [156], "Mixed-Modality": [158], "NIAH-with": [159], "state-of-the-art": [160], "(LongVila,": [163], "LlavaVideo,": [164], "VideoChat-Flash,": [165], "Qwen2.5-VL)": [166], "show": [167], "by": [174], "up": [175], "8.3x": [177], "at": [178, 188], "1M": [179], "while": [181], "maintaining": [182], "accuracy.": [183], "Our": [184], "code": [185], "is": [186], "available": [187], "https://aka.ms/MMInference.": [189]}, "primary_topic": {"id": "https://openalex.org/T11714", "display_name": "Multimodal Machine Learning Applications", "score": 0.9984999895095825, "subfield": {"id": "https://openalex.org/subfields/1707", "display_name": "Computer Vision and Pattern Recognition"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, "topics": [{"id": "https://openalex.org/T11714", "display_name": "Multimodal Machine Learning Applications", "score": 0.9984999895095825, "subfield": {"id": "https://openalex.org/subfields/1707", "display_name": "Computer Vision and Pattern Recognition"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, {"id": "https://openalex.org/T10036", "display_name": "Advanced Neural Network Applications", "score": 0.0008999999845400453, "subfield": {"id": "https://openalex.org/subfields/1707", "display_name": "Computer Vision and Pattern Recognition"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, {"id": "https://openalex.org/T11605", "display_name": "Visual Attention and Saliency Detection", "score": 9.999999747378752e-05, "subfield": {"id": "https://openalex.org/subfields/1707", "display_name": "Computer Vision and Pattern Recognition"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}], "primary_location": {"id": "pmh:oai:arXiv.org:2504.16083", "is_oa": true, "landing_page_url": "http://arxiv.org/abs/2504.16083", "pdf_url": "https://arxiv.org/pdf/2504.16083", "source": {"id": "https://openalex.org/S4306400194", "display_name": "arXiv (Cornell University)", "issn_l": "2331-8422", "issn": ["2331-8422"], "is_oa": true, "is_in_doaj": false, "is_core": false, "listed_in": [], "host_organization": "https://openalex.org/I205783295", "host_organization_name": "Cornell University", "host_organization_lineage": ["https://openalex.org/I205783295"], "host_organization_lineage_names": [], "type": "repository"}, "license": null, "license_id": null, "version": "submittedVersion", "is_accepted": false, "is_published": false, "raw_source_name": null, "raw_type": "text"}, "cited_by_count": 0, "is_retracted": false}, {"id": "https://openalex.org/W7115713398", "doi": "https://doi.org/10.5281/zenodo.17952002", "title": "The Deterministic Computation Law. A Formal Mathematical Framework for Reproducible Artificial Intelligence", "publication_year": 2025, "publication_date": "2025-12-16", "type": "preprint", "language": "en", "authorships": [{"author_position": "first", "author": {"id": null, "display_name": "Kumar, Sanjay", "orcid": null, "observed_orcids": []}, "institutions": [], "countries": [], "is_corresponding": true, "raw_author_name": "Kumar, Sanjay", "raw_affiliation_strings": ["DeterministicAI Research Labs"], "raw_orcid": null, "affiliations": [{"raw_affiliation_string": "DeterministicAI Research Labs", "institution_ids": []}]}], "abstract_inverted_index": {"Modern": [0], "AI": [1], "systems": [2], "exhibit": [3], "substantial": [4], "nondeterminism": [5], "arising": [6], "from": [7, 56], "stochastic": [8], "sampling,": [9], "floating-point": [10], "instabilities,": [11], "nondeterministic": [12], "GPU": [13], "kernels,": [14], "race": [15], "conditions": [16], "in": [17, 37, 114], "parallel": [18], "execution,": [19], "and": [20, 29, 42, 65, 95, 117, 142], "probabilistic": [21], "internal": [22, 103], "mechanisms.": [23], "This": [24, 45, 147], "variability": [25], "prevents": [26], "reproducibility,": [27], "auditability,": [28], "scientific": [30], "verification": [31], "-": [32, 60, 68, 140], "properties": [33], "required": [34], "for": [35, 52, 144, 154], "deployment": [36], "scientific,": [38], "medical,": [39], "financial,": [40], "legal,": [41], "safety-critical": [43], "contexts.": [44], "paper": [46], "introduces": [47], "a": [48, 82, 89, 97, 150], "first-principles": [49], "mathematical": [50], "foundation": [51, 153], "reproducible": [53, 102, 145], "computation.": [54, 146], "Beginning": [55], "three": [57, 138], "minimal": [58], "axioms": [59, 139], "Input": [61], "Determinism,": [62], "Representation": [63], "Invariance,": [64], "Replayable": [66], "Reasoning": [67], "we": [69], "derive": [70], "the": [71, 121, 131, 137], "Deterministic": [72], "Computation": [73], "Law": [74], "(DCL):": [75], "R": [76], "=": [77], "H(D(P))": [78], "where": [79], "D": [80], "is": [81, 96, 130], "canonicalization": [83], "operator": [84, 100], "mapping": [85], "problem": [86], "representations": [87], "into": [88], "quotient": [90, 111], "space": [91], "of": [92, 120, 159], "canonical": [93], "forms,": [94], "deterministic": [98, 155], "reasoning": [99], "implementing": [101], "state": [104, 115], "transitions.": [105], "We": [106, 123], "formally": [107], "develop": [108], "equivalence": [109], "relations,": [110], "constructions,": [112], "determinism": [113], "evolution,": [116], "categorical": [118], "interpretations": [119], "theory.": [122], "provide": [124], "full": [125], "proofs": [126], "showing": [127], "that": [128], "DCL": [129], "unique": [132], "computational": [133], "structure": [134], "consistent": [135], "with": [136], "necessary": [141], "sufficient": [143], "framework": [148], "offers": [149], "mathematically": [151], "rigorous": [152], "artificial": [156], "intelligence,": [157], "independent": [158], "architecture,": [160], "training": [161], "method,": [162], "model": [163], "class,": [164], "or": [165], "implementation": [166], "strategy.": [167]}, "primary_topic": {"id": "https://openalex.org/T13904", "display_name": "Artificial Intelligence Applications", "score": 0.14030000567436218, "subfield": {"id": "https://openalex.org/subfields/1702", "display_name": "Artificial Intelligence"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, "topics": [{"id": "https://openalex.org/T13904", "display_name": "Artificial Intelligence Applications", "score": 0.14030000567436218, "subfield": {"id": "https://openalex.org/subfields/1702", "display_name": "Artificial Intelligence"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, {"id": "https://openalex.org/T10126", "display_name": "Logic, programming, and type systems", "score": 0.13950000703334808, "subfield": {"id": "https://openalex.org/subfields/1702", "display_name": "Artificial Intelligence"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, {"id": "https://openalex.org/T11986", "display_name": "Scientific Computing and Data Management", "score": 0.11320000141859055, "subfield": {"id": "https://openalex.org/subfields/1802", "display_name": "Information Systems and Management"}, "field": {"id": "https://openalex.org/fields/18", "display_name": "Decision Sciences"}, "domain": {"id": "https://openalex.org/domains/2", "display_name": "Social Sciences"}}], "primary_location": {"id": "doi:10.5281/zenodo.17952002", "is_oa": true, "landing_page_url": "https://doi.org/10.5281/zenodo.17952002", "pdf_url": null, "source": {"id": "https://openalex.org/S4306400562", "display_name": "Zenodo (CERN European Organization for Nuclear Research)", "issn_l": null, "issn": null, "is_oa": true, "is_in_doaj": false, "is_core": false, "listed_in": [], "host_organization": "https://openalex.org/I67311998", "host_organization_name": "European Organization for Nuclear Research", "host_organization_lineage": ["https://openalex.org/I67311998"], "host_organization_lineage_names": [], "type": "repository"}, "license": "cc-by", "license_id": "https://openalex.org/licenses/cc-by", "version": null, "is_accepted": false, "is_published": false, "raw_source_name": null, "raw_type": "Preprint"}, "cited_by_count": 0, "is_retracted": false}, {"id": "https://openalex.org/W4414156680", "doi": "https://doi.org/10.36227/techrxiv.175744961.11927967/v1", "title": "Edge AI Enabled Real-Time Video Analytics: A Full-Lifecycle Perspective", "publication_year": 2025, "publication_date": "2025-09-09", "type": "preprint", "language": "en", "authorships": [{"author_position": "first", "author": {"id": "https://openalex.org/A5100410982", "display_name": "Xinyan Wang", "orcid": "https://orcid.org/0000-0002-7892-4276", "observed_orcids": ["https://orcid.org/0000-0002-7892-4276"]}, "institutions": [{"id": "https://openalex.org/I150229711", "display_name": "University of Electronic Science and Technology of China", "ror": "https://ror.org/04qr3zq92", "country_code": "CN", "type": "education", "lineage": ["https://openalex.org/I150229711"]}], "countries": ["CN"], "is_corresponding": true, "raw_author_name": "Xinyan Wang", "raw_affiliation_strings": ["College of Computer Science and Engineering , University of Electronic Science and Technology of China , Chengdu , Sichuan 610051 , China"], "raw_orcid": "https://orcid.org/0000-0002-7892-4276", "affiliations": [{"raw_affiliation_string": "College of Computer Science and Engineering , University of Electronic Science and Technology of China , Chengdu , Sichuan 610051 , China", "institution_ids": ["https://openalex.org/I150229711"]}]}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5103223723", "display_name": "Zhiwei Zhao", "orcid": "https://orcid.org/0000-0002-4555-6534", "observed_orcids": ["https://orcid.org/0000-0002-4555-6534"]}, "institutions": [{"id": "https://openalex.org/I150229711", "display_name": "University of Electronic Science and Technology of China", "ror": "https://ror.org/04qr3zq92", "country_code": "CN", "type": "education", "lineage": ["https://openalex.org/I150229711"]}], "countries": ["CN"], "is_corresponding": false, "raw_author_name": "Zhiwei Zhao", "raw_affiliation_strings": ["College of Computer Science and Engineering , University of Electronic Science and Technology of China , Chengdu , Sichuan 610051 , China"], "raw_orcid": null, "affiliations": [{"raw_affiliation_string": "College of Computer Science and Engineering , University of Electronic Science and Technology of China , Chengdu , Sichuan 610051 , China", "institution_ids": ["https://openalex.org/I150229711"]}]}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5100770003", "display_name": "Geyong Min", "orcid": "https://orcid.org/0000-0003-1395-7314", "observed_orcids": ["https://orcid.org/0000-0003-1395-7314"]}, "institutions": [{"id": "https://openalex.org/I150229711", "display_name": "University of Electronic Science and Technology of China", "ror": "https://ror.org/04qr3zq92", "country_code": "CN", "type": "education", "lineage": ["https://openalex.org/I150229711"]}], "countries": ["CN"], "is_corresponding": false, "raw_author_name": "Geyong Min", "raw_affiliation_strings": ["College of Computer Science and Engineering , University of Electronic Science and Technology of China , Chengdu , Sichuan 610051 , China"], "raw_orcid": null, "affiliations": [{"raw_affiliation_string": "College of Computer Science and Engineering , University of Electronic Science and Technology of China , Chengdu , Sichuan 610051 , China", "institution_ids": ["https://openalex.org/I150229711"]}]}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5078821760", "display_name": "Zi Wang", "orcid": "https://orcid.org/0000-0001-8635-8334", "observed_orcids": ["https://orcid.org/0000-0001-8635-8334"]}, "institutions": [{"id": "https://openalex.org/I150229711", "display_name": "University of Electronic Science and Technology of China", "ror": "https://ror.org/04qr3zq92", "country_code": "CN", "type": "education", "lineage": ["https://openalex.org/I150229711"]}, {"id": "https://openalex.org/I23923803", "display_name": "University of Exeter", "ror": "https://ror.org/03yghzc09", "country_code": "GB", "type": "education", "lineage": ["https://openalex.org/I23923803"]}], "countries": ["CN", "GB"], "is_corresponding": false, "raw_author_name": "Zi Wang", "raw_affiliation_strings": ["College of Computer Science and Engineering , University of Electronic Science and Technology of China , Chengdu , Sichuan 610051 , China", "Department of Computer Science , University of Exeter , EX4 4PY Exeter , U.K"], "raw_orcid": null, "affiliations": [{"raw_affiliation_string": "College of Computer Science and Engineering , University of Electronic Science and Technology of China , Chengdu , Sichuan 610051 , China", "institution_ids": ["https://openalex.org/I150229711"]}, {"raw_affiliation_string": "Department of Computer Science , University of Exeter , EX4 4PY Exeter , U.K", "institution_ids": ["https://openalex.org/I23923803"]}]}, {"author_position": "last", "author": {"id": "https://openalex.org/A5100407591", "display_name": "Zhe Wang", "orcid": "https://orcid.org/0000-0001-9727-5569", "observed_orcids": ["https://orcid.org/0000-0001-9727-5569", "https://orcid.org/0009-0008-2790-2473"]}, "institutions": [{"id": "https://openalex.org/I150229711", "display_name": "University of Electronic Science and Technology of China", "ror": "https://ror.org/04qr3zq92", "country_code": "CN", "type": "education", "lineage": ["https://openalex.org/I150229711"]}, {"id": "https://openalex.org/I23923803", "display_name": "University of Exeter", "ror": "https://ror.org/03yghzc09", "country_code": "GB", "type": "education", "lineage": ["https://openalex.org/I23923803"]}], "countries": ["CN", "GB"], "is_corresponding": false, "raw_author_name": "Zhe Wang", "raw_affiliation_strings": ["College of Computer Science and Engineering , University of Electronic Science and Technology of China , Chengdu , Sichuan 610051 , China", "Department of Computer Science , University of Exeter , EX4 4PY Exeter , U.K"], "raw_orcid": null, "affiliations": [{"raw_affiliation_string": "College of Computer Science and Engineering , University of Electronic Science and Technology of China , Chengdu , Sichuan 610051 , China", "institution_ids": ["https://openalex.org/I150229711"]}, {"raw_affiliation_string": "Department of Computer Science , University of Exeter , EX4 4PY Exeter , U.K", "institution_ids": ["https://openalex.org/I23923803"]}]}], "abstract_inverted_index": {"The": [0], "unprecedented": [1], "surge": [2], "of": [3, 27, 59, 105, 120, 141, 190, 213], "video": [4, 38], "data": [5], "from": [6, 55, 67, 82, 124, 147], "diverse": [7], "domains": [8], "such": [9], "as": [10, 95], "intelligent": [11], "transportation,": [12], "security": [13, 104], "surveillance,": [14], "and": [15, 73, 80, 103, 160, 180, 218, 231], "embodied": [16], "perception": [17], "(EP)": [18], "has": [19, 41, 53, 93], "created": [20], "an": [21], "urgent": [22], "need": [23], "for": [24, 47, 169, 215], "efficient": [25], "extraction": [26], "useful": [28], "information": [29], "to": [30, 85, 127, 182, 235], "enable": [31], "smart": [32], "decision-making.": [33], "As": [34], "a": [35, 43, 96, 148], "result,": [36], "real-time": [37, 51, 145, 194], "analytics": [39], "(VA)": [40], "become": [42], "key": [44, 228], "enabling": [45], "technique": [46], "these": [48], "applications.": [49], "Recently,": [50], "VA": [52, 65, 92, 106, 115, 146, 170, 195], "benefited": [54], "the": [56, 83, 86, 101, 118, 131, 137, 187, 210, 220, 227], "rapid": [57], "advancements": [58, 189], "artificial": [60], "intelligence": [61, 81], "(AI).": [62], "However,": [63], "cloud-based": [64], "suffers": [66], "considerable": [68], "latency,": [69], "substantial": [70], "bandwidth": [71], "consumption,": [72], "potential": [74], "privacy": [75], "leakage.": [76], "By": [77], "moving": [78], "service": [79], "cloud": [84], "network": [87], "edge,": [88], "edge": [89, 112, 142, 191], "AI": [90, 113, 121, 143, 167, 192], "enabled": [91, 114, 144, 193], "emerged": [94], "transformative": [97], "paradigm": [98], "that": [99, 165], "enhances": [100], "efficiency": [102], "systems.": [107], "Unfortunately,": [108], "existing": [109], "surveys": [110], "on": [111], "rarely": [116], "consider": [117], "lifecycle": [119], "models,": [122], "ranging": [123], "training,": [125], "deployment": [126, 155, 211], "inference.": [128], "Motivated": [129], "by": [130, 197], "above": [132], "gap,": [133], "this": [134, 240], "paper": [135], "presents": [136], "first": [138], "comprehensive": [139], "survey": [140], "fulllifecycle": [149], "perspective,": [150], "covering": [151], "edge-based": [152], "training": [153], "paradigms,": [154], "strategies,": [156], "post-deployment": [157], "adaptation": [158], "mechanisms,": [159], "inference": [161, 222], "optimizations.": [162], "Furthermore,": [163], "considering": [164], "traditional": [166], "models": [168, 200], "still": [171], "face": [172], "significant": [173], "challenges": [174, 212, 230], "in": [175, 239], "multimodal": [176], "reasoning,": [177], "contextual": [178], "understanding,": [179], "generalization": [181], "unseen": [183], "scenarios,": [184], "we": [185, 203, 225], "review": [186], "recent": [188], "powered": [196], "large": [198], "language": [199], "(LLMs).": [201], "Specifically,": [202], "present": [204], "several": [205], "representative": [206], "vision": [207], "tasks,": [208], "outline": [209], "LLMs": [214], "edgebased": [216], "VA,": [217], "discuss": [219, 232], "corresponding": [221], "solutions.": [223], "Finally,": [224], "highlight": [226], "research": [229], "future": [233], "opportunities": [234], "guide": [236], "continuous": [237], "efforts": [238], "field.": [241]}, "primary_topic": {"id": "https://openalex.org/T10273", "display_name": "IoT and Edge/Fog Computing", "score": 0.9678000211715698, "subfield": {"id": "https://openalex.org/subfields/1705", "display_name": "Computer Networks and Communications"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, "topics": [{"id": "https://openalex.org/T10273", "display_name": "IoT and Edge/Fog Computing", "score": 0.9678000211715698, "subfield": {"id": "https://openalex.org/subfields/1705", "display_name": "Computer Networks and Communications"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, {"id": "https://openalex.org/T10036", "display_name": "Advanced Neural Network Applications", "score": 0.02850000001490116, "subfield": {"id": "https://openalex.org/subfields/1707", "display_name": "Computer Vision and Pattern Recognition"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, {"id": "https://openalex.org/T10331", "display_name": "Video Surveillance and Tracking Methods", "score": 0.0005000000237487257, "subfield": {"id": "https://openalex.org/subfields/1707", "display_name": "Computer Vision and Pattern Recognition"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}], "primary_location": {"id": "doi:10.36227/techrxiv.175744961.11927967/v1", "is_oa": true, "landing_page_url": "https://doi.org/10.36227/techrxiv.175744961.11927967/v1", "pdf_url": "https://www.techrxiv.org/doi/pdf/10.36227/techrxiv.175744961.11927967", "source": null, "license": "cc-by", "license_id": "https://openalex.org/licenses/cc-by", "version": "acceptedVersion", "is_accepted": true, "is_published": false, "raw_source_name": null, "raw_type": "posted-content"}, "cited_by_count": 1, "is_retracted": false}, {"id": "https://openalex.org/W4416392741", "doi": "https://doi.org/10.48550/arxiv.2510.07356", "title": "ConCuR: Conciseness Makes State-of-the-Art Kernel Generation", "publication_year": 2025, "publication_date": "2025-10-08", "type": "preprint", "language": "en", "authorships": [{"author_position": "first", "author": {"id": null, "display_name": "Kong, Lingcheng", "orcid": null, "observed_orcids": []}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Kong, Lingcheng", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": null, "display_name": "Wei, Jiateng", "orcid": null, "observed_orcids": []}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Wei, Jiateng", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": null, "display_name": "Shen, Hanzhang", "orcid": null, "observed_orcids": []}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Shen, Hanzhang", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "last", "author": {"id": "https://openalex.org/A5100332013", "display_name": "Huan Wang", "orcid": "https://orcid.org/0000-0002-6480-9568", "observed_orcids": ["https://orcid.org/0000-0002-6480-9568", "https://orcid.org/0009-0000-5290-3533"]}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Wang, Huan", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}], "abstract_inverted_index": {"GPU": [0], "kernel": [1, 22, 52, 120, 147, 177, 197], "generation": [2, 23, 53, 87, 178, 198], "by": [3, 74], "LLMs": [4, 49], "has": [5], "recently": [6], "experienced": [7], "rapid": [8], "development,": [9], "leveraging": [10, 44], "test-time": [11], "scaling": [12], "and": [13, 36, 65, 99, 118, 140, 157, 183, 187], "reinforcement": [14], "learning": [15], "techniques.": [16], "However,": [17], "a": [18, 61, 75, 111, 170], "key": [19], "challenge": [20, 40], "for": [21, 146], "is": [24, 105], "the": [25, 51, 106, 126, 135, 163, 174, 196, 201], "scarcity": [26], "of": [27, 88, 115, 176], "high-quality": [28, 32, 67], "data,": [29], "as": [30, 149, 151, 155, 169], "most": [31], "kernels": [33, 69], "are": [34], "proprietary": [35], "not": [37], "open-source.": [38], "This": [39], "prevents": [41], "us": [42], "from": [43], "supervised": [45], "fine-tuning": [46], "to": [47, 50, 122, 172], "align": [48], "task.": [54], "To": [55], "address": [56], "this": [57, 92], "challenge,": [58], "we": [59, 94, 160], "develop": [60], "pipeline": [62, 189], "that": [63, 78, 162], "generates": [64], "curates": [66], "CUDA": [68, 119], "with": [70], "reasoning": [71, 82, 165], "traces,": [72], "motivated": [73], "critical": [76], "observation": [77], "concise": [79], "yet": [80], "informative": [81], "traces": [83], "result": [84], "in": [85, 195, 200], "robust": [86], "high-performance": [89], "kernels.": [90], "Using": [91], "pipeline,": [93], "construct": [95], "our": [96, 101, 123, 129, 184], "dataset": [97, 113], "ConCuR": [98], "introduce": [100], "model": [102, 108, 130], "KernelCoder,": [103], "which": [104], "first": [107], "trained": [109], "on": [110], "curated": [112], "consisting": [114], "PyTorch,": [116], "reasoning,": [117], "pairs,": [121], "knowledge.": [124], "In": [125], "KernelBench": [127], "setup,": [128], "achieves": [131], "significant": [132], "improvements": [133], "over": [134], "existing": [136], "top-performing": [137], "model,": [138], "QwQ-32B,": [139], "outperforms": [141], "all": [142], "open-source": [143], "models": [144, 153], "fine-tuned": [145], "generation,": [148], "well": [150], "frontier": [152], "such": [154], "DeepSeek-V3.1-Think": [156], "Claude-4-sonnet.": [158], "Finally,": [159], "show": [161], "average": [164], "length": [166], "can": [167, 190], "serve": [168], "metric": [171], "assess": [173], "difficulty": [175], "tasks.": [179], "The": [180], "observations,": [181], "metrics,": [182], "data": [185, 194], "collection": [186], "curation": [188], "help": [191], "obtain": [192], "better": [193], "task": [199], "future.": [202]}, "primary_topic": {"id": "https://openalex.org/T10054", "display_name": "Parallel Computing and Optimization Techniques", "score": 0.6435999870300293, "subfield": {"id": "https://openalex.org/subfields/1708", "display_name": "Hardware and Architecture"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, "topics": [{"id": "https://openalex.org/T10054", "display_name": "Parallel Computing and Optimization Techniques", "score": 0.6435999870300293, "subfield": {"id": "https://openalex.org/subfields/1708", "display_name": "Hardware and Architecture"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, {"id": "https://openalex.org/T10260", "display_name": "Software Engineering Research", "score": 0.28029999136924744, "subfield": {"id": "https://openalex.org/subfields/1710", "display_name": "Information Systems"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, {"id": "https://openalex.org/T10028", "display_name": "Topic Modeling", "score": 0.014800000004470348, "subfield": {"id": "https://openalex.org/subfields/1702", "display_name": "Artificial Intelligence"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}], "primary_location": {"id": "pmh:oai:arXiv.org:2510.07356", "is_oa": true, "landing_page_url": "http://arxiv.org/abs/2510.07356", "pdf_url": "https://arxiv.org/pdf/2510.07356", "source": {"id": "https://openalex.org/S4306400194", "display_name": "arXiv (Cornell University)", "issn_l": "2331-8422", "issn": ["2331-8422"], "is_oa": true, "is_in_doaj": false, "is_core": false, "listed_in": [], "host_organization": "https://openalex.org/I205783295", "host_organization_name": "Cornell University", "host_organization_lineage": ["https://openalex.org/I205783295"], "host_organization_lineage_names": [], "type": "repository"}, "license": null, "license_id": null, "version": "submittedVersion", "is_accepted": false, "is_published": false, "raw_source_name": null, "raw_type": "text"}, "cited_by_count": 0, "is_retracted": false}, {"id": "https://openalex.org/W4414830449", "doi": "https://doi.org/10.48550/arxiv.2504.08242", "title": "Jupiter: Fast and Resource-Efficient Collaborative Inference of Generative LLMs on Edge Devices", "publication_year": 2025, "publication_date": "2025-04-11", "type": "preprint", "language": "en", "authorships": [{"author_position": "first", "author": {"id": "https://openalex.org/A5069427601", "display_name": "Shengyuan Ye", "orcid": "https://orcid.org/0000-0001-8867-0655", "observed_orcids": ["https://orcid.org/0000-0001-8867-0655"]}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Ye, Shengyuan", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5106355629", "display_name": "Bei Ouyang", "orcid": "https://orcid.org/0009-0001-4408-2832", "observed_orcids": ["https://orcid.org/0009-0001-4408-2832"]}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Ouyang, Bei", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5055161955", "display_name": "Liekang Zeng", "orcid": "https://orcid.org/0000-0003-4800-8768", "observed_orcids": ["https://orcid.org/0000-0003-4800-8768"]}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Zeng, Liekang", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5106355630", "display_name": "Tianyi Qian", "orcid": "https://orcid.org/0009-0001-2802-2258", "observed_orcids": ["https://orcid.org/0009-0001-2802-2258"]}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Qian, Tianyi", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5100730785", "display_name": "Xiaowen Chu", "orcid": "https://orcid.org/0000-0001-9745-4372", "observed_orcids": ["https://orcid.org/0000-0001-9745-4372"]}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Chu, Xiaowen", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5039176528", "display_name": "Jian Tang", "orcid": "https://orcid.org/0000-0003-4418-0114", "observed_orcids": ["https://orcid.org/0000-0003-4418-0114"]}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Tang, Jian", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "last", "author": {"id": "https://openalex.org/A5100385709", "display_name": "Chen Xu", "orcid": "https://orcid.org/0000-0002-3070-9358", "observed_orcids": ["https://orcid.org/0000-0002-3070-9358"]}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Chen, Xu", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}], "abstract_inverted_index": {"Generative": [0], "large": [1], "language": [2], "models": [3], "(LLMs)": [4], "have": [5], "garnered": [6], "significant": [7], "attention": [8], "due": [9], "to": [10, 33, 70, 143, 170, 214], "their": [11], "exceptional": [12], "capabilities": [13], "in": [14, 20, 53], "various": [15, 208], "AI": [16, 122], "tasks.": [17], "Traditionally": [18], "deployed": [19], "cloud": [21], "datacenters,": [22], "LLMs": [23], "are": [24], "now": [25], "increasingly": [26], "moving": [27], "towards": [28], "more": [29], "accessible": [30], "edge": [31, 48, 68, 89, 121, 209], "platforms": [32], "protect": [34], "sensitive": [35], "user": [36], "data": [37], "and": [38, 58, 86, 118, 137, 150, 163], "ensure": [39], "privacy": [40], "preservation.": [41], "The": [42], "limited": [43], "computational": [44], "resources": [45], "of": [46, 75, 88, 147], "individual": [47, 76], "devices,": [49, 77], "however,": [50], "can": [51], "result": [52], "excessively": [54], "prolonged": [55], "inference": [56, 192], "latency": [57, 217], "overwhelmed": [59], "memory": [60], "usage.": [61], "While": [62], "existing": [63], "research": [64], "has": [65], "explored": [66], "collaborative": [67, 120], "computing": [69], "break": [71], "the": [72, 97, 101, 144, 148], "resource": [73, 172], "wall": [74], "these": [78], "solutions": [79], "yet": [80], "suffer": [81], "from": [82], "massive": [83], "communication": [84], "overhead": [85], "under-utilization": [87], "resources.": [90], "Furthermore,": [91], "they": [92], "focus": [93], "exclusively": [94], "on": [95, 197], "optimizing": [96], "prefill": [98, 149, 154], "phase,": [99, 155], "neglecting": [100], "crucial": [102], "autoregressive": [103], "decoding": [104, 151, 183], "phase": [105], "for": [106, 124], "generative": [107, 125], "LLMs.": [108], "To": [109], "address": [110], "that,": [111], "we": [112], "propose": [113], "Jupiter,": [114], "a": [115, 130, 135, 158, 165], "fast,": [116], "scalable,": [117], "resource-efficient": [119], "system": [123, 140], "LLM": [126], "inference.": [127], "Jupiter": [128, 156, 176, 202], "introduces": [129], "flexible": [131], "pipelined": [132], "architecture": [133], "as": [134], "principle": [136], "differentiates": [138], "its": [139], "design": [141], "according": [142], "differentiated": [145], "characteristics": [146], "phases.": [152], "For": [153, 174], "submits": [157], "novel": [159], "intra-sequence": [160], "pipeline": [161, 181], "parallelism": [162, 167], "develops": [164], "meticulous": [166], "planning": [168], "strategy": [169], "maximize": [171], "efficiency;": [173], "decoding,": [175, 188], "devises": [177], "an": [178], "effective": [179], "outline-based": [180], "parallel": [182], "mechanism": [184], "combined": [185], "with": [186], "speculative": [187], "which": [189], "further": [190], "magnifies": [191], "acceleration.": [193], "Extensive": [194], "evaluation": [195], "based": [196], "realistic": [198], "implementation": [199], "demonstrates": [200], "that": [201], "remarkably": [203], "outperforms": [204], "state-of-the-art": [205], "approaches": [206], "under": [207], "environment": [210], "setups,": [211], "achieving": [212], "up": [213], "26.1x": [215], "end-to-end": [216], "reduction": [218], "while": [219], "rendering": [220], "on-par": [221], "generation": [222], "quality.": [223]}, "primary_topic": {"id": "https://openalex.org/T10273", "display_name": "IoT and Edge/Fog Computing", "score": 0.9624000191688538, "subfield": {"id": "https://openalex.org/subfields/1705", "display_name": "Computer Networks and Communications"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, "topics": [{"id": "https://openalex.org/T10273", "display_name": "IoT and Edge/Fog Computing", "score": 0.9624000191688538, "subfield": {"id": "https://openalex.org/subfields/1705", "display_name": "Computer Networks and Communications"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, {"id": "https://openalex.org/T10054", "display_name": "Parallel Computing and Optimization Techniques", "score": 0.0071000000461936, "subfield": {"id": "https://openalex.org/subfields/1708", "display_name": "Hardware and Architecture"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, {"id": "https://openalex.org/T10181", "display_name": "Natural Language Processing Techniques", "score": 0.005400000140070915, "subfield": {"id": "https://openalex.org/subfields/1702", "display_name": "Artificial Intelligence"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}], "primary_location": {"id": "pmh:oai:arXiv.org:2504.08242", "is_oa": true, "landing_page_url": "http://arxiv.org/abs/2504.08242", "pdf_url": "https://arxiv.org/pdf/2504.08242", "source": {"id": "https://openalex.org/S4306400194", "display_name": "arXiv (Cornell University)", "issn_l": "2331-8422", "issn": ["2331-8422"], "is_oa": true, "is_in_doaj": false, "is_core": false, "listed_in": [], "host_organization": "https://openalex.org/I205783295", "host_organization_name": "Cornell University", "host_organization_lineage": ["https://openalex.org/I205783295"], "host_organization_lineage_names": [], "type": "repository"}, "license": null, "license_id": null, "version": "submittedVersion", "is_accepted": false, "is_published": false, "raw_source_name": null, "raw_type": "text"}, "cited_by_count": 0, "is_retracted": false}, {"id": "https://openalex.org/W4415145225", "doi": "https://doi.org/10.48550/arxiv.2505.15380", "title": "Accelerating Autoregressive Speech Synthesis Inference With Speech Speculative Decoding", "publication_year": 2025, "publication_date": "2025-05-21", "type": "preprint", "language": "en", "authorships": [{"author_position": "first", "author": {"id": "https://openalex.org/A5093400546", "display_name": "Zijian Lin", "orcid": "https://orcid.org/0000-0001-7653-4496", "observed_orcids": ["https://orcid.org/0000-0001-7653-4496"]}, "institutions": [{"id": "https://openalex.org/I4405263052", "display_name": "Tsinghua Shenzhen International Graduate School", "ror": "https://ror.org/01vceef04", "country_code": "CN", "type": "education", "lineage": ["https://openalex.org/I4405263052", "https://openalex.org/I99065089"]}, {"id": "https://openalex.org/I99065089", "display_name": "Tsinghua University", "ror": "https://ror.org/03cve4549", "country_code": "CN", "type": "education", "lineage": ["https://openalex.org/I99065089"]}], "countries": ["CN"], "is_corresponding": false, "raw_author_name": "Lin, Zijian", "raw_affiliation_strings": ["Shenzhen International Graduate School , Tsinghua University , China"], "raw_orcid": null, "affiliations": [{"raw_affiliation_string": "Shenzhen International Graduate School , Tsinghua University , China", "institution_ids": ["https://openalex.org/I4405263052", "https://openalex.org/I99065089"]}]}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5100354733", "display_name": "Yang Zhang", "orcid": "https://orcid.org/0000-0002-8540-1254", "observed_orcids": ["https://orcid.org/0000-0002-8540-1254", "https://orcid.org/0000-0003-4593-8310"]}, "institutions": [{"id": "https://openalex.org/I2250653659", "display_name": "Tencent (China)", "ror": "https://ror.org/00hhjss72", "country_code": "CN", "type": "company", "lineage": ["https://openalex.org/I2250653659"]}], "countries": ["CN"], "is_corresponding": false, "raw_author_name": "Zhang, Yang", "raw_affiliation_strings": ["Platform and Content Group , Tencent , China"], "raw_orcid": null, "affiliations": [{"raw_affiliation_string": "Platform and Content Group , Tencent , China", "institution_ids": ["https://openalex.org/I2250653659"]}]}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5017808406", "display_name": "Yougen Yuan", "orcid": "https://orcid.org/0000-0002-2490-566X", "observed_orcids": ["https://orcid.org/0000-0002-2490-566X"]}, "institutions": [{"id": "https://openalex.org/I2250653659", "display_name": "Tencent (China)", "ror": "https://ror.org/00hhjss72", "country_code": "CN", "type": "company", "lineage": ["https://openalex.org/I2250653659"]}], "countries": ["CN"], "is_corresponding": false, "raw_author_name": "Yuan, Yougen", "raw_affiliation_strings": ["Platform and Content Group , Tencent , China"], "raw_orcid": null, "affiliations": [{"raw_affiliation_string": "Platform and Content Group , Tencent , China", "institution_ids": ["https://openalex.org/I2250653659"]}]}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5049927248", "display_name": "Yuming Yan", "orcid": "https://orcid.org/0009-0005-4596-1600", "observed_orcids": ["https://orcid.org/0009-0005-4596-1600"]}, "institutions": [{"id": "https://openalex.org/I2250653659", "display_name": "Tencent (China)", "ror": "https://ror.org/00hhjss72", "country_code": "CN", "type": "company", "lineage": ["https://openalex.org/I2250653659"]}], "countries": ["CN"], "is_corresponding": false, "raw_author_name": "Yan, Yuming", "raw_affiliation_strings": ["Platform and Content Group , Tencent , China"], "raw_orcid": null, "affiliations": [{"raw_affiliation_string": "Platform and Content Group , Tencent , China", "institution_ids": ["https://openalex.org/I2250653659"]}]}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5063136754", "display_name": "Jinjiang Liu", "orcid": "https://orcid.org/0000-0002-1543-9193", "observed_orcids": ["https://orcid.org/0000-0002-1543-9193"]}, "institutions": [{"id": "https://openalex.org/I4405263052", "display_name": "Tsinghua Shenzhen International Graduate School", "ror": "https://ror.org/01vceef04", "country_code": "CN", "type": "education", "lineage": ["https://openalex.org/I4405263052", "https://openalex.org/I99065089"]}, {"id": "https://openalex.org/I99065089", "display_name": "Tsinghua University", "ror": "https://ror.org/03cve4549", "country_code": "CN", "type": "education", "lineage": ["https://openalex.org/I99065089"]}], "countries": ["CN"], "is_corresponding": false, "raw_author_name": "Liu, Jinjiang", "raw_affiliation_strings": ["Shenzhen International Graduate School , Tsinghua University , China"], "raw_orcid": null, "affiliations": [{"raw_affiliation_string": "Shenzhen International Graduate School , Tsinghua University , China", "institution_ids": ["https://openalex.org/I4405263052", "https://openalex.org/I99065089"]}]}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5102869280", "display_name": "Zhiyong Wu", "orcid": "https://orcid.org/0000-0001-8533-0524", "observed_orcids": ["https://orcid.org/0000-0001-8533-0524"]}, "institutions": [{"id": "https://openalex.org/I4405263052", "display_name": "Tsinghua Shenzhen International Graduate School", "ror": "https://ror.org/01vceef04", "country_code": "CN", "type": "education", "lineage": ["https://openalex.org/I4405263052", "https://openalex.org/I99065089"]}, {"id": "https://openalex.org/I99065089", "display_name": "Tsinghua University", "ror": "https://ror.org/03cve4549", "country_code": "CN", "type": "education", "lineage": ["https://openalex.org/I99065089"]}], "countries": ["CN"], "is_corresponding": false, "raw_author_name": "Wu, Zhiyong", "raw_affiliation_strings": ["Shenzhen International Graduate School , Tsinghua University , China"], "raw_orcid": null, "affiliations": [{"raw_affiliation_string": "Shenzhen International Graduate School , Tsinghua University , China", "institution_ids": ["https://openalex.org/I4405263052", "https://openalex.org/I99065089"]}]}, {"author_position": "middle", "author": {"id": null, "display_name": "Hu, Pengfei", "orcid": null, "observed_orcids": []}, "institutions": [{"id": "https://openalex.org/I2250653659", "display_name": "Tencent (China)", "ror": "https://ror.org/00hhjss72", "country_code": "CN", "type": "company", "lineage": ["https://openalex.org/I2250653659"]}], "countries": ["CN"], "is_corresponding": false, "raw_author_name": "Hu, Pengfei", "raw_affiliation_strings": ["Platform and Content Group , Tencent , China"], "raw_orcid": null, "affiliations": [{"raw_affiliation_string": "Platform and Content Group , Tencent , China", "institution_ids": ["https://openalex.org/I2250653659"]}]}, {"author_position": "last", "author": {"id": null, "display_name": "Yu, Qun", "orcid": null, "observed_orcids": []}, "institutions": [{"id": "https://openalex.org/I2250653659", "display_name": "Tencent (China)", "ror": "https://ror.org/00hhjss72", "country_code": "CN", "type": "company", "lineage": ["https://openalex.org/I2250653659"]}], "countries": ["CN"], "is_corresponding": false, "raw_author_name": "Yu, Qun", "raw_affiliation_strings": ["Platform and Content Group , Tencent , China"], "raw_orcid": null, "affiliations": [{"raw_affiliation_string": "Platform and Content Group , Tencent , China", "institution_ids": ["https://openalex.org/I2250653659"]}]}], "abstract_inverted_index": {"Modern": [0], "autoregressive": [1, 50, 96], "speech": [2, 51], "synthesis": [3, 52], "models": [4, 7, 22], "leveraging": [5], "language": [6], "have": [8], "demonstrated": [9], "remarkable": [10], "performance.": [11], "However,": [12], "the": [13, 74, 78, 108, 114, 118], "sequential": [14], "nature": [15], "of": [16, 91, 110, 117], "next": [17], "token": [18, 65], "prediction": [19], "in": [20, 30, 71, 112], "these": [21], "leads": [23], "to": [24, 62], "significant": [25, 89], "latency,": [26], "hindering": [27], "their": [28], "deployment": [29], "scenarios": [31], "where": [32], "inference": [33], "speed": [34], "is": [35], "critical.": [36], "In": [37], "this": [38], "work,": [39], "we": [40], "propose": [41], "Speech": [42], "Speculative": [43], "Decoding": [44], "(SSD),": [45], "a": [46, 58, 88], "novel": [47], "framework": [48], "for": [49], "acceleration.": [53], "Specifically,": [54], "our": [55], "method": [56], "employs": [57], "lightweight": [59], "draft": [60], "model": [61, 76, 120], "generate": [63], "candidate": [64], "sequences,": [66], "which": [67], "are": [68], "subsequently": [69], "verified": [70], "parallel": [72], "by": [73], "target": [75, 119], "using": [77], "proposed": [79], "SSD": [80, 86, 111], "framework.": [81], "Experimental": [82], "results": [83], "demonstrate": [84], "that": [85], "achieves": [87], "speedup": [90], "1.4x": [92], "compared": [93], "with": [94], "conventional": [95], "decoding,": [97], "while": [98, 121], "maintaining": [99], "high": [100], "fidelity": [101], "and": [102], "naturalness.": [103], "Subjective": [104], "evaluations": [105], "further": [106], "validate": [107], "effectiveness": [109], "preserving": [113], "perceptual": [115], "quality": [116], "accelerating": [122], "inference.": [123]}, "primary_topic": {"id": "https://openalex.org/T10201", "display_name": "Speech Recognition and Synthesis", "score": 0.9994999766349792, "subfield": {"id": "https://openalex.org/subfields/1702", "display_name": "Artificial Intelligence"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, "topics": [{"id": "https://openalex.org/T10201", "display_name": "Speech Recognition and Synthesis", "score": 0.9994999766349792, "subfield": {"id": "https://openalex.org/subfields/1702", "display_name": "Artificial Intelligence"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, {"id": "https://openalex.org/T10860", "display_name": "Speech and Audio Processing", "score": 0.00019999999494757503, "subfield": {"id": "https://openalex.org/subfields/1711", "display_name": "Signal Processing"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, {"id": "https://openalex.org/T10028", "display_name": "Topic Modeling", "score": 0.0, "subfield": {"id": "https://openalex.org/subfields/1702", "display_name": "Artificial Intelligence"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}], "primary_location": {"id": "pmh:oai:arXiv.org:2505.15380", "is_oa": true, "landing_page_url": "http://arxiv.org/abs/2505.15380", "pdf_url": "https://arxiv.org/pdf/2505.15380", "source": {"id": "https://openalex.org/S4306400194", "display_name": "arXiv (Cornell University)", "issn_l": "2331-8422", "issn": ["2331-8422"], "is_oa": true, "is_in_doaj": false, "is_core": false, "listed_in": [], "host_organization": "https://openalex.org/I205783295", "host_organization_name": "Cornell University", "host_organization_lineage": ["https://openalex.org/I205783295"], "host_organization_lineage_names": [], "type": "repository"}, "license": null, "license_id": null, "version": "submittedVersion", "is_accepted": false, "is_published": false, "raw_source_name": null, "raw_type": "text"}, "cited_by_count": 0, "is_retracted": false}, {"id": "https://openalex.org/W4415089005", "doi": "https://doi.org/10.48550/arxiv.2509.11815", "title": "SpecVLM: Fast Speculative Decoding in Vision-Language Models", "publication_year": 2025, "publication_date": "2025-09-15", "type": "preprint", "language": "en", "authorships": [{"author_position": "first", "author": {"id": "https://openalex.org/A5003944732", "display_name": "Haiduo Huang", "orcid": "https://orcid.org/0009-0000-6726-9221", "observed_orcids": ["https://orcid.org/0009-0000-6726-9221"]}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Huang, Haiduo", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5115565809", "display_name": "Fuwei Yang", "orcid": null, "observed_orcids": []}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Yang, Fuwei", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5100534154", "display_name": "Zhenhua Liu", "orcid": "https://orcid.org/0009-0001-1413-3133", "observed_orcids": ["https://orcid.org/0009-0001-1413-3133", "https://orcid.org/0009-0008-7656-0633"]}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Liu, Zhenhua", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": null, "display_name": "Yin, Xuanwu", "orcid": null, "observed_orcids": []}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Yin, Xuanwu", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5052630982", "display_name": "Dong Li", "orcid": "https://orcid.org/0009-0003-4039-434X", "observed_orcids": ["https://orcid.org/0009-0003-4039-434X"]}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Li, Dong", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": null, "display_name": "Ren, Pengju", "orcid": null, "observed_orcids": []}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Ren, Pengju", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "last", "author": {"id": "https://openalex.org/A5027115167", "display_name": "Emad Barsoum", "orcid": "https://orcid.org/0000-0002-4097-8690", "observed_orcids": ["https://orcid.org/0000-0002-4097-8690"]}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Barsoum, Emad", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}], "abstract_inverted_index": {"Speculative": [0], "decoding": [1, 55], "is": [2, 28, 202], "a": [3, 61, 67, 133, 150], "powerful": [4], "way": [5], "to": [6, 17, 101], "accelerate": [7], "autoregressive": [8, 78], "large": [9], "language": [10], "models": [11, 19], "(LLMs),": [12], "but": [13], "directly": [14], "porting": [15], "it": [16], "vision-language": [18], "(VLMs)": [20], "faces": [21], "unique": [22], "systems": [23], "constraints:": [24], "the": [25, 48, 122, 159, 193], "prefill": [26], "stage": [27], "dominated": [29], "by": [30], "visual": [31, 89], "tokens": [32], "whose": [33], "count": [34], "scales": [35], "with": [36, 86, 125], "image": [37], "resolution": [38], "and": [39, 45, 58, 80, 98, 104, 129, 136, 142, 183, 188], "video": [40], "length,": [41, 164], "inflating": [42], "both": [43], "compute": [44], "memory,": [46], "especially": [47], "key-value": [49], "(KV)": [50], "cache.": [51], "We": [52], "study": [53], "speculative": [54, 166], "for": [56], "VLMs": [57], "introduce": [59], "SpecVLM,": [60], "practical": [62], "system": [63], "that": [64, 91, 120], "(1)": [65], "establishes": [66], "strong": [68], "EAGLE-2-style": [69], "baseline,": [70], "EagleVLM,": [71], "delivering": [72], "1.5--2.3x": [73], "end-to-end": [74, 176], "speedups": [75, 177], "over": [76, 186], "full": [77], "inference,": [79], "(2)": [81], "further": [82], "accelerates": [83], "VLM": [84], "inference": [85], "an": [87, 116], "elastic": [88], "compressor": [90], "adaptively": [92], "selects": [93], "among": [94], "pruning,": [95], "pooling,": [96], "convolution,": [97], "resampler": [99], "primitives": [100], "balance": [102], "FLOPs/parameters": [103], "accuracy": [105], "per": [106], "input.": [107], "To": [108], "avoid": [109], "costly": [110], "offline": [111], "distillation": [112, 118], "corpora,": [113], "we": [114], "propose": [115], "online-logit": [117], "protocol": [119, 148], "trains": [121], "draft": [123, 160], "model": [124], "on-the-fly": [126], "teacher": [127], "logits": [128], "penultimate": [130], "features": [131], "using": [132], "combined": [134], "cross-entropy": [135], "Smooth": [137], "L1": [138], "objective,": [139], "eliminating": [140], "storage": [141], "preprocessing": [143], "while": [144, 191], "remaining": [145], "compute-efficient.": [146], "This": [147], "reveals": [149], "training-time": [151], "scaling": [152], "effect:": [153], "longer": [154], "online": [155], "training": [156], "monotonically": [157], "increases": [158], "model's": [161, 195], "average": [162], "accepted": [163], "improving": [165], "efficiency.": [167], "Empirically,": [168], "SpecVLM": [169], "achieves": [170], "additional": [171], "acceleration,": [172], "culminating": [173], "in": [174], "2.5--2.9x": [175], "within": [178], "5": [179], "epochs": [180], "across": [181], "LLaVA": [182], "MMMU,": [184], "consistently": [185], "resolutions": [187], "task": [189], "difficulties,": [190], "preserving": [192], "target": [194], "output": [196], "distribution": [197], "(lossless": [198], "decoding).": [199], "Our": [200], "code": [201], "available": [203], "at": [204], "https://github.com/haiduo/SpecVLM.": [205]}, "primary_topic": {"id": "https://openalex.org/T11714", "display_name": "Multimodal Machine Learning Applications", "score": 0.7509999871253967, "subfield": {"id": "https://openalex.org/subfields/1707", "display_name": "Computer Vision and Pattern Recognition"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, "topics": [{"id": "https://openalex.org/T11714", "display_name": "Multimodal Machine Learning Applications", "score": 0.7509999871253967, "subfield": {"id": "https://openalex.org/subfields/1707", "display_name": "Computer Vision and Pattern Recognition"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, {"id": "https://openalex.org/T10181", "display_name": "Natural Language Processing Techniques", "score": 0.11680000275373459, "subfield": {"id": "https://openalex.org/subfields/1702", "display_name": "Artificial Intelligence"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, {"id": "https://openalex.org/T10036", "display_name": "Advanced Neural Network Applications", "score": 0.10189999639987946, "subfield": {"id": "https://openalex.org/subfields/1707", "display_name": "Computer Vision and Pattern Recognition"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}], "primary_location": {"id": "pmh:oai:arXiv.org:2509.11815", "is_oa": true, "landing_page_url": "http://arxiv.org/abs/2509.11815", "pdf_url": "https://arxiv.org/pdf/2509.11815", "source": {"id": "https://openalex.org/S4306400194", "display_name": "arXiv (Cornell University)", "issn_l": "2331-8422", "issn": ["2331-8422"], "is_oa": true, "is_in_doaj": false, "is_core": false, "listed_in": [], "host_organization": "https://openalex.org/I205783295", "host_organization_name": "Cornell University", "host_organization_lineage": ["https://openalex.org/I205783295"], "host_organization_lineage_names": [], "type": "repository"}, "license": null, "license_id": null, "version": "submittedVersion", "is_accepted": false, "is_published": false, "raw_source_name": null, "raw_type": "text"}, "cited_by_count": 0, "is_retracted": false}, {"id": "https://openalex.org/W4417166484", "doi": "https://doi.org/10.48550/arxiv.2504.18583", "title": "PARD: Accelerating LLM Inference with Low-Cost PARallel Draft Model Adaptation", "publication_year": 2025, "publication_date": "2025-04-23", "type": "preprint", "language": "en", "authorships": [{"author_position": "first", "author": {"id": "https://openalex.org/A5119805120", "display_name": "Zihao An", "orcid": "https://orcid.org/0009-0001-5287-4813", "observed_orcids": ["https://orcid.org/0009-0001-5287-4813"]}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "An, Zihao", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5049409322", "display_name": "Huajun Bai", "orcid": "https://orcid.org/0000-0001-6017-9776", "observed_orcids": ["https://orcid.org/0000-0001-6017-9776"]}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Bai, Huajun", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5079691252", "display_name": "Ziqiong Liu", "orcid": null, "observed_orcids": []}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Liu, Ziqiong", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5100659913", "display_name": "Dong Li", "orcid": "https://orcid.org/0000-0002-4912-8107", "observed_orcids": ["https://orcid.org/0000-0002-4912-8107"]}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Li, Dong", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "last", "author": {"id": "https://openalex.org/A5027115167", "display_name": "Emad Barsoum", "orcid": "https://orcid.org/0000-0002-4097-8690", "observed_orcids": ["https://orcid.org/0000-0002-4097-8690"]}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Barsoum, Emad", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}], "abstract_inverted_index": {"The": [0], "autoregressive": [1, 158], "nature": [2], "of": [3, 53, 100, 129, 140, 153], "large": [4], "language": [5], "models": [6, 102, 160, 167], "(LLMs)": [7], "fundamentally": [8], "limits": [9], "inference": [10, 118, 194], "speed,": [11], "as": [12, 32], "each": [13, 60, 108], "forward": [14, 127], "pass": [15, 128], "generates": [16], "only": [17], "a": [18, 33, 37, 55, 75, 89, 125, 144], "single": [19, 90, 126], "token": [20, 42, 84], "and": [21, 82], "is": [22, 210, 217], "often": [23], "bottlenecked": [24], "by": [25, 119, 183], "memory": [26], "bandwidth.": [27], "Speculative": [28], "decoding": [29, 78], "has": [30], "emerged": [31], "promising": [34], "solution,": [35], "adopting": [36], "draft-then-verify": [38], "strategy": [39], "to": [40, 93, 161, 199], "accelerate": [41], "generation.": [43], "While": [44], "the": [45, 130, 136, 151, 174, 192], "EAGLE": [46], "series": [47], "achieves": [48, 197], "strong": [49], "acceleration,": [50], "its": [51], "requirement": [52], "training": [54, 106, 137, 181], "separate": [56, 105], "draft": [57, 91, 131, 159, 166, 179], "head": [58], "for": [59, 107], "target": [61, 101], "model": [62, 92, 180], "introduces": [63], "substantial": [64], "adaptation": [65, 112, 138], "costs.": [66, 113], "In": [67], "this": [68], "work,": [69], "we": [70, 142], "propose": [71, 143], "\\textbf{PARD": [72], "(PARallel": [73], "Draft)},": [74], "novel": [76], "speculative": [77], "method": [79, 177], "featuring": [80], "\\textit{target-independence}": [81], "\\textit{parallel": [83], "prediction}.": [85], "Specifically,": [86], "PARD": [87, 115, 196], "enables": [88], "be": [94, 162], "applied": [95], "across": [96], "an": [97], "entire": [98], "family": [99], "without": [103], "requiring": [104], "variant,": [109], "thereby": [110], "minimizing": [111], "Meanwhile,": [114], "substantially": [116], "accelerates": [117], "predicting": [120], "multiple": [121], "future": [122], "tokens": [123, 206], "within": [124], "phase.": [132], "To": [133], "further": [134], "reduce": [135], "cost": [139], "PARD,": [141], "COnditional": [145], "Drop-token": [146], "(COD)": [147], "mechanism": [148], "based": [149], "on": [150, 202], "integrity": [152], "prefix": [154], "key-value": [155], "states,": [156], "enabling": [157], "adapted": [163], "into": [164], "parallel": [165], "at": [168, 219], "low-cost.": [169], "Our": [170, 215], "experiments": [171], "show": [172], "that": [173], "proposed": [175], "COD": [176], "improves": [178], "efficiency": [182], "\\textbf{3$\\times$}": [184], "compared": [185], "with": [186], "traditional": [187], "masked": [188], "prediction": [189], "training.": [190], "On": [191], "\\texttt{vLLM}": [193], "framework,": [195], "up": [198], "\\textbf{3.67$\\times$}": [200], "speedup": [201], "LLaMA3.1-8B,": [203], "reaching": [204], "\\textbf{264.88}": [205], "per": [207], "second,": [208], "which": [209], "\\textbf{1.15$\\times$}": [211], "faster": [212], "than": [213], "EAGLE-3.": [214], "code": [216], "available": [218], "https://github.com/AMD-AIG-AIMA/PARD.": [220]}, "primary_topic": {"id": "https://openalex.org/T10181", "display_name": "Natural Language Processing Techniques", "score": 0.8359000086784363, "subfield": {"id": "https://openalex.org/subfields/1702", "display_name": "Artificial Intelligence"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, "topics": [{"id": "https://openalex.org/T10181", "display_name": "Natural Language Processing Techniques", "score": 0.8359000086784363, "subfield": {"id": "https://openalex.org/subfields/1702", "display_name": "Artificial Intelligence"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, {"id": "https://openalex.org/T10028", "display_name": "Topic Modeling", "score": 0.12710000574588776, "subfield": {"id": "https://openalex.org/subfields/1702", "display_name": "Artificial Intelligence"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, {"id": "https://openalex.org/T10036", "display_name": "Advanced Neural Network Applications", "score": 0.00839999970048666, "subfield": {"id": "https://openalex.org/subfields/1707", "display_name": "Computer Vision and Pattern Recognition"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}], "primary_location": {"id": "pmh:oai:arXiv.org:2504.18583", "is_oa": true, "landing_page_url": "http://arxiv.org/abs/2504.18583", "pdf_url": "https://arxiv.org/pdf/2504.18583", "source": {"id": "https://openalex.org/S4306400194", "display_name": "arXiv (Cornell University)", "issn_l": "2331-8422", "issn": ["2331-8422"], "is_oa": true, "is_in_doaj": false, "is_core": false, "listed_in": [], "host_organization": "https://openalex.org/I205783295", "host_organization_name": "Cornell University", "host_organization_lineage": ["https://openalex.org/I205783295"], "host_organization_lineage_names": [], "type": "repository"}, "license": null, "license_id": null, "version": "submittedVersion", "is_accepted": false, "is_published": false, "raw_source_name": null, "raw_type": "text"}, "cited_by_count": 0, "is_retracted": false}, {"id": "https://openalex.org/W7117127821", "doi": "https://doi.org/10.5281/zenodo.18029804", "title": "A Tiered Memory Architecture for Semantic OS: From 4-bit KV Compression to Cold-Storage Swapping", "publication_year": 2025, "publication_date": "2025-12-23", "type": "preprint", "language": null, "authorships": [{"author_position": "first", "author": {"id": "https://openalex.org/A5120881706", "display_name": "Zhongren Wang", "orcid": null, "observed_orcids": []}, "institutions": [], "countries": [], "is_corresponding": true, "raw_author_name": "Wang, zhongren", "raw_affiliation_strings": [], "raw_orcid": "https://orcid.org/0009-0002-5225-091X", "affiliations": []}], "abstract_inverted_index": {"We": [0, 52], "propose": [1], "a": [2, 32, 61, 66, 77, 84, 100, 127], "novel": [3], "memory": [4, 18, 35, 94], "management": [5], "architecture": [6, 92], "for": [7], "the": [8, 16, 115, 121, 124], "Semantic": [9, 67, 128], "Operating": [10], "System": [11], "(Semantic": [12], "OS).": [13], "To": [14], "address": [15], "GPU": [17], "bottleneck": [19], "and": [20, 45, 50], "logical": [21, 108], "instability": [22], "of": [23, 103], "large": [24, 78], "generative": [25], "models": [26], "under": [27], "long-context": [28], "scenarios,": [29], "we": [30], "introduce": [31], "tiered": [33], "semantic": [34, 57, 86, 116], "model:": [36], "KV": [37], "Caches": [38], "are": [39], "compressed": [40], "via": [41], "4-bit": [42], "symmetric": [43], "quantization": [44], "dynamically": [46], "swapped": [47], "between": [48], "RAM": [49], "disk.": [51], "further": [53], "design": [54], "an": [55], "asymmetric": [56], "addressing": [58], "mechanism,": [59], "where": [60], "lightweight": [62], "model": [63, 79], "acts": [64], "as": [65], "Memory": [68], "Management": [69], "Unit": [70], "(S-MMU)": [71], "to": [72, 105], "extract": [73], "user": [74], "intent,": [75], "while": [76, 98], "performs": [80], "high-precision": [81], "recall": [82, 101], "in": [83], "shared": [85], "space.": [87], "Experiments": [88], "show": [89], "that": [90], "our": [91], "reduces": [93], "footprint": [95], "by": [96], "4\u00d7": [97], "maintaining": [99], "score": [102], "0.7390\u2014sufficient": [104], "trigger": [106], "reliable": [107], "routing.": [109], "Crucially,": [110], "when": [111], "queries": [112], "fall": [113], "outside": [114], "domain": [117], "(e.g.,": [118], "\u201cWhat": [119], "is": [120], "weather": [122], "today?\u201d),": [123], "system": [125], "issues": [126], "Kernel": [129], "Crash": [130], "Report,": [131], "preventing": [132], "hallucination": [133], "through": [134], "verifiable": [135], "rejection.": [136]}, "primary_topic": {"id": "https://openalex.org/T10054", "display_name": "Parallel Computing and Optimization Techniques", "score": 0.26809999346733093, "subfield": {"id": "https://openalex.org/subfields/1708", "display_name": "Hardware and Architecture"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, "topics": [{"id": "https://openalex.org/T10054", "display_name": "Parallel Computing and Optimization Techniques", "score": 0.26809999346733093, "subfield": {"id": "https://openalex.org/subfields/1708", "display_name": "Hardware and Architecture"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, {"id": "https://openalex.org/T10028", "display_name": "Topic Modeling", "score": 0.1378999948501587, "subfield": {"id": "https://openalex.org/subfields/1702", "display_name": "Artificial Intelligence"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, {"id": "https://openalex.org/T10181", "display_name": "Natural Language Processing Techniques", "score": 0.07270000129938126, "subfield": {"id": "https://openalex.org/subfields/1702", "display_name": "Artificial Intelligence"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}], "primary_location": {"id": "doi:10.5281/zenodo.18029804", "is_oa": true, "landing_page_url": "https://doi.org/10.5281/zenodo.18029804", "pdf_url": null, "source": {"id": "https://openalex.org/S4306400562", "display_name": "Zenodo (CERN European Organization for Nuclear Research)", "issn_l": null, "issn": null, "is_oa": true, "is_in_doaj": false, "is_core": false, "listed_in": [], "host_organization": "https://openalex.org/I67311998", "host_organization_name": "European Organization for Nuclear Research", "host_organization_lineage": ["https://openalex.org/I67311998"], "host_organization_lineage_names": [], "type": "repository"}, "license": "cc-by", "license_id": "https://openalex.org/licenses/cc-by", "version": null, "is_accepted": false, "is_published": false, "raw_source_name": null, "raw_type": "Preprint"}, "cited_by_count": 0, "is_retracted": false}, {"id": "https://openalex.org/W4414591728", "doi": "https://doi.org/10.48550/arxiv.2509.09754", "title": "LAVa: Layer-wise KV Cache Eviction with Dynamic Budget Allocation", "publication_year": 2025, "publication_date": "2025-09-11", "type": "preprint", "language": "en", "authorships": [{"author_position": "first", "author": {"id": null, "display_name": "Shen, Yiqun", "orcid": null, "observed_orcids": []}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Shen, Yiqun", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5109463615", "display_name": "Song Yuan", "orcid": null, "observed_orcids": []}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Yuan, Song", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": null, "display_name": "Zhang, Zhengze", "orcid": null, "observed_orcids": []}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Zhang, Zhengze", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": null, "display_name": "Wang, Xiaoliang", "orcid": null, "observed_orcids": []}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Wang, Xiaoliang", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5060116992", "display_name": "Daxin Jiang", "orcid": "https://orcid.org/0000-0002-6657-5806", "observed_orcids": ["https://orcid.org/0000-0002-6657-5806"]}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Jiang, Daxin", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "last", "author": {"id": "https://openalex.org/A5065000009", "display_name": "Cam-Tu Nguyen", "orcid": "https://orcid.org/0009-0006-9484-6876", "observed_orcids": ["https://orcid.org/0009-0006-9484-6876"]}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Cam-Tu, Nguyen", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}], "abstract_inverted_index": {"KV": [0], "Cache": [1], "is": [2, 95, 181], "commonly": [3], "used": [4], "to": [5, 70], "accelerate": [6], "LLM": [7], "inference": [8], "with": [9, 79, 123], "long": [10], "contexts,": [11], "yet": [12], "its": [13, 131], "high": [14], "memory": [15], "demand": [16], "drives": [17], "the": [18, 60, 96, 117], "need": [19], "for": [20, 44, 100, 145], "cache": [21, 45, 72, 101], "compression.": [22], "Existing": [23], "compression": [24, 46, 78, 169], "methods,": [25, 110], "however,": [26], "are": [27, 143], "largely": [28], "heuristic": [29], "and": [30, 65, 103, 128], "lack": [31], "dynamic": [32, 80, 91, 104, 140, 152, 168], "budget": [33, 105], "allocation.": [34], "To": [35], "address": [36], "this": [37], "limitation,": [38], "we": [39, 58, 88], "introduce": [40], "a": [41, 67, 137, 156, 166], "unified": [42, 98], "framework": [43], "by": [47, 84], "minimizing": [48], "information": [49], "loss": [50, 64], "in": [51, 159], "Transformer": [52], "residual": [53], "streams.": [54], "Building": [55], "on": [56, 114], "it,": [57], "analyze": [59], "layer": [61, 92, 141], "attention": [62], "output": [63], "derive": [66], "new": [68, 138], "metric": [69], "compare": [71], "entries": [73], "across": [74, 176], "heads,": [75], "enabling": [76], "layer-wise": [77], "head": [81, 153], "budgets.": [82, 93], "Additionally,": [83], "contrasting": [85], "cross-layer": [86], "information,": [87], "also": [89], "achieve": [90], "LAVa": [94, 171], "first": [97], "strategy": [99], "eviction": [102], "allocation": [106], "that,": [107], "unlike": [108], "prior": [109], "does": [111], "not": [112], "rely": [113], "training": [115], "or": [116], "combination": [118], "of": [119], "multiple": [120], "strategies.": [121], "Experiments": [122], "benchmarks": [124], "(LongBench,": [125], "Needle-In-A-Haystack,": [126], "Ruler,": [127], "InfiniteBench)": [129], "demonstrate": [130], "superiority.": [132], "Moreover,": [133], "our": [134], "experiments": [135], "reveal": [136], "insight:": [139], "budgets": [142, 154], "crucial": [144], "generation": [146], "tasks": [147, 161], "(e.g.,": [148, 162], "code": [149, 180], "completion),": [150], "while": [151], "play": [155], "key": [157], "role": [158], "extraction": [160], "extractive": [163], "QA).": [164], "As": [165], "fully": [167], "method,": [170], "consistently": [172], "maintains": [173], "top": [174], "performance": [175], "task": [177], "types.": [178], "Our": [179], "available": [182], "at": [183], "https://github.com/MGDDestiny/Lava.": [184]}, "primary_topic": {"id": "https://openalex.org/T10028", "display_name": "Topic Modeling", "score": 0.4578000009059906, "subfield": {"id": "https://openalex.org/subfields/1702", "display_name": "Artificial Intelligence"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, "topics": [{"id": "https://openalex.org/T10028", "display_name": "Topic Modeling", "score": 0.4578000009059906, "subfield": {"id": "https://openalex.org/subfields/1702", "display_name": "Artificial Intelligence"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, {"id": "https://openalex.org/T10181", "display_name": "Natural Language Processing Techniques", "score": 0.2709999978542328, "subfield": {"id": "https://openalex.org/subfields/1702", "display_name": "Artificial Intelligence"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, {"id": "https://openalex.org/T10036", "display_name": "Advanced Neural Network Applications", "score": 0.1356000006198883, "subfield": {"id": "https://openalex.org/subfields/1707", "display_name": "Computer Vision and Pattern Recognition"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}], "primary_location": {"id": "pmh:oai:arXiv.org:2509.09754", "is_oa": true, "landing_page_url": "http://arxiv.org/abs/2509.09754", "pdf_url": "https://arxiv.org/pdf/2509.09754", "source": {"id": "https://openalex.org/S4306400194", "display_name": "arXiv (Cornell University)", "issn_l": "2331-8422", "issn": ["2331-8422"], "is_oa": true, "is_in_doaj": false, "is_core": false, "listed_in": [], "host_organization": "https://openalex.org/I205783295", "host_organization_name": "Cornell University", "host_organization_lineage": ["https://openalex.org/I205783295"], "host_organization_lineage_names": [], "type": "repository"}, "license": null, "license_id": null, "version": "submittedVersion", "is_accepted": false, "is_published": false, "raw_source_name": null, "raw_type": "text"}, "cited_by_count": 0, "is_retracted": false}, {"id": "https://openalex.org/W4416872067", "doi": "https://doi.org/10.48550/arxiv.2508.09192", "title": "Diffusion LLMs Can Do Faster-Than-AR Inference via Discrete Diffusion Forcing", "publication_year": 2025, "publication_date": "2025-08-08", "type": "preprint", "language": "en", "authorships": [{"author_position": "first", "author": {"id": "https://openalex.org/A5059661008", "display_name": "Xu Wang", "orcid": "https://orcid.org/0000-0002-8043-7135", "observed_orcids": ["https://orcid.org/0000-0002-8043-7135"]}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Wang, Xu", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5040301444", "display_name": "Chenkai Xu", "orcid": "https://orcid.org/0000-0002-5695-8926", "observed_orcids": ["https://orcid.org/0000-0002-5695-8926"]}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Xu, Chenkai", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": null, "display_name": "Jin, Yijie", "orcid": null, "observed_orcids": []}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Jin, Yijie", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5101250200", "display_name": "Jiachun Jin", "orcid": null, "observed_orcids": []}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Jin, Jiachun", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5100397087", "display_name": "Hao Zhang", "orcid": "https://orcid.org/0009-0007-1175-5918", "observed_orcids": ["https://orcid.org/0009-0007-1175-5918"]}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Zhang, Hao", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "last", "author": {"id": "https://openalex.org/A5102623510", "display_name": "Zhijie Deng", "orcid": null, "observed_orcids": []}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Deng, Zhijie", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}], "abstract_inverted_index": {"Diffusion": [0], "Large": [1], "Language": [2], "Models": [3], "(dLLMs)": [4], "have": [5, 36], "emerged": [6], "as": [7], "a": [8, 26, 54, 127, 134], "promising": [9], "alternative": [10], "to": [11, 21, 75, 156], "autoregressive": [12, 73], "(AR)": [13], "LLMs": [14, 43], "for": [15, 91, 108], "text": [16], "generation,": [17], "with": [18, 67, 115], "the": [19, 32, 98, 163], "potential": [20], "decode": [22], "multiple": [23], "tokens": [24, 84], "in": [25], "single": [27], "iteration.": [28], "However,": [29], "none": [30], "of": [31, 44, 82, 88], "existing": [33], "open-source": [34], "dLLMs": [35, 66, 100, 142, 158], "achieved": [37], "superior": [38], "inference": [39, 147], "speed": [40, 148], "over": [41], "AR": [42], "similar": [45], "size.": [46], "This": [47], "paper": [48], "breaks": [49], "this": [50, 96], "barrier": [51], "based": [52, 120], "on": [53, 121, 153], "simple": [55], "and": [56, 138, 151, 161], "effective": [57], "strategy": [58], "named": [59], "discrete": [60], "diffusion": [61], "forcing": [62], "(D2F).": [63], "D2F": [64, 111, 141], "equips": [65], "two": [68], "key": [69], "capabilities:": [70], "(1)": [71], "block-wise": [72], "generation": [74], "enable": [76], "KV": [77], "cache": [78], "utilization;": [79], "(2)": [80], "prediction": [81], "following": [83], "without": [85], "requiring": [86], "completion": [87], "prior": [89], "blocks": [90], "inter-block": [92], "parallel": [93, 129], "decoding.": [94], "In": [95], "way,": [97], "vanilla": [99, 157], "are": [101], "refurbished": [102], "into": [103], "an": [104, 116], "AR-diffusion": [105], "hybrid": [106], "paradigm": [107], "efficient": [109], "inference.": [110], "can": [112, 165], "be": [113, 166], "implemented": [114], "asymmetric": [117], "distillation": [118], "process": [119], "pre-trained": [122], "dLLMs.": [123], "We": [124], "further": [125], "propose": [126], "pipelined": [128], "decoding": [130], "algorithm,": [131], "which": [132], "enables": [133], "trade-off": [135], "between": [136], "efficiency": [137], "efficacy.": [139], "Empirically,": [140], "achieve": [143], "more": [144, 167], "than": [145, 149, 168], "$\\mathbf{2.5\\times}$": [146], "LLaMA3": [150], "Qwen2.5": [152], "GSM8K.": [154], "Compared": [155], "like": [159], "LLaDA": [160], "Dream,": [162], "acceleration": [164], "$\\mathbf{50\\times}$": [169], "while": [170], "maintaining": [171], "comparable": [172], "output": [173], "quality.": [174], "The": [175], "code": [176], "is": [177], "available": [178], "at": [179], "https://github.com/zhijie-group/Discrete-Diffusion-Forcing.": [180]}, "primary_topic": {"id": "https://openalex.org/T10181", "display_name": "Natural Language Processing Techniques", "score": 0.6060000061988831, "subfield": {"id": "https://openalex.org/subfields/1702", "display_name": "Artificial Intelligence"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, "topics": [{"id": "https://openalex.org/T10181", "display_name": "Natural Language Processing Techniques", "score": 0.6060000061988831, "subfield": {"id": "https://openalex.org/subfields/1702", "display_name": "Artificial Intelligence"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, {"id": "https://openalex.org/T10028", "display_name": "Topic Modeling", "score": 0.3483999967575073, "subfield": {"id": "https://openalex.org/subfields/1702", "display_name": "Artificial Intelligence"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, {"id": "https://openalex.org/T10775", "display_name": "Generative Adversarial Networks and Image Synthesis", "score": 0.010900000110268593, "subfield": {"id": "https://openalex.org/subfields/1707", "display_name": "Computer Vision and Pattern Recognition"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}], "primary_location": {"id": "pmh:oai:arXiv.org:2508.09192", "is_oa": true, "landing_page_url": "http://arxiv.org/abs/2508.09192", "pdf_url": "https://arxiv.org/pdf/2508.09192", "source": {"id": "https://openalex.org/S4306400194", "display_name": "arXiv (Cornell University)", "issn_l": "2331-8422", "issn": ["2331-8422"], "is_oa": true, "is_in_doaj": false, "is_core": false, "listed_in": [], "host_organization": "https://openalex.org/I205783295", "host_organization_name": "Cornell University", "host_organization_lineage": ["https://openalex.org/I205783295"], "host_organization_lineage_names": [], "type": "repository"}, "license": null, "license_id": null, "version": "submittedVersion", "is_accepted": false, "is_published": false, "raw_source_name": null, "raw_type": "text"}, "cited_by_count": 0, "is_retracted": false}, {"id": "https://openalex.org/W7104000765", "doi": "https://doi.org/10.48550/arxiv.2511.02230", "title": "Continuum: Efficient and Robust Multi-Turn LLM Agent Scheduling with KV Cache Time-to-Live", "publication_year": 2025, "publication_date": "2025-11-04", "type": "preprint", "language": null, "authorships": [{"author_position": "first", "author": {"id": null, "display_name": "Li, Hanchen", "orcid": null, "observed_orcids": []}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Li, Hanchen", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": null, "display_name": "He, Runyuan", "orcid": null, "observed_orcids": []}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "He, Runyuan", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": null, "display_name": "Mang, Qiuyang", "orcid": null, "observed_orcids": []}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Mang, Qiuyang", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": null, "display_name": "Zhang, Qizheng", "orcid": null, "observed_orcids": []}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Zhang, Qizheng", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": null, "display_name": "Mao, Huanzhi", "orcid": null, "observed_orcids": []}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Mao, Huanzhi", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": null, "display_name": "Chen, Xiaokun", "orcid": null, "observed_orcids": []}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Chen, Xiaokun", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": null, "display_name": "Zhou, Hangrui", "orcid": null, "observed_orcids": []}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Zhou, Hangrui", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": null, "display_name": "Zhang, Huanchen", "orcid": null, "observed_orcids": []}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Zhang, Huanchen", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": null, "display_name": "Cheung, Alvin", "orcid": null, "observed_orcids": []}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Cheung, Alvin", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": null, "display_name": "Gonzalez, Joseph", "orcid": null, "observed_orcids": []}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Gonzalez, Joseph", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "last", "author": {"id": null, "display_name": "Stoica, Ion", "orcid": null, "observed_orcids": []}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Stoica, Ion", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}], "abstract_inverted_index": {"KV": [0, 18, 42, 66, 146, 159, 185], "cache": [1, 19, 67, 147, 160, 186], "management": [2], "is": [3], "essential": [4], "for": [5, 28, 137, 145, 214], "efficient": [6], "LLM": [7, 33], "inference.": [8], "To": [9], "maximize": [10], "utilization,": [11], "existing": [12], "inference": [13], "engines": [14], "evict": [15], "finished": [16], "requests'": [17], "if": [20], "new": [21], "requests": [22, 150], "are": [23], "waiting.": [24], "This": [25], "policy": [26], "breaks": [27], "agentic": [29, 215], "workloads,": [30], "which": [31], "interleave": [32], "calls": [34, 49], "with": [35, 164, 204, 224], "tools,": [36], "introducing": [37, 142], "pauses": [38], "that": [39, 151, 233], "prevent": [40], "effective": [41], "reuse": [43], "across": [44], "turns.": [45], "Since": [46], "many": [47, 73], "tool": [48, 110, 123, 153], "have": [50], "much": [51], "shorter": [52], "durations": [53], "than": [54], "human": [55], "response": [56], "multi-turn": [57, 138, 209], "chatbot,": [58], "it": [59], "would": [60], "be": [61, 188], "promising": [62], "to": [63, 79, 105, 116, 132, 191], "retain": [64], "the": [65, 82, 95, 106, 113, 158, 170, 181, 184, 236], "in": [68, 161], "during": [69], "these": [70], "tools.": [71], "However,": [72], "challenges": [74], "remain.": [75], "First,": [76], "we": [77], "need": [78], "consider": [80], "both": [81], "potential": [83, 174], "cost": [84, 172], "of": [85, 109, 122], "recomputation": [86], "or": [87], "reloading": [88], "(if": [89], "offloading": [90], "enabled)": [91], "as": [92, 94], "well": [93], "increasing": [96], "queueing": [97, 175], "delays": [98], "after": [99], "eviction": [100], "from": [101], "GPU.": [102], "Second,": [103], "due": [104], "internal": [107], "variance": [108], "call": [111, 124], "durations,": [112], "method": [114], "needs": [115], "remain": [117], "robust": [118, 197], "under": [119, 199], "limited": [120], "predictability": [121], "durations.": [125], "We": [126], "present": [127], "Continnum,": [128], "a": [129, 165], "serving": [130], "system": [131], "optimize": [133], "job": [134, 238], "completion": [135, 239], "time": [136], "agent": [139], "workloads": [140], "by": [141, 169, 178, 241], "time-to-live": [143, 166], "mechanism": [144], "retention.": [148], "For": [149], "generate": [152], "calls,": [154], "Continnum": [155, 207, 234], "selectively": [156], "pins": [157], "GPU": [162, 194], "memory": [163], "value": [167], "determined": [168], "reload": [171], "and": [173, 211, 229], "delay": [176, 213], "induced": [177], "eviction.": [179], "When": [180, 202], "TTL": [182], "expires,": [183], "can": [187], "automatically": [189], "evicted": [190], "free": [192], "up": [193], "memory,": [195], "providing": [196], "performance": [198], "edge": [200], "cases.": [201], "combined": [203], "program-level": [205], "first-come-first-serve,": [206], "preserves": [208], "continuity,": [210], "reduces": [212], "workflows.": [216], "Evaluations": [217], "on": [218], "real-world": [219], "agents": [220], "(SWE-Bench,": [221], "BFCL,": [222], "OpenHand)": [223], "Llama-3.1": [225], "8B/70B,": [226], "Gemma-3": [227], "12B,": [228], "GLM-4.5": [230], "355B": [231], "shows": [232], "improves": [235], "average": [237], "times": [240], "over": [242], "8x": [243], "while": [244], "improving": [245], "throughput.": [246]}, "primary_topic": {"id": "https://openalex.org/T10054", "display_name": "Parallel Computing and Optimization Techniques", "score": 0.37940001487731934, "subfield": {"id": "https://openalex.org/subfields/1708", "display_name": "Hardware and Architecture"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, "topics": [{"id": "https://openalex.org/T10054", "display_name": "Parallel Computing and Optimization Techniques", "score": 0.37940001487731934, "subfield": {"id": "https://openalex.org/subfields/1708", "display_name": "Hardware and Architecture"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, {"id": "https://openalex.org/T12127", "display_name": "Software System Performance and Reliability", "score": 0.2549000084400177, "subfield": {"id": "https://openalex.org/subfields/1705", "display_name": "Computer Networks and Communications"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, {"id": "https://openalex.org/T10181", "display_name": "Natural Language Processing Techniques", "score": 0.045499999076128006, "subfield": {"id": "https://openalex.org/subfields/1702", "display_name": "Artificial Intelligence"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}], "primary_location": {"id": "pmh:oai:arXiv.org:2511.02230", "is_oa": true, "landing_page_url": "https://arxiv.org/abs/2511.02230", "pdf_url": "https://arxiv.org/pdf/2511.02230", "source": {"id": "https://openalex.org/S4306400194", "display_name": "arXiv (Cornell University)", "issn_l": "2331-8422", "issn": ["2331-8422"], "is_oa": true, "is_in_doaj": false, "is_core": false, "listed_in": [], "host_organization": "https://openalex.org/I205783295", "host_organization_name": "Cornell University", "host_organization_lineage": ["https://openalex.org/I205783295"], "host_organization_lineage_names": [], "type": "repository"}, "license": "cc-by", "license_id": "https://openalex.org/licenses/cc-by", "version": "submittedVersion", "is_accepted": false, "is_published": false, "raw_source_name": null, "raw_type": "text"}, "cited_by_count": 0, "is_retracted": false}, {"id": "https://openalex.org/W4416414696", "doi": "https://doi.org/10.48550/arxiv.2511.09844", "title": "Steering Pretrained Drafters during Speculative Decoding", "publication_year": 2025, "publication_date": "2025-11-13", "type": "preprint", "language": "en", "authorships": [{"author_position": "first", "author": {"id": "https://openalex.org/A5115003228", "display_name": "Fr\u00e9d\u00e9ric Berdoz", "orcid": null, "observed_orcids": []}, "institutions": [{"id": "https://openalex.org/I35440088", "display_name": "ETH Zurich", "ror": "https://ror.org/05a28rw58", "country_code": "CH", "type": "education", "lineage": ["https://openalex.org/I2799323385", "https://openalex.org/I35440088"]}], "countries": ["CH"], "is_corresponding": false, "raw_author_name": "Berdoz, Fr\u00e9d\u00e9ric", "raw_affiliation_strings": ["ETH Zurich"], "raw_orcid": null, "affiliations": [{"raw_affiliation_string": "ETH Zurich", "institution_ids": ["https://openalex.org/I35440088"]}]}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5120414886", "display_name": "Peer Rheinboldt", "orcid": null, "observed_orcids": []}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Rheinboldt, Peer", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "last", "author": {"id": "https://openalex.org/A5078339613", "display_name": "Roger P. Wattenhofer", "orcid": "https://orcid.org/0000-0002-6339-3134", "observed_orcids": ["https://orcid.org/0000-0002-6339-3134"]}, "institutions": [{"id": "https://openalex.org/I35440088", "display_name": "ETH Zurich", "ror": "https://ror.org/05a28rw58", "country_code": "CH", "type": "education", "lineage": ["https://openalex.org/I2799323385", "https://openalex.org/I35440088"]}], "countries": ["CH"], "is_corresponding": false, "raw_author_name": "Wattenhofer, Roger", "raw_affiliation_strings": ["ETH Zurich"], "raw_orcid": null, "affiliations": [{"raw_affiliation_string": "ETH Zurich", "institution_ids": ["https://openalex.org/I35440088"]}]}], "abstract_inverted_index": {"Speculative": [0], "decoding": [1], "accelerates": [2], "language": [3], "model": [4], "inference": [5], "by": [6, 95, 134], "separating": [7], "generation": [8, 66], "into": [9, 113], "fast": [10], "drafting": [11, 31, 72], "and": [12, 25, 111, 141, 161], "parallel": [13], "verification.": [14], "Its": [15], "main": [16], "limitation": [17], "is": [18, 74], "drafter-verifier": [19], "misalignment,": [20], "which": [21], "limits": [22], "token": [23], "acceptance": [24, 60, 90], "reduces": [26], "overall": [27], "effectiveness.": [28], "While": [29], "small": [30], "heads": [32], "trained": [33], "from": [34, 106], "scratch": [35], "compensate": [36], "with": [37], "speed,": [38], "they": [39], "struggle": [40], "when": [41, 46, 71], "verification": [42, 78], "dominates": [43], "latency": [44, 73], "or": [45, 79], "inputs": [47], "are": [48], "out": [49], "of": [50, 92, 131], "distribution.": [51], "In": [52, 82], "contrast,": [53], "pretrained": [54, 93, 115, 162], "drafters,": [55], "though": [56], "slower,": [57], "achieve": [58], "higher": [59], "rates": [61, 91], "thanks": [62], "to": [63, 77, 87, 118, 136, 158], "stronger": [64], "standalone": [65], "capabilities,": [67], "making": [68], "them": [69], "competitive": [70], "negligible": [75, 149], "relative": [76], "communication": [80], "overhead.": [81, 151], "this": [83], "work,": [84], "we": [85], "aim": [86], "improve": [88], "the": [89, 107, 114, 129], "drafters": [94], "introducing": [96], "a": [97, 102], "lightweight": [98], "dynamic": [99], "alignment": [100, 121], "mechanism:": [101], "steering": [103], "vector": [104], "computed": [105], "verifier's": [108], "hidden": [109], "states": [110], "injected": [112], "drafter.": [116], "Compared": [117], "existing": [119, 159], "offline": [120], "methods": [122], "such": [123], "as": [124], "distillation,": [125], "our": [126, 153], "approach": [127, 154], "boosts": [128], "number": [130], "accepted": [132], "tokens": [133], "up": [135], "35\\%": [137], "under": [138, 143], "standard": [139], "sampling": [140], "22\\%": [142], "greedy": [144], "sampling,": [145], "all": [146], "while": [147], "incurring": [148], "computational": [150], "Importantly,": [152], "can": [155], "be": [156], "retrofitted": [157], "architectures": [160], "models,": [163], "enabling": [164], "rapid": [165], "adoption.": [166]}, "primary_topic": {"id": "https://openalex.org/T10181", "display_name": "Natural Language Processing Techniques", "score": 0.6845999956130981, "subfield": {"id": "https://openalex.org/subfields/1702", "display_name": "Artificial Intelligence"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, "topics": [{"id": "https://openalex.org/T10181", "display_name": "Natural Language Processing Techniques", "score": 0.6845999956130981, "subfield": {"id": "https://openalex.org/subfields/1702", "display_name": "Artificial Intelligence"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, {"id": "https://openalex.org/T10028", "display_name": "Topic Modeling", "score": 0.2782999873161316, "subfield": {"id": "https://openalex.org/subfields/1702", "display_name": "Artificial Intelligence"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, {"id": "https://openalex.org/T10201", "display_name": "Speech Recognition and Synthesis", "score": 0.009999999776482582, "subfield": {"id": "https://openalex.org/subfields/1702", "display_name": "Artificial Intelligence"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}], "primary_location": {"id": "pmh:oai:arXiv.org:2511.09844", "is_oa": true, "landing_page_url": "http://arxiv.org/abs/2511.09844", "pdf_url": "https://arxiv.org/pdf/2511.09844", "source": {"id": "https://openalex.org/S4306400194", "display_name": "arXiv (Cornell University)", "issn_l": "2331-8422", "issn": ["2331-8422"], "is_oa": true, "is_in_doaj": false, "is_core": false, "listed_in": [], "host_organization": "https://openalex.org/I205783295", "host_organization_name": "Cornell University", "host_organization_lineage": ["https://openalex.org/I205783295"], "host_organization_lineage_names": [], "type": "repository"}, "license": "cc-by", "license_id": "https://openalex.org/licenses/cc-by", "version": "submittedVersion", "is_accepted": false, "is_published": false, "raw_source_name": null, "raw_type": "text"}, "cited_by_count": 0, "is_retracted": false}, {"id": "https://openalex.org/W4415948549", "doi": "https://doi.org/10.48550/arxiv.2510.15980", "title": "Cognitive Load Traces as Symbolic and Visual Accounts of Deep Model Cognition", "publication_year": 2025, "publication_date": "2025-10-13", "type": "preprint", "language": null, "authorships": [{"author_position": "first", "author": {"id": "https://openalex.org/A5100407529", "display_name": "Dong Liu", "orcid": "https://orcid.org/0000-0003-3988-7879", "observed_orcids": ["https://orcid.org/0000-0003-3988-7879"]}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Liu, Dong", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "last", "author": {"id": null, "display_name": "Yu, Yanxuan", "orcid": null, "observed_orcids": []}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Yu, Yanxuan", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}], "abstract_inverted_index": {"We": [0, 73], "propose": [1, 74], "\\textbf{Cognitive": [2], "Load": [3, 17], "Traces}": [4], "(CLTs)": [5], "as": [6, 25, 39, 62], "a": [7, 40], "mid-level": [8], "interpretability": [9], "framework": [10], "for": [11], "deep": [12], "models,": [13], "inspired": [14], "by": [15, 114], "Cognitive": [16], "Theory": [18], "in": [19], "human": [20], "cognition.": [21], "CLTs": [22, 38, 100], "are": [23], "defined": [24], "symbolic,": [26], "temporally": [27], "varying": [28], "functions": [29], "that": [30, 85, 99, 110], "quantify": [31], "model-internal": [32], "resource": [33], "allocation.": [34], "Formally,": [35], "we": [36], "represent": [37], "three-component": [41], "stochastic": [42], "process": [43], "$(\\mathrm{IL}_t,": [44], "\\mathrm{EL}_t,": [45], "\\mathrm{GL}_t)$,": [46], "corresponding": [47], "to": [48], "\\emph{Intrinsic},": [49], "\\emph{Extraneous},": [50], "and": [51, 70, 78, 95, 106], "\\emph{Germane}": [52], "load.": [53], "Each": [54], "component": [55], "is": [56], "instantiated": [57], "through": [58], "measurable": [59], "proxies": [60], "such": [61], "attention": [63], "entropy,": [64], "KV-cache": [65], "miss": [66], "ratio,": [67], "representation": [68], "dispersion,": [69], "decoding": [71], "stability.": [72], "both": [75], "symbolic": [76], "formulations": [77], "visualization": [79], "methods": [80], "(load": [81], "curves,": [82], "simplex": [83], "diagrams)": [84], "enable": [86, 107], "interpretable": [87], "analysis": [88], "of": [89], "reasoning": [90, 94, 112], "dynamics.": [91], "Experiments": [92], "on": [93], "planning": [96], "benchmarks": [97], "show": [98], "predict": [101], "error-onset,": [102], "reveal": [103], "cognitive": [104], "strategies,": [105], "load-guided": [108], "interventions": [109], "improve": [111], "efficiency": [113], "15-30\\%": [115], "while": [116], "maintaining": [117], "accuracy.": [118]}, "primary_topic": {"id": "https://openalex.org/T12026", "display_name": "Explainable Artificial Intelligence (XAI)", "score": 0.9990000128746033, "subfield": {"id": "https://openalex.org/subfields/1702", "display_name": "Artificial Intelligence"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, "topics": [{"id": "https://openalex.org/T12026", "display_name": "Explainable Artificial Intelligence (XAI)", "score": 0.9990000128746033, "subfield": {"id": "https://openalex.org/subfields/1702", "display_name": "Artificial Intelligence"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, {"id": "https://openalex.org/T11516", "display_name": "Visual and Cognitive Learning Processes", "score": 0.00019999999494757503, "subfield": {"id": "https://openalex.org/subfields/3205", "display_name": "Experimental and Cognitive Psychology"}, "field": {"id": "https://openalex.org/fields/32", "display_name": "Psychology"}, "domain": {"id": "https://openalex.org/domains/2", "display_name": "Social Sciences"}}, {"id": "https://openalex.org/T10036", "display_name": "Advanced Neural Network Applications", "score": 9.999999747378752e-05, "subfield": {"id": "https://openalex.org/subfields/1707", "display_name": "Computer Vision and Pattern Recognition"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}], "primary_location": {"id": "pmh:oai:arXiv.org:2510.15980", "is_oa": true, "landing_page_url": "http://arxiv.org/abs/2510.15980", "pdf_url": "https://arxiv.org/pdf/2510.15980", "source": {"id": "https://openalex.org/S4306400194", "display_name": "arXiv (Cornell University)", "issn_l": "2331-8422", "issn": ["2331-8422"], "is_oa": true, "is_in_doaj": false, "is_core": false, "listed_in": [], "host_organization": "https://openalex.org/I205783295", "host_organization_name": "Cornell University", "host_organization_lineage": ["https://openalex.org/I205783295"], "host_organization_lineage_names": [], "type": "repository"}, "license": null, "license_id": null, "version": "submittedVersion", "is_accepted": false, "is_published": false, "raw_source_name": null, "raw_type": "text"}, "cited_by_count": 0, "is_retracted": false}, {"id": "https://openalex.org/W7117117156", "doi": "https://doi.org/10.48550/arxiv.2512.19179", "title": "CascadeInfer: Length-Aware Scheduling of LLM Serving with Low Latency and Load Balancing", "publication_year": 2025, "publication_date": "2025-12-22", "type": "preprint", "language": null, "authorships": [{"author_position": "first", "author": {"id": "https://openalex.org/A5121229933", "display_name": "Yitao Yuan", "orcid": "https://orcid.org/0009-0005-5427-4829", "observed_orcids": ["https://orcid.org/0009-0005-5427-4829"]}, "institutions": [{"id": "https://openalex.org/I20231570", "display_name": "Peking University", "ror": "https://ror.org/02v51f717", "country_code": "CN", "type": "education", "lineage": ["https://openalex.org/I20231570"]}], "countries": ["CN"], "is_corresponding": false, "raw_author_name": "Yuan, Yitao", "raw_affiliation_strings": ["Peking University", "ScitiX AI"], "raw_orcid": null, "affiliations": [{"raw_affiliation_string": "Peking University", "institution_ids": ["https://openalex.org/I20231570"]}, {"raw_affiliation_string": "ScitiX AI", "institution_ids": []}]}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5101140193", "display_name": "Chenqi Zhao", "orcid": "https://orcid.org/0009-0006-9540-1940", "observed_orcids": ["https://orcid.org/0009-0006-9540-1940"]}, "institutions": [{"id": "https://openalex.org/I20231570", "display_name": "Peking University", "ror": "https://ror.org/02v51f717", "country_code": "CN", "type": "education", "lineage": ["https://openalex.org/I20231570"]}], "countries": ["CN"], "is_corresponding": false, "raw_author_name": "Zhao, Chenqi", "raw_affiliation_strings": ["Peking University"], "raw_orcid": null, "affiliations": [{"raw_affiliation_string": "Peking University", "institution_ids": ["https://openalex.org/I20231570"]}]}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5121131782", "display_name": "Bohan Zhao", "orcid": null, "observed_orcids": []}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Zhao, Bohan", "raw_affiliation_strings": ["ScitiX AI"], "raw_orcid": null, "affiliations": [{"raw_affiliation_string": "ScitiX AI", "institution_ids": []}]}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5121128285", "display_name": "Zane Cao", "orcid": null, "observed_orcids": []}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Cao, Zane", "raw_affiliation_strings": ["ScitiX AI"], "raw_orcid": null, "affiliations": [{"raw_affiliation_string": "ScitiX AI", "institution_ids": []}]}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5121185961", "display_name": "Yongchao He", "orcid": null, "observed_orcids": []}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "He, Yongchao", "raw_affiliation_strings": ["ScitiX AI"], "raw_orcid": null, "affiliations": [{"raw_affiliation_string": "ScitiX AI", "institution_ids": []}]}, {"author_position": "last", "author": {"id": "https://openalex.org/A5121193143", "display_name": "Wenfei Wu", "orcid": "https://orcid.org/0000-0002-1357-3137", "observed_orcids": ["https://orcid.org/0000-0002-1357-3137"]}, "institutions": [{"id": "https://openalex.org/I20231570", "display_name": "Peking University", "ror": "https://ror.org/02v51f717", "country_code": "CN", "type": "education", "lineage": ["https://openalex.org/I20231570"]}], "countries": ["CN"], "is_corresponding": false, "raw_author_name": "Wu, Wenfei", "raw_affiliation_strings": ["Peking University"], "raw_orcid": null, "affiliations": [{"raw_affiliation_string": "Peking University", "institution_ids": ["https://openalex.org/I20231570"]}]}], "abstract_inverted_index": {"Efficiently": [0], "harnessing": [1], "GPU": [2, 61], "compute": [3], "is": [4], "critical": [5], "to": [6, 30, 83, 118, 163, 170, 179, 183], "improving": [7, 173], "user": [8], "experience": [9], "and": [10, 63, 139, 145, 165], "reducing": [11], "operational": [12], "costs": [13], "in": [14], "large": [15], "language": [16], "model": [17], "(LLM)": [18], "services.": [19], "However,": [20], "current": [21], "inference": [22], "engine": [23], "schedulers": [24], "overlook": [25], "the": [26, 80, 121, 125, 154, 184], "attention": [27], "backend's": [28], "sensitivity": [29], "request-length": [31], "heterogeneity": [32], "within": [33, 98, 140], "a": [34, 52, 69, 99, 105, 114, 143], "batch.": [35], "As": [36], "state-of-the-art": [37, 185], "models": [38], "now": [39], "support": [40], "context": [41], "windows": [42], "exceeding": [43], "128K": [44], "tokens,": [45], "this": [46], "once-tolerable": [47], "inefficiency": [48], "has": [49], "escalated": [50], "into": [51, 92], "primary": [53], "system": [54, 71, 175], "bottleneck,": [55], "causing": [56], "severe": [57], "performance": [58], "degradation": [59], "through": [60, 110], "underutilization": [62], "increased": [64], "latency.": [65], "We": [66], "present": [67], "CascadeInfer,": [68], "runtime": [70, 129], "that": [72], "dynamically": [73], "reschedules": [74], "requests": [75, 97, 108], "across": [76, 138], "multiple": [77], "instances": [78, 91], "serving": [79], "same": [81, 155], "LLM": [82], "mitigate": [84], "per-instance": [85], "length": [86, 101], "heterogeneity.": [87], "CascadeInfer": [88, 112, 157], "partitions": [89], "these": [90], "length-specialized": [93], "groups,": [94, 141], "each": [95], "handling": [96], "designated": [100], "range,": [102], "naturally": [103], "forming": [104], "pipeline": [106], "as": [107], "flow": [109], "them.": [111], "devises": [113], "dynamic": [115], "programming": [116], "algorithm": [117], "efficiently": [119], "find": [120], "stage": [122], "partition": [123], "with": [124, 133], "best": [126], "QoE,": [127], "employs": [128], "range": [130], "refinement": [131], "together": [132], "decentralized": [134], "load": [135], "(re)balance": [136], "both": [137], "achieving": [142], "balanced": [144], "efficient": [146], "multi-instance": [147, 186], "service.": [148], "Our": [149], "evaluation": [150], "shows": [151], "that,": [152], "under": [153], "configuration,": [156], "reduces": [158], "end-to-end": [159], "latency": [160, 167], "by": [161, 168, 177], "up": [162, 169, 178], "67%": [164], "tail": [166], "69%,": [171], "while": [172], "overall": [174], "throughput": [176], "2.89": [180], "times": [181], "compared": [182], "scheduling": [187], "systems.": [188]}, "primary_topic": {"id": "https://openalex.org/T10054", "display_name": "Parallel Computing and Optimization Techniques", "score": 0.40529999136924744, "subfield": {"id": "https://openalex.org/subfields/1708", "display_name": "Hardware and Architecture"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, "topics": [{"id": "https://openalex.org/T10054", "display_name": "Parallel Computing and Optimization Techniques", "score": 0.40529999136924744, "subfield": {"id": "https://openalex.org/subfields/1708", "display_name": "Hardware and Architecture"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, {"id": "https://openalex.org/T10101", "display_name": "Cloud Computing and Resource Management", "score": 0.30660000443458557, "subfield": {"id": "https://openalex.org/subfields/1710", "display_name": "Information Systems"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, {"id": "https://openalex.org/T10715", "display_name": "Distributed and Parallel Computing Systems", "score": 0.1014999970793724, "subfield": {"id": "https://openalex.org/subfields/1705", "display_name": "Computer Networks and Communications"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}], "primary_location": {"id": "pmh:oai:arXiv.org:2512.19179", "is_oa": true, "landing_page_url": "https://arxiv.org/abs/2512.19179", "pdf_url": "https://arxiv.org/pdf/2512.19179", "source": {"id": "https://openalex.org/S4306400194", "display_name": "arXiv (Cornell University)", "issn_l": "2331-8422", "issn": ["2331-8422"], "is_oa": true, "is_in_doaj": false, "is_core": false, "listed_in": [], "host_organization": "https://openalex.org/I205783295", "host_organization_name": "Cornell University", "host_organization_lineage": ["https://openalex.org/I205783295"], "host_organization_lineage_names": [], "type": "repository"}, "license": "cc-by", "license_id": "https://openalex.org/licenses/cc-by", "version": "submittedVersion", "is_accepted": false, "is_published": false, "raw_source_name": null, "raw_type": "text"}, "cited_by_count": 0, "is_retracted": false}, {"id": "https://openalex.org/W4412722459", "doi": "https://doi.org/10.1109/mm.2025.3592570", "title": "Speculative Decoding on the SN40L Reconfigurable Dataflow Unit", "publication_year": 2025, "publication_date": "2025-07-29", "type": "article", "language": "en", "authorships": [{"author_position": "first", "author": {"id": null, "display_name": "Darshan Gandhi", "orcid": "https://orcid.org/0009-0002-9442-4228", "observed_orcids": ["https://orcid.org/0009-0002-9442-4228"]}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Darshan Gandhi", "raw_affiliation_strings": ["SambaNova Systems, Inc., Palo Alto, CA, USA", "SambaNova Systems Inc, Palo Alto, CA, USA"], "raw_orcid": "https://orcid.org/0009-0002-9442-4228", "affiliations": [{"raw_affiliation_string": "SambaNova Systems, Inc., Palo Alto, CA, USA", "institution_ids": []}, {"raw_affiliation_string": "SambaNova Systems Inc, Palo Alto, CA, USA", "institution_ids": []}]}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5114635547", "display_name": "Pushkar Nandkar", "orcid": "https://orcid.org/0009-0004-5810-2162", "observed_orcids": ["https://orcid.org/0009-0004-5810-2162"]}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Pushkar Nandkar", "raw_affiliation_strings": ["SambaNova Systems, Inc., Palo Alto, CA, USA", "SambaNova Systems Inc, Palo Alto, CA, USA"], "raw_orcid": "https://orcid.org/0009-0004-5810-2162", "affiliations": [{"raw_affiliation_string": "SambaNova Systems, Inc., Palo Alto, CA, USA", "institution_ids": []}, {"raw_affiliation_string": "SambaNova Systems Inc, Palo Alto, CA, USA", "institution_ids": []}]}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5081247884", "display_name": "Nasim Farahini", "orcid": "https://orcid.org/0009-0000-5072-4152", "observed_orcids": ["https://orcid.org/0009-0000-5072-4152"]}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Nasim Farahini", "raw_affiliation_strings": ["SambaNova Systems, Inc., Palo Alto, CA, USA", "SambaNova Systems Inc, Palo Alto, CA, USA"], "raw_orcid": "https://orcid.org/0009-0000-5072-4152", "affiliations": [{"raw_affiliation_string": "SambaNova Systems, Inc., Palo Alto, CA, USA", "institution_ids": []}, {"raw_affiliation_string": "SambaNova Systems Inc, Palo Alto, CA, USA", "institution_ids": []}]}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5075185447", "display_name": "H\u00e5kan Zeffer", "orcid": "https://orcid.org/0009-0001-3947-0478", "observed_orcids": ["https://orcid.org/0009-0001-3947-0478"]}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "H\u00e5kan Zeffer", "raw_affiliation_strings": ["SambaNova Systems, Inc., Palo Alto, CA, USA", "SambaNova Systems Inc, Palo Alto, CA, USA"], "raw_orcid": "https://orcid.org/0009-0001-3947-0478", "affiliations": [{"raw_affiliation_string": "SambaNova Systems, Inc., Palo Alto, CA, USA", "institution_ids": []}, {"raw_affiliation_string": "SambaNova Systems Inc, Palo Alto, CA, USA", "institution_ids": []}]}, {"author_position": "middle", "author": {"id": null, "display_name": "John Long", "orcid": "https://orcid.org/0000-0001-7406-1410", "observed_orcids": ["https://orcid.org/0000-0001-7406-1410"]}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "John Long", "raw_affiliation_strings": ["SambaNova Systems, Inc., Palo Alto, CA, USA", "SambaNova Systems Inc, Palo Alto, CA, USA"], "raw_orcid": "https://orcid.org/0000-0001-7406-1410", "affiliations": [{"raw_affiliation_string": "SambaNova Systems, Inc., Palo Alto, CA, USA", "institution_ids": []}, {"raw_affiliation_string": "SambaNova Systems Inc, Palo Alto, CA, USA", "institution_ids": []}]}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5119102216", "display_name": "Samuel Rydh", "orcid": "https://orcid.org/0009-0000-6582-9416", "observed_orcids": ["https://orcid.org/0009-0000-6582-9416"]}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Samuel Rydh", "raw_affiliation_strings": ["SambaNova Systems, Inc., Palo Alto, CA, USA", "SambaNova Systems Inc, Palo Alto, CA, USA"], "raw_orcid": "https://orcid.org/0009-0000-6582-9416", "affiliations": [{"raw_affiliation_string": "SambaNova Systems, Inc., Palo Alto, CA, USA", "institution_ids": []}, {"raw_affiliation_string": "SambaNova Systems Inc, Palo Alto, CA, USA", "institution_ids": []}]}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5114635548", "display_name": "Matheen Musaddiq", "orcid": "https://orcid.org/0009-0003-9118-385X", "observed_orcids": ["https://orcid.org/0009-0003-9118-385X"]}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Matheen Musaddiq", "raw_affiliation_strings": ["SambaNova Systems, Inc., Palo Alto, CA, USA", "SambaNova Systems Inc, Palo Alto, CA, USA"], "raw_orcid": "https://orcid.org/0009-0003-9118-385X", "affiliations": [{"raw_affiliation_string": "SambaNova Systems, Inc., Palo Alto, CA, USA", "institution_ids": []}, {"raw_affiliation_string": "SambaNova Systems Inc, Palo Alto, CA, USA", "institution_ids": []}]}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5033875105", "display_name": "Tuowen Zhao", "orcid": "https://orcid.org/0000-0002-6347-7135", "observed_orcids": ["https://orcid.org/0000-0002-6347-7135"]}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Tuowen Zhao", "raw_affiliation_strings": ["SambaNova Systems, Inc., Palo Alto, CA, USA", "SambaNova Systems Inc, Palo Alto, CA, USA"], "raw_orcid": "https://orcid.org/0000-0002-6347-7135", "affiliations": [{"raw_affiliation_string": "SambaNova Systems, Inc., Palo Alto, CA, USA", "institution_ids": []}, {"raw_affiliation_string": "SambaNova Systems Inc, Palo Alto, CA, USA", "institution_ids": []}]}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5098316663", "display_name": "Joshua Brot", "orcid": "https://orcid.org/0009-0007-6959-765X", "observed_orcids": ["https://orcid.org/0009-0007-6959-765X"]}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Joshua Brot", "raw_affiliation_strings": ["SambaNova Systems, Inc., Palo Alto, CA, USA", "SambaNova Systems Inc, Palo Alto, CA, USA"], "raw_orcid": "https://orcid.org/0009-0007-6959-765X", "affiliations": [{"raw_affiliation_string": "SambaNova Systems, Inc., Palo Alto, CA, USA", "institution_ids": []}, {"raw_affiliation_string": "SambaNova Systems Inc, Palo Alto, CA, USA", "institution_ids": []}]}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5114635549", "display_name": "Reid Goodbar", "orcid": null, "observed_orcids": []}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Reid Goodbar", "raw_affiliation_strings": ["SambaNova Systems, Inc., Palo Alto, CA, USA", "SambaNova Systems Inc, Palo Alto, CA, USA"], "raw_orcid": null, "affiliations": [{"raw_affiliation_string": "SambaNova Systems, Inc., Palo Alto, CA, USA", "institution_ids": []}, {"raw_affiliation_string": "SambaNova Systems Inc, Palo Alto, CA, USA", "institution_ids": []}]}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5100928662", "display_name": "Yun Du", "orcid": "https://orcid.org/0009-0001-6723-4384", "observed_orcids": ["https://orcid.org/0009-0001-6723-4384"]}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Yun Du", "raw_affiliation_strings": ["SambaNova Systems, Inc., Palo Alto, CA, USA", "SambaNova Systems Inc, Palo Alto, CA, USA"], "raw_orcid": "https://orcid.org/0009-0001-6723-4384", "affiliations": [{"raw_affiliation_string": "SambaNova Systems, Inc., Palo Alto, CA, USA", "institution_ids": []}, {"raw_affiliation_string": "SambaNova Systems Inc, Palo Alto, CA, USA", "institution_ids": []}]}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5112914114", "display_name": "Mingran Wang", "orcid": "https://orcid.org/0009-0006-8119-2201", "observed_orcids": ["https://orcid.org/0009-0006-8119-2201"]}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Mingran Wang", "raw_affiliation_strings": ["SambaNova Systems, Inc., Palo Alto, CA, USA", "SambaNova Systems Inc, Palo Alto, CA, USA"], "raw_orcid": "https://orcid.org/0009-0006-8119-2201", "affiliations": [{"raw_affiliation_string": "SambaNova Systems, Inc., Palo Alto, CA, USA", "institution_ids": []}, {"raw_affiliation_string": "SambaNova Systems Inc, Palo Alto, CA, USA", "institution_ids": []}]}, {"author_position": "last", "author": {"id": "https://openalex.org/A5111746097", "display_name": "Raghu Prabhakar", "orcid": "https://orcid.org/0000-0003-0230-4377", "observed_orcids": ["https://orcid.org/0000-0003-0230-4377"]}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Raghu Prabhakar", "raw_affiliation_strings": ["SambaNova Systems, Inc., Palo Alto, CA, USA", "SambaNova Systems Inc, Palo Alto, CA, USA"], "raw_orcid": "https://orcid.org/0000-0003-0230-4377", "affiliations": [{"raw_affiliation_string": "SambaNova Systems, Inc., Palo Alto, CA, USA", "institution_ids": []}, {"raw_affiliation_string": "SambaNova Systems Inc, Palo Alto, CA, USA", "institution_ids": []}]}], "abstract_inverted_index": null, "primary_topic": {"id": "https://openalex.org/T10904", "display_name": "Embedded Systems Design Techniques", "score": 0.4805000126361847, "subfield": {"id": "https://openalex.org/subfields/1708", "display_name": "Hardware and Architecture"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, "topics": [{"id": "https://openalex.org/T10904", "display_name": "Embedded Systems Design Techniques", "score": 0.4805000126361847, "subfield": {"id": "https://openalex.org/subfields/1708", "display_name": "Hardware and Architecture"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, {"id": "https://openalex.org/T10054", "display_name": "Parallel Computing and Optimization Techniques", "score": 0.45969998836517334, "subfield": {"id": "https://openalex.org/subfields/1708", "display_name": "Hardware and Architecture"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, {"id": "https://openalex.org/T10829", "display_name": "Interconnection Networks and Systems", "score": 0.014399999752640724, "subfield": {"id": "https://openalex.org/subfields/1705", "display_name": "Computer Networks and Communications"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}], "primary_location": {"id": "doi:10.1109/mm.2025.3592570", "is_oa": false, "landing_page_url": "https://doi.org/10.1109/mm.2025.3592570", "pdf_url": null, "source": {"id": "https://openalex.org/S59697426", "display_name": "IEEE Micro", "issn_l": "0272-1732", "issn": ["0272-1732", "1937-4143"], "is_oa": false, "is_in_doaj": false, "is_core": true, "listed_in": ["cwts-core", "jufo-3", "norway-2"], "host_organization": "https://openalex.org/P4310319808", "host_organization_name": "Institute of Electrical and Electronics Engineers", "host_organization_lineage": ["https://openalex.org/P4310319808"], "host_organization_lineage_names": ["Institute of Electrical and Electronics Engineers"], "type": "journal"}, "license": null, "license_id": null, "version": "publishedVersion", "is_accepted": true, "is_published": true, "raw_source_name": "IEEE Micro", "raw_type": "journal-article"}, "cited_by_count": 0, "is_retracted": false}, {"id": "https://openalex.org/W4417102179", "doi": "https://doi.org/10.48550/arxiv.2508.02123", "title": "Understanding the Essence: Delving into Annotator Prototype Learning for Multi-Class Annotation Aggregation", "publication_year": 2025, "publication_date": "2025-08-04", "type": "preprint", "language": "en", "authorships": [{"author_position": "first", "author": {"id": "https://openalex.org/A5100774581", "display_name": "Ju Chen", "orcid": "https://orcid.org/0009-0001-6957-7292", "observed_orcids": ["https://orcid.org/0009-0001-6957-7292", "https://orcid.org/0009-0005-9731-3365"]}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Chen, Ju", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5065134183", "display_name": "Jun Feng", "orcid": "https://orcid.org/0000-0002-9790-5353", "observed_orcids": ["https://orcid.org/0000-0002-9790-5353"]}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Feng, Jun", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "last", "author": {"id": "https://openalex.org/A5036219211", "display_name": "Shenyu Zhang", "orcid": "https://orcid.org/0009-0005-3864-8907", "observed_orcids": ["https://orcid.org/0009-0005-3864-8907"]}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Zhang, Shenyu", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}], "abstract_inverted_index": {"Multi-class": [0], "classification": [1], "annotations": [2], "have": [3], "significantly": [4], "advanced": [5], "AI": [6], "applications,": [7], "with": [8], "truth": [9], "inference": [10], "serving": [11], "as": [12, 143], "a": [13, 31, 49, 65, 88, 100, 115, 133, 144, 164, 184, 193], "critical": [14], "technique": [15], "for": [16, 72], "aggregating": [17], "noisy": [18], "and": [19, 63, 102, 159, 166, 192], "biased": [20], "annotations.": [21], "Existing": [22], "state-of-the-art": [23], "methods": [24, 36], "typically": [25], "model": [26], "each": [27, 74], "annotator's": [28, 75], "expertise": [29, 77, 126, 138], "using": [30], "confusion": [32, 59, 67, 120, 135], "matrix.": [33], "However,": [34], "these": [35, 84, 149], "suffer": [37], "from": [38], "two": [39], "widely": [40], "recognized": [41], "issues:": [42], "1)": [43], "when": [44, 53], "most": [45], "annotators": [46], "label": [47], "only": [48], "few": [50], "tasks,": [51], "or": [52], "classes": [54], "are": [55, 61], "imbalanced,": [56], "the": [57, 124, 137, 156, 189], "estimated": [58], "matrices": [60], "unreliable,": [62], "2)": [64], "single": [66, 134], "matrix": [68], "often": [69], "remains": [70], "inadequate": [71], "capturing": [73], "full": [76], "patterns": [78, 127], "across": [79], "all": [80, 129], "tasks.": [81], "To": [82], "address": [83], "issues,": [85, 162], "we": [86, 110], "propose": [87], "novel": [89], "confusion-matrix-based": [90], "method,": [91], "PTBCC": [92, 180], "(ProtoType": [93], "learning-driven": [94, 153], "Bayesian": [95], "Classifier": [96], "Combination),": [97], "to": [98, 183], "introduce": [99], "reliable": [101], "richer": [103, 165], "annotator": [104, 140], "estimation": [105], "by": [106, 202], "prototype": [107, 119, 152], "learning.": [108], "Specifically,": [109], "assume": [111], "that": [112, 179], "there": [113], "exists": [114], "set": [116], "$S$": [117], "of": [118, 128, 170], "matrices,": [121], "which": [122], "capture": [123], "inherent": [125], "annotators.": [130, 171], "Rather": [131], "than": [132], "matrix,": [136], "per": [139], "is": [141], "extended": [142], "Dirichlet": [145], "prior": [146], "distribution": [147], "over": [148, 203], "prototypes.": [150], "This": [151], "mechanism": [154], "circumvents": [155], "data": [157], "sparsity": [158], "class": [160], "imbalance": [161], "ensuring": [163], "more": [167], "flexible": [168], "characterization": [169], "Extensive": [172], "experiments": [173], "on": [174], "11": [175], "real-world": [176], "datasets": [177], "demonstrate": [178], "achieves": [181], "up": [182], "15%": [185], "accuracy": [186, 197], "improvement": [187], "in": [188], "best": [190], "case,": [191], "3%": [194], "higher": [195], "average": [196], "while": [198], "reducing": [199], "computational": [200], "cost": [201], "90%.": [204]}, "primary_topic": {"id": "https://openalex.org/T12535", "display_name": "Machine Learning and Data Classification", "score": 0.9775000214576721, "subfield": {"id": "https://openalex.org/subfields/1702", "display_name": "Artificial Intelligence"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, "topics": [{"id": "https://openalex.org/T12535", "display_name": "Machine Learning and Data Classification", "score": 0.9775000214576721, "subfield": {"id": "https://openalex.org/subfields/1702", "display_name": "Artificial Intelligence"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, {"id": "https://openalex.org/T12072", "display_name": "Machine Learning and Algorithms", "score": 0.004399999976158142, "subfield": {"id": "https://openalex.org/subfields/1702", "display_name": "Artificial Intelligence"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, {"id": "https://openalex.org/T11704", "display_name": "Mobile Crowdsensing and Crowdsourcing", "score": 0.0024999999441206455, "subfield": {"id": "https://openalex.org/subfields/1706", "display_name": "Computer Science Applications"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}], "primary_location": {"id": "pmh:oai:arXiv.org:2508.02123", "is_oa": true, "landing_page_url": "http://arxiv.org/abs/2508.02123", "pdf_url": "https://arxiv.org/pdf/2508.02123", "source": {"id": "https://openalex.org/S4306400194", "display_name": "arXiv (Cornell University)", "issn_l": "2331-8422", "issn": ["2331-8422"], "is_oa": true, "is_in_doaj": false, "is_core": false, "listed_in": [], "host_organization": "https://openalex.org/I205783295", "host_organization_name": "Cornell University", "host_organization_lineage": ["https://openalex.org/I205783295"], "host_organization_lineage_names": [], "type": "repository"}, "license": null, "license_id": null, "version": "submittedVersion", "is_accepted": false, "is_published": false, "raw_source_name": null, "raw_type": "text"}, "cited_by_count": 0, "is_retracted": false}, {"id": "https://openalex.org/W7115590400", "doi": "https://doi.org/10.48550/arxiv.2512.11718", "title": "Speculative Decoding Speed-of-Light: Optimal Lower Bounds via Branching Random Walks", "publication_year": 2025, "publication_date": "2025-12-12", "type": "preprint", "language": null, "authorships": [{"author_position": "first", "author": {"id": null, "display_name": "Pankratov, Sergey", "orcid": null, "observed_orcids": []}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Pankratov, Sergey", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "last", "author": {"id": null, "display_name": "Alistarh, Dan", "orcid": null, "observed_orcids": []}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Alistarh, Dan", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}], "abstract_inverted_index": {"Speculative": [0], "generation": [1, 53, 65], "has": [2], "emerged": [3], "as": [4, 100], "a": [5, 60], "promising": [6], "technique": [7], "to": [8, 19, 74], "accelerate": [9], "inference": [10], "in": [11, 168], "large": [12], "language": [13], "models": [14, 157], "(LLMs)": [15], "by": [16, 58], "leveraging": [17], "parallelism": [18], "verify": [20], "multiple": [21], "draft": [22, 78], "tokens": [23, 92], "simultaneously.": [24], "However,": [25], "the": [26, 30, 41, 47, 63, 76, 88, 111, 116, 120, 127, 137, 146, 163], "fundamental": [27], "limits": [28, 138], "on": [29, 46, 155], "achievable": [31], "speedup": [32], "remain": [33], "poorly": [34], "understood.": [35], "In": [36], "this": [37], "work,": [38], "we": [39], "establish": [40], "first": [42], "``tight''": [43], "lower": [44], "bounds": [45, 167], "runtime": [48], "of": [49, 91, 119, 139, 148, 165], "any": [50], "deterministic": [51], "speculative": [52, 96, 150], "algorithm.": [54], "This": [55, 131], "is": [56, 98, 110, 115, 126], "achieved": [57], "drawing": [59], "parallel": [61, 140], "between": [62], "token": [64, 141], "process": [66], "and": [67, 124, 143], "branching": [68], "random": [69], "walks,": [70], "which": [71], "allows": [72], "us": [73], "analyze": [75], "optimal": [77], "tree": [79], "selection": [80], "problem.": [81], "We": [82], "prove,": [83], "under": [84], "basic": [85], "assumptions,": [86], "that": [87], "expected": [89, 117, 128], "number": [90], "successfully": [93], "predicted": [94], "per": [95], "iteration": [97], "bounded": [99], "$\\mathbb{E}[X]": [101], "\\leq": [102], "(\u03bc+": [103], "\u03bc_{(2)})\\log(P": [104], ")/\u03bc^2": [105], "+": [106], "O(1)$,": [107], "where": [108], "$P$": [109], "verifier's": [112, 121], "capacity,": [113], "$\u03bc$": [114], "entropy": [118], "output": [122], "distribution,": [123], "$\u03bc_{(2)}$": [125], "second": [129], "log-moment.": [130], "result": [132], "provides": [133], "new": [134], "insights": [135], "into": [136], "generation,": [142], "could": [144], "guide": [145], "design": [147], "future": [149], "decoding": [151], "systems.": [152], "Empirical": [153], "evaluations": [154], "Llama": [156], "validate": [158], "our": [159, 166], "theoretical": [160], "predictions,": [161], "confirming": [162], "tightness": [164], "practical": [169], "settings.": [170]}, "primary_topic": {"id": "https://openalex.org/T10181", "display_name": "Natural Language Processing Techniques", "score": 0.8823000192642212, "subfield": {"id": "https://openalex.org/subfields/1702", "display_name": "Artificial Intelligence"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, "topics": [{"id": "https://openalex.org/T10181", "display_name": "Natural Language Processing Techniques", "score": 0.8823000192642212, "subfield": {"id": "https://openalex.org/subfields/1702", "display_name": "Artificial Intelligence"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, {"id": "https://openalex.org/T10028", "display_name": "Topic Modeling", "score": 0.07209999859333038, "subfield": {"id": "https://openalex.org/subfields/1702", "display_name": "Artificial Intelligence"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, {"id": "https://openalex.org/T11269", "display_name": "Algorithms and Data Compression", "score": 0.008899999782443047, "subfield": {"id": "https://openalex.org/subfields/1702", "display_name": "Artificial Intelligence"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}], "primary_location": {"id": "doi:10.48550/arxiv.2512.11718", "is_oa": true, "landing_page_url": "https://doi.org/10.48550/arxiv.2512.11718", "pdf_url": null, "source": {"id": "https://openalex.org/S4306400194", "display_name": "arXiv (Cornell University)", "issn_l": "2331-8422", "issn": ["2331-8422"], "is_oa": true, "is_in_doaj": false, "is_core": false, "listed_in": [], "host_organization": "https://openalex.org/I205783295", "host_organization_name": "Cornell University", "host_organization_lineage": ["https://openalex.org/I205783295"], "host_organization_lineage_names": [], "type": "repository"}, "license": "cc-by", "license_id": "https://openalex.org/licenses/cc-by", "version": null, "is_accepted": false, "is_published": false, "raw_source_name": null, "raw_type": "Preprint"}, "cited_by_count": 0, "is_retracted": false}, {"id": "https://openalex.org/W7118822696", "doi": "https://doi.org/10.71465/fair524", "title": "Cross-Hardware Optimization Strategies for Large-Scale Recommendation Model Inference in Production Systems", "publication_year": 2025, "publication_date": "2025-12-11", "type": "article", "language": null, "authorships": [{"author_position": "first", "author": {"id": "https://openalex.org/A5121513870", "display_name": "Zijian Shen", "orcid": "https://orcid.org/0009-0001-1157-6510", "observed_orcids": ["https://orcid.org/0009-0001-1157-6510"]}, "institutions": [{"id": "https://openalex.org/I74973139", "display_name": "Carnegie Mellon University", "ror": "https://ror.org/05x2bcf33", "country_code": "US", "type": "education", "lineage": ["https://openalex.org/I74973139"]}], "countries": ["US"], "is_corresponding": false, "raw_author_name": "Zijian Shen", "raw_affiliation_strings": ["Carnegie Mellon University, United States"], "raw_orcid": null, "affiliations": [{"raw_affiliation_string": "Carnegie Mellon University, United States", "institution_ids": ["https://openalex.org/I74973139"]}]}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5121502904", "display_name": "Zimeng Wang", "orcid": null, "observed_orcids": []}, "institutions": [{"id": "https://openalex.org/I185649097", "display_name": "New England College", "ror": "https://ror.org/059aewc73", "country_code": "US", "type": "education", "lineage": ["https://openalex.org/I185649097"]}], "countries": ["US"], "is_corresponding": false, "raw_author_name": "Zimeng Wang", "raw_affiliation_strings": ["New England College, United States"], "raw_orcid": null, "affiliations": [{"raw_affiliation_string": "New England College, United States", "institution_ids": ["https://openalex.org/I185649097"]}]}, {"author_position": "last", "author": {"id": "https://openalex.org/A5103778573", "display_name": "Yang Liu", "orcid": "https://orcid.org/0009-0008-5087-8133", "observed_orcids": ["https://orcid.org/0009-0008-5087-8133"]}, "institutions": [{"id": "https://openalex.org/I107077323", "display_name": "Worcester Polytechnic Institute", "ror": "https://ror.org/05ejpqr48", "country_code": "US", "type": "education", "lineage": ["https://openalex.org/I107077323"]}], "countries": ["US"], "is_corresponding": false, "raw_author_name": "Yang Liu", "raw_affiliation_strings": ["Worcester Polytechnic Institute, United States"], "raw_orcid": null, "affiliations": [{"raw_affiliation_string": "Worcester Polytechnic Institute, United States", "institution_ids": ["https://openalex.org/I107077323"]}]}], "abstract_inverted_index": {"Large-scale": [0], "recommendation": [1, 27, 51, 95, 203], "systems": [2], "have": [3], "become": [4], "indispensable": [5], "components": [6], "of": [7, 13, 26, 161], "modern": [8], "digital": [9], "platforms,": [10], "processing": [11, 63, 67, 71], "billions": [12], "user": [14], "interactions": [15], "daily": [16], "to": [17], "deliver": [18], "personalized": [19], "content": [20], "and": [21, 74, 89, 124, 166, 182], "services.": [22], "The": [23, 97, 169], "computational": [24], "demands": [25], "model": [28, 52, 85, 108, 145, 162], "inference": [29, 177], "in": [30, 84, 131, 175, 206], "production": [31, 133, 208], "environments": [32], "present": [33], "significant": [34], "challenges,": [35], "particularly": [36], "when": [37], "deploying": [38], "across": [39, 61, 147], "heterogeneous": [40, 139, 207], "hardware": [41, 112, 149, 164, 180], "architectures.": [42], "This": [43], "review": [44, 170], "examines": [45], "cross-hardware": [46, 156], "optimization": [47, 157, 190], "strategies": [48, 123], "for": [49, 94, 110], "large-scale": [50, 202], "inference,": [53], "focusing": [54], "on": [55], "techniques": [56, 91, 194], "that": [57, 106, 128, 142, 154, 188], "enable": [58, 143], "efficient": [59, 201], "deployment": [60, 146, 184, 205], "graphics": [62], "units": [64, 68, 72], "(GPUs),": [65], "central": [66], "(CPUs),": [69], "tensor": [70], "(TPUs),": [73], "field-programmable": [75], "gate": [76], "arrays": [77], "(FPGAs).": [78], "We": [79, 118, 186], "systematically": [80], "analyze": [81], "recent": [82], "advances": [83], "compression,": [86], "including": [87], "quantization": [88], "pruning": [90], "specifically": [92], "designed": [93], "models.": [96], "paper": [98], "explores": [99], "hardware-aware": [100], "neural": [101], "architecture": [102], "search": [103], "(NAS)": [104], "methods": [105], "optimize": [107], "structures": [109], "target": [111], "platforms": [113], "while": [114], "maintaining": [115], "prediction": [116], "accuracy.": [117], "investigate": [119], "dynamic": [120], "resource": [121], "allocation": [122], "load": [125], "balancing": [126], "mechanisms": [127], "improve": [129], "throughput": [130], "multi-device": [132], "systems.": [134], "Additionally,": [135], "we": [136], "examine": [137], "emerging": [138], "computing": [140], "frameworks": [141], "seamless": [144], "diverse": [148], "infrastructures.": [150], "Our": [151], "analysis": [152], "reveals": [153], "successful": [155], "requires": [158], "careful": [159], "consideration": [160], "architecture,": [163], "characteristics,": [165], "system-level": [167], "constraints.": [168], "identifies": [171], "critical": [172], "research": [173], "gaps": [174], "real-time": [176], "optimization,": [178], "automated": [179], "selection,": [181], "energy-efficient": [183], "strategies.": [185], "conclude": [187], "integrated": [189], "approaches": [191], "combining": [192], "multiple": [193], "offer": [195], "the": [196], "most": [197], "promising": [198], "path": [199], "toward": [200], "system": [204], "environments.": [209]}, "primary_topic": {"id": "https://openalex.org/T10203", "display_name": "Recommender Systems and Techniques", "score": 0.5702999830245972, "subfield": {"id": "https://openalex.org/subfields/1710", "display_name": "Information Systems"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, "topics": [{"id": "https://openalex.org/T10203", "display_name": "Recommender Systems and Techniques", "score": 0.5702999830245972, "subfield": {"id": "https://openalex.org/subfields/1710", "display_name": "Information Systems"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, {"id": "https://openalex.org/T10036", "display_name": "Advanced Neural Network Applications", "score": 0.3756999969482422, "subfield": {"id": "https://openalex.org/subfields/1707", "display_name": "Computer Vision and Pattern Recognition"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, {"id": "https://openalex.org/T10054", "display_name": "Parallel Computing and Optimization Techniques", "score": 0.02410000003874302, "subfield": {"id": "https://openalex.org/subfields/1708", "display_name": "Hardware and Architecture"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}], "primary_location": {"id": "doi:10.71465/fair524", "is_oa": true, "landing_page_url": "https://doi.org/10.71465/fair524", "pdf_url": "https://sprcopen.org/index.php/FAIR/article/download/524/479", "source": {"id": "https://openalex.org/S5407049411", "display_name": "Frontiers in Artificial Intelligence Research", "issn_l": "3079-6342", "issn": ["3079-6342", "3079-6350"], "is_oa": false, "is_in_doaj": false, "is_core": false, "listed_in": [], "host_organization": null, "host_organization_name": null, "host_organization_lineage": [], "host_organization_lineage_names": [], "type": "journal"}, "license": "cc-by-nc-nd", "license_id": "https://openalex.org/licenses/cc-by-nc-nd", "version": "publishedVersion", "is_accepted": true, "is_published": true, "raw_source_name": "Frontiers in Artificial Intelligence Research", "raw_type": "journal-article"}, "cited_by_count": 1, "is_retracted": false}, {"id": "https://openalex.org/W4414691249", "doi": "https://doi.org/10.48550/arxiv.2507.09201", "title": "SLIM: A Heterogeneous Accelerator for Edge Inference of Sparse Large Language Model via Adaptive Thresholding", "publication_year": 2025, "publication_date": "2025-07-12", "type": "preprint", "language": "en", "authorships": [{"author_position": "first", "author": {"id": "https://openalex.org/A5039571679", "display_name": "Weihong Xu", "orcid": "https://orcid.org/0000-0003-3766-3353", "observed_orcids": ["https://orcid.org/0000-0003-3766-3353"]}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Xu, Weihong", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": null, "display_name": "Choi, Haein", "orcid": null, "observed_orcids": []}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Choi, Haein", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": null, "display_name": "Hsu, Po-kai", "orcid": null, "observed_orcids": []}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Hsu, Po-kai", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5054894631", "display_name": "Shimeng Yu", "orcid": "https://orcid.org/0000-0002-0068-3652", "observed_orcids": ["https://orcid.org/0000-0002-0068-3652", "https://orcid.org/0000-0002-2125-7062"]}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Yu, Shimeng", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "last", "author": {"id": "https://openalex.org/A5112841571", "display_name": "Tajana Rosing", "orcid": "https://orcid.org/0000-0002-6954-997X", "observed_orcids": ["https://orcid.org/0000-0002-6954-997X"]}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Rosing, Tajana", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}], "abstract_inverted_index": {"Large": [0], "language": [1], "models": [2], "(LLMs)": [3], "have": [4], "demonstrated": [5], "exceptional": [6], "proficiency": [7], "in": [8, 31, 59, 122, 137], "understanding": [9], "and": [10, 28, 35, 115, 126, 148, 164], "generating": [11], "human": [12], "language,": [13], "but": [14], "efficient": [15], "inference": [16, 45], "on": [17, 77], "resource-constrained": [18], "embedded": [19], "devices": [20], "remains": [21], "challenging": [22], "due": [23], "to": [24, 46, 53, 101], "large": [25], "model": [26], "sizes": [27], "memory-intensive": [29, 132], "operations": [30, 134], "feedforward": [32], "network": [33], "(FFN)": [34], "multi-head": [36], "attention": [37], "(MHA)": [38], "layers.": [39], "While": [40], "existing": [41], "accelerators": [42], "offload": [43], "LLM": [44, 60, 75, 82, 178], "expensive": [47], "heterogeneous": [48, 107], "computing": [49, 183], "systems,": [50], "they": [51], "fail": [52], "exploit": [54], "the": [55], "significant": [56], "sparsity": [57, 83, 92], "inherent": [58], "operations,": [61], "leaving": [62], "hardware": [63, 108], "resources": [64], "underutilized.": [65], "We": [66], "propose": [67], "SLIM,": [68], "an": [69, 85], "algorithm-hardware": [70], "co-design": [71], "optimized": [72], "for": [73, 181], "sparse": [74], "serving": [76], "edge": [78, 182], "devices.": [79], "SLIM": [80], "exploits": [81], "through": [84], "adaptive": [86], "thresholding": [87], "algorithm": [88], "that": [89], "enables": [90], "runtime-configurable": [91], "with": [93], "negligible": [94], "accuracy": [95], "loss,": [96], "fetching": [97], "only": [98], "activated": [99], "neurons": [100], "dramatically": [102], "reduce": [103], "data": [104, 146], "movement.": [105], "Our": [106, 151], "architecture": [109], "strategically": [110], "combines": [111], "near-storage": [112], "processing": [113], "(NSP)": [114], "processing-in-memory": [116], "(PIM):": [117], "FFN": [118], "weights": [119], "are": [120, 135], "stored": [121], "high-density": [123], "3D": [124], "NAND": [125], "computed": [127], "using": [128], "NSP": [129], "units,": [130], "while": [131, 172], "MHA": [133], "processed": [136], "PIM": [138], "modules.": [139], "This": [140], "design": [141], "significantly": [142], "reduces": [143], "memory": [144], "footprint,": [145], "movement,": [147], "energy": [149, 167], "consumption.": [150], "comprehensive": [152], "evaluation": [153], "demonstrates": [154], "SLIM's": [155], "effectiveness,": [156], "achieving": [157], "13-18x": [158], "throughput": [159], "improvements": [160], "over": [161, 169], "SSD-GPU": [162], "systems": [163, 171], "9-10x": [165], "better": [166], "efficiency": [168], "DRAM-GPU": [170], "maintaining": [173], "low": [174], "latency,": [175], "making": [176], "cost-effective": [177], "deployment": [179], "viable": [180], "environments.": [184]}, "primary_topic": {"id": "https://openalex.org/T10054", "display_name": "Parallel Computing and Optimization Techniques", "score": 0.5985999703407288, "subfield": {"id": "https://openalex.org/subfields/1708", "display_name": "Hardware and Architecture"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, "topics": [{"id": "https://openalex.org/T10054", "display_name": "Parallel Computing and Optimization Techniques", "score": 0.5985999703407288, "subfield": {"id": "https://openalex.org/subfields/1708", "display_name": "Hardware and Architecture"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, {"id": "https://openalex.org/T10036", "display_name": "Advanced Neural Network Applications", "score": 0.17020000517368317, "subfield": {"id": "https://openalex.org/subfields/1707", "display_name": "Computer Vision and Pattern Recognition"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, {"id": "https://openalex.org/T10502", "display_name": "Advanced Memory and Neural Computing", "score": 0.05350000038743019, "subfield": {"id": "https://openalex.org/subfields/2208", "display_name": "Electrical and Electronic Engineering"}, "field": {"id": "https://openalex.org/fields/22", "display_name": "Engineering"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}], "primary_location": {"id": "pmh:oai:arXiv.org:2507.09201", "is_oa": true, "landing_page_url": "http://arxiv.org/abs/2507.09201", "pdf_url": "https://arxiv.org/pdf/2507.09201", "source": {"id": "https://openalex.org/S4306400194", "display_name": "arXiv (Cornell University)", "issn_l": "2331-8422", "issn": ["2331-8422"], "is_oa": true, "is_in_doaj": false, "is_core": false, "listed_in": [], "host_organization": "https://openalex.org/I205783295", "host_organization_name": "Cornell University", "host_organization_lineage": ["https://openalex.org/I205783295"], "host_organization_lineage_names": [], "type": "repository"}, "license": null, "license_id": null, "version": "submittedVersion", "is_accepted": false, "is_published": false, "raw_source_name": null, "raw_type": "text"}, "cited_by_count": 0, "is_retracted": false}, {"id": "https://openalex.org/W4416371376", "doi": "https://doi.org/10.48550/arxiv.2503.10714", "title": "ZSMerge: Zero-Shot KV Cache Compression for Memory-Efficient Long-Context LLMs", "publication_year": 2025, "publication_date": "2025-03-13", "type": "preprint", "language": "en", "authorships": [{"author_position": "first", "author": {"id": "https://openalex.org/A5090522322", "display_name": "Xin Liu", "orcid": "https://orcid.org/0000-0002-0270-375X", "observed_orcids": ["https://orcid.org/0000-0002-0270-375X"]}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Liu, Xin", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5054245364", "display_name": "Xudong Wang", "orcid": "https://orcid.org/0000-0002-6575-5127", "observed_orcids": ["https://orcid.org/0000-0002-6575-5127"]}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Wang, Xudong", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5029601647", "display_name": "Pei Liu", "orcid": "https://orcid.org/0000-0002-6854-4333", "observed_orcids": ["https://orcid.org/0000-0002-6854-4333"]}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Liu, Pei", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "last", "author": {"id": "https://openalex.org/A5067691107", "display_name": "Guoming Tang", "orcid": "https://orcid.org/0000-0001-9801-1055", "observed_orcids": ["https://orcid.org/0000-0001-9801-1055"]}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Tang, Guoming", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}], "abstract_inverted_index": {"The": [0, 166], "linear": [1], "growth": [2], "of": [3, 146], "key-value": [4, 138], "(KV)": [5], "cache": [6, 29, 61, 67, 139], "memory": [7, 75, 116, 142], "and": [8, 99, 118], "quadratic": [9], "computational": [10], "in": [11, 23], "attention": [12, 97], "mechanisms": [13], "complexity": [14], "pose": [15], "significant": [16], "bottlenecks": [17], "for": [18, 65, 137], "large": [19], "language": [20], "models": [21], "(LLMs)": [22], "long-context": [24], "processing.": [25], "While": [26], "existing": [27], "KV": [28, 60], "optimization": [30], "methods": [31], "address": [32], "these": [33], "challenges": [34], "through": [35, 95], "token": [36, 80], "pruning": [37], "or": [38, 47], "feature": [39], "merging,": [40], "they": [41], "often": [42], "incur": [43], "irreversible": [44], "information": [45], "loss": [46], "require": [48], "costly": [49], "parameter": [50], "retraining.": [51, 112], "To": [52], "this": [53], "end,": [54], "we": [55], "propose": [56], "ZSMerge,": [57], "a": [58, 87, 101, 133], "dynamic": [59], "compression": [62, 135], "framework": [63], "designed": [64], "efficient": [66], "management,": [68], "featuring": [69], "three": [70], "key": [71], "operations:": [72], "(1)": [73], "fine-grained": [74], "allocation": [76], "guided": [77], "by": [78], "multi-dimensional": [79], "importance": [81], "metrics": [82], "at": [83, 158, 170], "head-level": [84], "granularity,": [85], "(2)": [86], "residual": [88], "merging": [89], "mechanism": [90, 104], "that": [91, 162], "preserves": [92], "critical": [93], "context": [94], "compensated": [96], "scoring,": [98], "(3)": [100], "zero-shot": [102], "adaptation": [103], "compatible": [105], "with": [106, 121, 154], "diverse": [107], "LLM": [108], "architectures": [109], "without": [110], "requiring": [111], "ZSMerge": [113], "significantly": [114], "enhances": [115], "efficiency": [117], "inference": [119], "speed": [120], "negligible": [122], "performance": [123], "degradation": [124], "across": [125], "LLMs.": [126], "When": [127], "applied": [128], "to": [129, 144], "LLaMA2-7B,": [130], "it": [131], "demonstrates": [132], "20:1": [134], "ratio": [136], "retention": [140], "(reducing": [141], "footprint": [143], "5\\%": [145], "baseline)": [147], "while": [148], "sustaining": [149], "comparable": [150], "generation": [151], "quality,": [152], "coupled": [153], "triple": [155], "throughput": [156], "gains": [157], "extreme": [159], "54k-token": [160], "contexts": [161], "eliminate": [163], "out-of-memory": [164], "failures.": [165], "code": [167], "is": [168], "available": [169], "https://github.com/SusCom-Lab/ZSMerge.": [171]}, "primary_topic": {"id": "https://openalex.org/T10028", "display_name": "Topic Modeling", "score": 0.5971999764442444, "subfield": {"id": "https://openalex.org/subfields/1702", "display_name": "Artificial Intelligence"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, "topics": [{"id": "https://openalex.org/T10028", "display_name": "Topic Modeling", "score": 0.5971999764442444, "subfield": {"id": "https://openalex.org/subfields/1702", "display_name": "Artificial Intelligence"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, {"id": "https://openalex.org/T10181", "display_name": "Natural Language Processing Techniques", "score": 0.29760000109672546, "subfield": {"id": "https://openalex.org/subfields/1702", "display_name": "Artificial Intelligence"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, {"id": "https://openalex.org/T10036", "display_name": "Advanced Neural Network Applications", "score": 0.05090000107884407, "subfield": {"id": "https://openalex.org/subfields/1707", "display_name": "Computer Vision and Pattern Recognition"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}], "primary_location": {"id": "pmh:oai:arXiv.org:2503.10714", "is_oa": true, "landing_page_url": "http://arxiv.org/abs/2503.10714", "pdf_url": "https://arxiv.org/pdf/2503.10714", "source": {"id": "https://openalex.org/S4306400194", "display_name": "arXiv (Cornell University)", "issn_l": "2331-8422", "issn": ["2331-8422"], "is_oa": true, "is_in_doaj": false, "is_core": false, "listed_in": [], "host_organization": "https://openalex.org/I205783295", "host_organization_name": "Cornell University", "host_organization_lineage": ["https://openalex.org/I205783295"], "host_organization_lineage_names": [], "type": "repository"}, "license": null, "license_id": null, "version": "submittedVersion", "is_accepted": false, "is_published": false, "raw_source_name": null, "raw_type": "text"}, "cited_by_count": 1, "is_retracted": false}, {"id": "https://openalex.org/W7118021545", "doi": "https://doi.org/10.48550/arxiv.2512.23858", "title": "Yggdrasil: Bridging Dynamic Speculation and Static Runtime for Latency-Optimal Tree-Based LLM Decoding", "publication_year": 2025, "publication_date": "2025-12-29", "type": "preprint", "language": null, "authorships": [{"author_position": "first", "author": {"id": "https://openalex.org/A5091217469", "display_name": "Yue Guan", "orcid": "https://orcid.org/0009-0005-7433-2627", "observed_orcids": ["https://orcid.org/0009-0005-7433-2627"]}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Guan, Yue", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5121801880", "display_name": "Changming Yu", "orcid": null, "observed_orcids": []}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Yu, Changming", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5121757028", "display_name": "Shihan Fang", "orcid": null, "observed_orcids": []}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Fang, Shihan", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5121755920", "display_name": "Weiming Hu", "orcid": null, "observed_orcids": []}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Hu, Weiming", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5041947969", "display_name": "Zaifeng Pan", "orcid": "https://orcid.org/0000-0002-6759-2616", "observed_orcids": ["https://orcid.org/0000-0002-6759-2616"]}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Pan, Zaifeng", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5121749910", "display_name": "Zheng Wang", "orcid": null, "observed_orcids": []}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Wang, Zheng", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5121816858", "display_name": "Zihan Liu", "orcid": null, "observed_orcids": []}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Liu, Zihan", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5140632867", "display_name": "Yangjie Zhou", "orcid": null, "observed_orcids": []}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Zhou, Yangjie", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5121808620", "display_name": "Yufei Ding", "orcid": null, "observed_orcids": []}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Ding, Yufei", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5135688432", "display_name": "Minyi Guo", "orcid": null, "observed_orcids": []}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Guo, Minyi", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "last", "author": {"id": "https://openalex.org/A5003939279", "display_name": "Jingwen Leng", "orcid": "https://orcid.org/0000-0002-5660-5493", "observed_orcids": ["https://orcid.org/0000-0002-5660-5493"]}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Leng, Jingwen", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}], "abstract_inverted_index": {"Speculative": [0], "decoding": [1, 41], "improves": [2], "LLM": [3], "inference": [4], "by": [5], "generating": [6], "and": [7, 27, 46, 66, 76], "verifying": [8], "multiple": [9, 86], "tokens": [10], "in": [11], "parallel,": [12], "but": [13], "existing": [14], "systems": [15], "suffer": [16], "from": [17], "suboptimal": [18], "performance": [19], "due": [20], "to": [21, 69, 79], "a": [22, 34, 59], "mismatch": [23], "between": [24], "dynamic": [25], "speculation": [26], "static": [28, 56], "runtime": [29], "assumptions.": [30], "We": [31], "present": [32], "Yggdrasil,": [33], "co-designed": [35], "system": [36], "that": [37], "enables": [38], "latency-optimal": [39], "speculative": [40], "through": [42], "context-aware": [43], "tree": [44, 53], "drafting": [45], "compiler-friendly": [47], "execution.": [48], "Yggdrasil": [49, 72], "introduces": [50], "an": [51], "equal-growth": [52], "structure": [54], "for": [55, 63], "graph": [57], "compatibility,": [58], "latency-aware": [60], "optimization": [61], "objective": [62], "draft": [64], "selection,": [65], "stage-based": [67], "scheduling": [68], "reduce": [70], "overhead.": [71], "supports": [73], "unmodified": [74], "LLMs": [75], "achieves": [77], "up": [78], "$3.98\\times$": [80], "speedup": [81], "over": [82], "state-of-the-art": [83], "baselines": [84], "across": [85], "hardware": [87], "setups.": [88]}, "primary_topic": {"id": "https://openalex.org/T10054", "display_name": "Parallel Computing and Optimization Techniques", "score": 0.5371000170707703, "subfield": {"id": "https://openalex.org/subfields/1708", "display_name": "Hardware and Architecture"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, "topics": [{"id": "https://openalex.org/T10054", "display_name": "Parallel Computing and Optimization Techniques", "score": 0.5371000170707703, "subfield": {"id": "https://openalex.org/subfields/1708", "display_name": "Hardware and Architecture"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, {"id": "https://openalex.org/T10181", "display_name": "Natural Language Processing Techniques", "score": 0.32280001044273376, "subfield": {"id": "https://openalex.org/subfields/1702", "display_name": "Artificial Intelligence"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, {"id": "https://openalex.org/T10028", "display_name": "Topic Modeling", "score": 0.04740000143647194, "subfield": {"id": "https://openalex.org/subfields/1702", "display_name": "Artificial Intelligence"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}], "primary_location": {"id": "doi:10.48550/arxiv.2512.23858", "is_oa": true, "landing_page_url": "https://doi.org/10.48550/arxiv.2512.23858", "pdf_url": null, "source": {"id": "https://openalex.org/S4306400194", "display_name": "arXiv (Cornell University)", "issn_l": "2331-8422", "issn": ["2331-8422"], "is_oa": true, "is_in_doaj": false, "is_core": false, "listed_in": [], "host_organization": "https://openalex.org/I205783295", "host_organization_name": "Cornell University", "host_organization_lineage": ["https://openalex.org/I205783295"], "host_organization_lineage_names": [], "type": "repository"}, "license": null, "license_id": null, "version": null, "is_accepted": false, "is_published": false, "raw_source_name": null, "raw_type": "Preprint"}, "cited_by_count": 0, "is_retracted": false}, {"id": "https://openalex.org/W7117870265", "doi": "https://doi.org/10.1145/3787470.3787480", "title": "OmniRouter: Budget and Performance Controllable Multi-LLM Routing", "publication_year": 2025, "publication_date": "2025-12-30", "type": "article", "language": "en", "authorships": [{"author_position": "first", "author": {"id": "https://openalex.org/A5100929110", "display_name": "Kai Mei", "orcid": null, "observed_orcids": []}, "institutions": [{"id": "https://openalex.org/I102322142", "display_name": "Rutgers, The State University of New Jersey", "ror": "https://ror.org/05vt9qd57", "country_code": "US", "type": "education", "lineage": ["https://openalex.org/I102322142"]}], "countries": ["US"], "is_corresponding": false, "raw_author_name": "Kai Mei", "raw_affiliation_strings": ["Department of Computer Science, Rutgers University"], "raw_orcid": null, "affiliations": [{"raw_affiliation_string": "Department of Computer Science, Rutgers University", "institution_ids": ["https://openalex.org/I102322142"]}]}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5081309714", "display_name": "Wujiang Xu", "orcid": "https://orcid.org/0000-0002-3500-1068", "observed_orcids": ["https://orcid.org/0000-0002-3500-1068"]}, "institutions": [{"id": "https://openalex.org/I102322142", "display_name": "Rutgers, The State University of New Jersey", "ror": "https://ror.org/05vt9qd57", "country_code": "US", "type": "education", "lineage": ["https://openalex.org/I102322142"]}], "countries": ["US"], "is_corresponding": false, "raw_author_name": "Wujiang Xu", "raw_affiliation_strings": ["Department of Computer Science, Rutgers University"], "raw_orcid": null, "affiliations": [{"raw_affiliation_string": "Department of Computer Science, Rutgers University", "institution_ids": ["https://openalex.org/I102322142"]}]}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5121719133", "display_name": "Minghao Guo", "orcid": null, "observed_orcids": []}, "institutions": [{"id": "https://openalex.org/I102322142", "display_name": "Rutgers, The State University of New Jersey", "ror": "https://ror.org/05vt9qd57", "country_code": "US", "type": "education", "lineage": ["https://openalex.org/I102322142"]}], "countries": ["US"], "is_corresponding": false, "raw_author_name": "Minghao Guo", "raw_affiliation_strings": ["Department of Computer Science, Rutgers University"], "raw_orcid": null, "affiliations": [{"raw_affiliation_string": "Department of Computer Science, Rutgers University", "institution_ids": ["https://openalex.org/I102322142"]}]}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5103921409", "display_name": "Shuhang Lin", "orcid": null, "observed_orcids": []}, "institutions": [{"id": "https://openalex.org/I102322142", "display_name": "Rutgers, The State University of New Jersey", "ror": "https://ror.org/05vt9qd57", "country_code": "US", "type": "education", "lineage": ["https://openalex.org/I102322142"]}], "countries": ["US"], "is_corresponding": false, "raw_author_name": "Shuhang Lin", "raw_affiliation_strings": ["Department of Computer Science, Rutgers University"], "raw_orcid": null, "affiliations": [{"raw_affiliation_string": "Department of Computer Science, Rutgers University", "institution_ids": ["https://openalex.org/I102322142"]}]}, {"author_position": "last", "author": {"id": "https://openalex.org/A5066709505", "display_name": "Y. Zhang", "orcid": null, "observed_orcids": []}, "institutions": [{"id": "https://openalex.org/I102322142", "display_name": "Rutgers, The State University of New Jersey", "ror": "https://ror.org/05vt9qd57", "country_code": "US", "type": "education", "lineage": ["https://openalex.org/I102322142"]}], "countries": ["US"], "is_corresponding": false, "raw_author_name": "Yongfeng Zhang", "raw_affiliation_strings": ["Department of Computer Science, Rutgers University"], "raw_orcid": null, "affiliations": [{"raw_affiliation_string": "Department of Computer Science, Rutgers University", "institution_ids": ["https://openalex.org/I102322142"]}]}], "abstract_inverted_index": {"Large": [0], "language": [1, 41], "models": [2, 20, 42, 114, 124], "(LLMs)": [3], "deliver": [4], "superior": [5], "performance": [6, 133], "but": [7], "require": [8], "substantial": [9], "computational": [10, 210], "resources": [11], "and": [12, 146, 155, 223], "operate": [13], "with": [14, 26, 170], "relatively": [15], "low": [16], "efficiency,": [17], "while": [18, 56, 129, 189, 207], "smaller": [19], "can": [21], "efficiently": [22], "handle": [23], "simpler": [24], "tasks": [25], "fewer": [27], "resources.": [28], "LLMrouting": [29], "is": [30, 140], "a": [31, 44, 67, 98, 119, 136, 159], "crucial": [32], "paradigm": [33], "that": [34, 125, 165, 197], "dynamically": [35, 182], "selects": [36], "the": [37, 73, 115, 131, 144, 152, 177, 224], "most": [38], "suitable": [39], "large": [40], "from": [43], "pool": [45], "of": [46, 108, 148], "candidates": [47], "to": [48, 142, 191, 201, 217], "process": [49], "diverse": [50], "inputs,": [51], "ensuring": [52, 130], "optimal": [53, 69, 179], "resource": [54, 89], "utilization": [55], "maintaining": [57], "response": [58, 205], "quality.": [59], "Existing": [60], "routing": [61, 101, 116], "frameworks": [62], "typically": [63], "model": [64], "this": [65, 93], "as": [66, 118], "locally": [68], "decision-making": [70], "problem,": [71, 94, 122], "selecting": [72], "presumed": [74], "best-fit": [75], "LLM": [76, 105], "for": [77, 103, 162], "each": [78], "query": [79], "individually,": [80], "which": [81], "overlooks": [82], "global": [83], "budget": [84], "constraints,": [85], "resulting": [86], "in": [87, 204], "ineffective": [88], "allocation.": [90], "To": [91], "tackle": [92], "we": [95, 157], "introduce": [96], "OmniRouter,": [97], "fundamentally": [99], "controllable": [100], "framework": [102], "multi-": [104], "serving.": [106], "Instead": [107], "making": [109], "per-query": [110], "greedy": [111], "choices,": [112], "OmniRouter": [113, 198], "task": [117], "constrained": [120, 160], "optimization": [121], "assigning": [123], "minimize": [126], "total": [127], "cost": [128, 154], "required": [132], "level.": [134], "Specifically,": [135], "hybrid": [137], "retrieval-augmented": [138], "predictor": [139], "designed": [141], "predict": [143], "capabilities": [145], "costs": [147, 211], "LLMs.": [149], "After": [150], "obtaining": [151], "predicted": [153], "performance,": [156], "utilize": [158], "optimizer": [161], "cost-optimal": [163], "assignments": [164], "employs": [166], "Lagrangian": [167], "dual": [168], "decomposition": [169], "adaptive": [171], "multipliers.": [172], "It": [173], "iteratively": [174], "converges": [175], "toward": [176], "globally": [178], "query-model": [180], "allocation,": [181], "balancing": [183], "latency": [184], "minimization": [185], "against": [186], "quality": [187], "thresholds": [188], "adhering": [190], "heterogeneous": [192], "capacity": [193], "constraints.": [194], "Experiments": [195], "show": [196], "achieves": [199], "up": [200], "6.30%": [202], "improvement": [203], "accuracy": [206], "simultaneously": [208], "reducing": [209], "by": [212], "at": [213, 228], "least": [214], "10.15%": [215], "compared": [216], "competitive": [218], "router": [219], "baselines.": [220], "The": [221], "code": [222], "dataset": [225], "are": [226], "available": [227], "https:": [229], "//github.com/dongyuanjushi/OmniRouter.": [230]}, "primary_topic": {"id": "https://openalex.org/T10028", "display_name": "Topic Modeling", "score": 0.4569999873638153, "subfield": {"id": "https://openalex.org/subfields/1702", "display_name": "Artificial Intelligence"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, "topics": [{"id": "https://openalex.org/T10028", "display_name": "Topic Modeling", "score": 0.4569999873638153, "subfield": {"id": "https://openalex.org/subfields/1702", "display_name": "Artificial Intelligence"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, {"id": "https://openalex.org/T10181", "display_name": "Natural Language Processing Techniques", "score": 0.13099999725818634, "subfield": {"id": "https://openalex.org/subfields/1702", "display_name": "Artificial Intelligence"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, {"id": "https://openalex.org/T10101", "display_name": "Cloud Computing and Resource Management", "score": 0.11990000307559967, "subfield": {"id": "https://openalex.org/subfields/1710", "display_name": "Information Systems"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}], "primary_location": {"id": "doi:10.1145/3787470.3787480", "is_oa": false, "landing_page_url": "https://doi.org/10.1145/3787470.3787480", "pdf_url": null, "source": {"id": "https://openalex.org/S4210176598", "display_name": "ACM SIGKDD Explorations Newsletter", "issn_l": "1931-0145", "issn": ["1931-0145", "1931-0153"], "is_oa": false, "is_in_doaj": false, "is_core": true, "listed_in": ["cwts-core", "norway-1"], "host_organization": "https://openalex.org/P4310319798", "host_organization_name": "Association for Computing Machinery", "host_organization_lineage": ["https://openalex.org/P4310319798"], "host_organization_lineage_names": ["Association for Computing Machinery"], "type": "journal"}, "license": null, "license_id": null, "version": "publishedVersion", "is_accepted": true, "is_published": true, "raw_source_name": "ACM SIGKDD Explorations Newsletter", "raw_type": "journal-article"}, "cited_by_count": 2, "is_retracted": false}, {"id": "https://openalex.org/W4411309330", "doi": "https://doi.org/10.1145/3744244", "title": "HAPE: Hardware-Aware LLM Pruning For Efficient On-Device Inference Optimization", "publication_year": 2025, "publication_date": "2025-06-14", "type": "article", "language": "en", "authorships": [{"author_position": "first", "author": {"id": "https://openalex.org/A5013503446", "display_name": "Wenqian Zhao", "orcid": "https://orcid.org/0000-0001-9501-9254", "observed_orcids": ["https://orcid.org/0000-0001-9501-9254", "https://orcid.org/0009-0006-9542-5299"]}, "institutions": [{"id": "https://openalex.org/I177725633", "display_name": "Chinese University of Hong Kong", "ror": "https://ror.org/00t33hh48", "country_code": "HK", "type": "education", "lineage": ["https://openalex.org/I177725633"]}], "countries": ["HK"], "is_corresponding": false, "raw_author_name": "Wenqian Zhao", "raw_affiliation_strings": ["The Chinese University of Hong Kong", "The Chinese University of Hong Kong,  Hong Kong, Hong Kong"], "raw_orcid": "https://orcid.org/0000-0001-9501-9254", "affiliations": [{"raw_affiliation_string": "The Chinese University of Hong Kong", "institution_ids": ["https://openalex.org/I177725633"]}, {"raw_affiliation_string": "The Chinese University of Hong Kong,  Hong Kong, Hong Kong", "institution_ids": ["https://openalex.org/I177725633"]}]}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5092876716", "display_name": "Lancheng Zou", "orcid": "https://orcid.org/0009-0004-6820-7064", "observed_orcids": ["https://orcid.org/0009-0004-6820-7064"]}, "institutions": [{"id": "https://openalex.org/I177725633", "display_name": "Chinese University of Hong Kong", "ror": "https://ror.org/00t33hh48", "country_code": "HK", "type": "education", "lineage": ["https://openalex.org/I177725633"]}], "countries": ["HK"], "is_corresponding": false, "raw_author_name": "Lancheng Zou", "raw_affiliation_strings": ["The Chinese University of Hong Kong", "The Chinese University of Hong Kong,  Hong Kong Hong Kong"], "raw_orcid": "https://orcid.org/0009-0004-6820-7064", "affiliations": [{"raw_affiliation_string": "The Chinese University of Hong Kong", "institution_ids": ["https://openalex.org/I177725633"]}, {"raw_affiliation_string": "The Chinese University of Hong Kong,  Hong Kong Hong Kong", "institution_ids": ["https://openalex.org/I177725633"]}]}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5032567340", "display_name": "Zixiao Wang", "orcid": "https://orcid.org/0009-0000-8179-0996", "observed_orcids": ["https://orcid.org/0009-0000-8179-0996"]}, "institutions": [{"id": "https://openalex.org/I177725633", "display_name": "Chinese University of Hong Kong", "ror": "https://ror.org/00t33hh48", "country_code": "HK", "type": "education", "lineage": ["https://openalex.org/I177725633"]}], "countries": ["HK"], "is_corresponding": false, "raw_author_name": "Zixiao Wang", "raw_affiliation_strings": ["Computer Science and Engineering, The Chinese University of Hong Kong", "Computer Science and Engineering, The Chinese University of Hong Kong, Hong Kong, Hong Kong"], "raw_orcid": "https://orcid.org/0009-0000-8179-0996", "affiliations": [{"raw_affiliation_string": "Computer Science and Engineering, The Chinese University of Hong Kong", "institution_ids": ["https://openalex.org/I177725633"]}, {"raw_affiliation_string": "Computer Science and Engineering, The Chinese University of Hong Kong, Hong Kong, Hong Kong", "institution_ids": ["https://openalex.org/I177725633"]}]}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5033561795", "display_name": "Xufeng Yao", "orcid": "https://orcid.org/0000-0002-7994-6290", "observed_orcids": ["https://orcid.org/0000-0002-7994-6290"]}, "institutions": [{"id": "https://openalex.org/I177725633", "display_name": "Chinese University of Hong Kong", "ror": "https://ror.org/00t33hh48", "country_code": "HK", "type": "education", "lineage": ["https://openalex.org/I177725633"]}], "countries": ["HK"], "is_corresponding": false, "raw_author_name": "Xufeng Yao", "raw_affiliation_strings": ["Computer Science and Engineering, CUHK", "Computer Science and Engineering, CUHK, Hong Kong Hong Kong"], "raw_orcid": "https://orcid.org/0000-0002-7994-6290", "affiliations": [{"raw_affiliation_string": "Computer Science and Engineering, CUHK", "institution_ids": ["https://openalex.org/I177725633"]}, {"raw_affiliation_string": "Computer Science and Engineering, CUHK, Hong Kong Hong Kong", "institution_ids": ["https://openalex.org/I177725633"]}]}, {"author_position": "last", "author": {"id": "https://openalex.org/A5051340429", "display_name": "Bei Yu", "orcid": "https://orcid.org/0000-0001-6406-4810", "observed_orcids": ["https://orcid.org/0000-0001-6406-4810", "https://orcid.org/0000-0002-1557-7166"]}, "institutions": [{"id": "https://openalex.org/I177725633", "display_name": "Chinese University of Hong Kong", "ror": "https://ror.org/00t33hh48", "country_code": "HK", "type": "education", "lineage": ["https://openalex.org/I177725633"]}], "countries": ["HK"], "is_corresponding": false, "raw_author_name": "Bei Yu", "raw_affiliation_strings": ["The Chinese University of Hong Kong Department of Computer Science and Engineering", "The Chinese University of Hong Kong Department of Computer Science and Engineering,  Hong Kong Hong Kong"], "raw_orcid": "https://orcid.org/0000-0001-6406-4810", "affiliations": [{"raw_affiliation_string": "The Chinese University of Hong Kong Department of Computer Science and Engineering", "institution_ids": ["https://openalex.org/I177725633"]}, {"raw_affiliation_string": "The Chinese University of Hong Kong Department of Computer Science and Engineering,  Hong Kong Hong Kong", "institution_ids": ["https://openalex.org/I177725633"]}]}], "abstract_inverted_index": {"Over": [0], "the": [1, 42, 59, 65, 141, 196], "past": [2], "few": [3], "years,": [4], "large": [5, 57], "language": [6], "models": [7], "(LLMs)": [8], "have": [9, 76, 95], "demonstrated": [10], "remarkable": [11], "performance": [12], "and": [13, 32, 74, 92, 109, 145, 185, 201, 211], "versatility": [14, 91], "across": [15], "a": [16, 37, 97, 124, 192], "variety": [17], "of": [18, 62, 67, 80, 90, 111, 114], "complex": [19], "tasks.": [20], "However,": [21], "their": [22, 28], "deployment": [23, 66, 108], "has": [24], "been": [25], "challenged": [26], "by": [27, 163], "substantial": [29], "model": [30, 43], "size": [31], "computational": [33], "requirements.": [34], "Pruning": [35], "is": [36, 171], "effective": [38], "approach": [39], "to": [40, 84, 105, 122, 153], "make": [41], "parameters": [44], "sparse,": [45], "thereby": [46, 202], "acquire": [47], "inference": [48, 110, 130, 209], "acceleration.": [49], "While": [50], "not": [51], "everyone": [52], "requires": [53], "training": [54, 200], "or": [55], "fine-tuning": [56], "models,": [58], "diverse": [60, 208], "range": [61], "applications": [63], "necessitates": [64], "LLMs": [68], "on": [69, 214], "different": [70], "devices.": [71], "Model": [72], "pruning": [73, 99, 125, 135, 143, 151], "compression": [75], "emerged": [77], "as": [78], "areas": [79], "deep": [81], "research": [82], "interest": [83], "address": [85], "these": [86], "challenges.": [87], "In": [88], "consideration": [89], "practicality,": [93], "we": [94, 119, 160], "designed": [96], "hardware-aware": [98], "process": [100, 189], "for": [101, 198], "general-purpose": [102], "hardware/edge": [103], "devices": [104], "enable": [106], "efficient": [107], "LLMs.": [112], "Instead": [113], "considering": [115], "sparse": [116], "ratio": [117], "alone,": [118], "are": [120], "motivated": [121], "design": [123], "framework": [126, 139], "that": [127], "incorporates": [128], "genuine": [129], "speed-up": [131], "sensitivity": [132], "from": [133, 158], "each": [134], "structure.": [136], "Moreover,": [137], "our": [138], "breaks": [140], "layer-by-layer": [142], "setting": [144], "fuse": [146], "several": [147], "layers": [148], "into": [149], "one": [150], "stage": [152], "allow": [154], "cross-layer": [155], "optimization.": [156, 187], "Apart": [157], "that,": [159], "hold": [161], "pragmatism": [162], "conducting": [164], "compilation": [165], "optimization": [166], "during": [167], "pruning.": [168], "This": [169], "step": [170], "critical": [172], "because": [173], "most": [174], "sparsity": [175], "patterns": [176], "barely": [177], "show": [178], "distinct": [179], "speed": [180, 210], "acceleration": [181], "with": [182], "corresponding": [183], "dataflow": [184], "memory": [186], "Our": [188], "operates": [190], "within": [191], "post-training": [193], "framework,": [194], "obviating": [195], "need": [197], "additional": [199], "reducing": [203], "resource": [204], "requirements,": [205], "while": [206], "ensuring": [207], "accuracy": [212], "requirements": [213], "hardware.": [215]}, "primary_topic": {"id": "https://openalex.org/T10036", "display_name": "Advanced Neural Network Applications", "score": 0.9955999851226807, "subfield": {"id": "https://openalex.org/subfields/1707", "display_name": "Computer Vision and Pattern Recognition"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, "topics": [{"id": "https://openalex.org/T10036", "display_name": "Advanced Neural Network Applications", "score": 0.9955999851226807, "subfield": {"id": "https://openalex.org/subfields/1707", "display_name": "Computer Vision and Pattern Recognition"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, {"id": "https://openalex.org/T10054", "display_name": "Parallel Computing and Optimization Techniques", "score": 0.0008999999845400453, "subfield": {"id": "https://openalex.org/subfields/1708", "display_name": "Hardware and Architecture"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, {"id": "https://openalex.org/T10363", "display_name": "Low-power high-performance VLSI design", "score": 0.00039999998989515007, "subfield": {"id": "https://openalex.org/subfields/2208", "display_name": "Electrical and Electronic Engineering"}, "field": {"id": "https://openalex.org/fields/22", "display_name": "Engineering"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}], "primary_location": {"id": "doi:10.1145/3744244", "is_oa": true, "landing_page_url": "https://doi.org/10.1145/3744244", "pdf_url": "https://dl.acm.org/doi/pdf/10.1145/3744244", "source": {"id": "https://openalex.org/S105046310", "display_name": "ACM Transactions on Design Automation of Electronic Systems", "issn_l": "1084-4309", "issn": ["1084-4309", "1557-7309"], "is_oa": false, "is_in_doaj": false, "is_core": true, "listed_in": ["cwts-core", "jufo-1", "norway-1"], "host_organization": "https://openalex.org/P4310319798", "host_organization_name": "Association for Computing Machinery", "host_organization_lineage": ["https://openalex.org/P4310319798"], "host_organization_lineage_names": ["Association for Computing Machinery"], "type": "journal"}, "license": "cc-by", "license_id": "https://openalex.org/licenses/cc-by", "version": "publishedVersion", "is_accepted": true, "is_published": true, "raw_source_name": "ACM Transactions on Design Automation of Electronic Systems", "raw_type": "journal-article"}, "cited_by_count": 8, "is_retracted": false}, {"id": "https://openalex.org/W7108811324", "doi": "https://doi.org/10.5281/zenodo.17815797", "title": "Semantic-Aware Caching for Memory-Efficient Large Language Model Inference", "publication_year": 2025, "publication_date": "2025-12-04", "type": "article", "language": null, "authorships": [{"author_position": "first", "author": {"id": null, "display_name": "Revista, Zen", "orcid": "https://orcid.org/0009-0007-6299-2008", "observed_orcids": ["https://orcid.org/0009-0007-6299-2008"]}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Revista, Zen", "raw_affiliation_strings": [], "raw_orcid": "https://orcid.org/0009-0007-6299-2008", "affiliations": []}, {"author_position": "last", "author": {"id": null, "display_name": "IA, 10", "orcid": "https://orcid.org/0009-0007-6299-2008", "observed_orcids": ["https://orcid.org/0009-0007-6299-2008"]}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "IA, 10", "raw_affiliation_strings": [], "raw_orcid": "https://orcid.org/0009-0007-6299-2008", "affiliations": []}], "abstract_inverted_index": {"Large": [0], "Language": [1], "Models": [2], "(LLMs)": [3], "have": [4], "demonstrated": [5], "remarkable": [6], "capabilities": [7], "across": [8], "various": [9], "natural": [10], "language": [11], "processing": [12], "tasks,": [13], "yet": [14], "their": [15, 117, 141], "deployment": [16, 227], "in": [17, 143, 231], "real-world": [18], "applications": [19, 236], "is": [20, 39], "often": [21], "constrained": [22], "by": [23], "significant": [24], "memory": [25, 37, 91, 193], "requirements,": [26], "particularly": [27], "during": [28, 57], "the": [29, 40, 64, 80, 110, 146, 150, 216], "inference": [30, 71, 203], "phase.": [31], "A": [32], "major": [33], "contributor": [34], "to": [35, 53, 89, 123, 191, 221], "this": [36], "bottleneck": [38], "Key-Value": [41], "(KV)": [42], "cache,": [43], "which": [44], "stores": [45], "intermediate": [46], "attention": [47], "states": [48], "for": [49, 93, 159, 164, 235], "previously": [50], "processed": [51], "tokens": [52], "avoid": [54], "redundant": [55], "computations": [56], "autoregressive": [58], "decoding.": [59], "As": [60], "context": [61, 239], "windows": [62], "expand,": [63], "KV": [65, 113, 129], "cache": [66, 130], "grows": [67], "linearly,": [68], "making": [69], "long-context": [70], "a": [72, 86, 157, 208], "memory-bound": [73], "problem.": [74], "This": [75], "paper": [76, 206], "proposes": [77], "and": [78, 121, 137, 174, 202, 211, 225], "explores": [79], "concept": [81], "of": [82, 112, 149, 218, 228], "semantic-aware": [83, 107, 160, 219], "caching": [84, 98, 108, 220], "as": [85], "novel": [87], "approach": [88, 188], "improve": [90], "efficiency": [92], "LLM": [94], "inference.": [95], "Unlike": [96], "traditional": [97], "strategies": [99], "that": [100], "rely": [101], "on": [102, 116, 180], "recency": [103], "or": [104, 197, 234], "fixed-size": [105], "policies,": [106], "prioritizes": [109], "retention": [111], "pairs": [114], "based": [115, 179], "perceived": [118], "semantic": [119, 147, 181], "importance": [120], "relevance": [122], "future": [124], "generations.": [125], "We": [126, 183], "review": [127], "existing": [128], "optimization": [131], "techniques,": [132], "including": [133], "quantization,": [134], "eviction": [135], "strategies,": [136], "architectural": [138], "innovations,": [139], "highlighting": [140], "limitations": [142], "fully": [144], "leveraging": [145], "content": [148], "cached": [151], "information.": [152], "Our": [153], "proposed": [154], "methodology": [155], "outlines": [156], "framework": [158, 210], "caching,": [161], "detailing": [162], "mechanisms": [163], "dynamically": [165], "scoring": [166], "token": [167], "importance,": [168], "adaptively": [169], "evicting": [170], "less": [171], "critical": [172], "tokens,": [173], "potentially": [175], "applying": [176], "heterogeneous": [177], "compression": [178], "value.": [182], "discuss": [184], "how": [185], "such": [186], "an": [187], "can": [189], "lead": [190], "substantial": [192], "reductions": [194], "while": [195], "preserving": [196], "even": [198], "enhancing": [199], "model": [200], "quality": [201], "throughput.": [204], "The": [205], "presents": [207], "conceptual": [209], "discusses": [212], "expected": [213], "results,": [214], "demonstrating": [215], "potential": [217], "enable": [222], "more": [223], "efficient": [224], "scalable": [226], "LLMs,": [229], "especially": [230], "resource-constrained": [232], "environments": [233], "requiring": [237], "extensive": [238], "understanding.": [240]}, "primary_topic": {"id": "https://openalex.org/T10028", "display_name": "Topic Modeling", "score": 0.5503000020980835, "subfield": {"id": "https://openalex.org/subfields/1702", "display_name": "Artificial Intelligence"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, "topics": [{"id": "https://openalex.org/T10028", "display_name": "Topic Modeling", "score": 0.5503000020980835, "subfield": {"id": "https://openalex.org/subfields/1702", "display_name": "Artificial Intelligence"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, {"id": "https://openalex.org/T10181", "display_name": "Natural Language Processing Techniques", "score": 0.2671000063419342, "subfield": {"id": "https://openalex.org/subfields/1702", "display_name": "Artificial Intelligence"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, {"id": "https://openalex.org/T10036", "display_name": "Advanced Neural Network Applications", "score": 0.07739999890327454, "subfield": {"id": "https://openalex.org/subfields/1707", "display_name": "Computer Vision and Pattern Recognition"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}], "primary_location": {"id": "doi:10.5281/zenodo.17815797", "is_oa": true, "landing_page_url": "https://doi.org/10.5281/zenodo.17815797", "pdf_url": null, "source": {"id": "https://openalex.org/S4306400562", "display_name": "Zenodo (CERN European Organization for Nuclear Research)", "issn_l": null, "issn": null, "is_oa": true, "is_in_doaj": false, "is_core": false, "listed_in": [], "host_organization": "https://openalex.org/I67311998", "host_organization_name": "European Organization for Nuclear Research", "host_organization_lineage": ["https://openalex.org/I67311998"], "host_organization_lineage_names": [], "type": "repository"}, "license": "cc-by", "license_id": "https://openalex.org/licenses/cc-by", "version": null, "is_accepted": false, "is_published": false, "raw_source_name": null, "raw_type": "JournalArticle"}, "cited_by_count": 0, "is_retracted": false}, {"id": "https://openalex.org/W4412828027", "doi": "https://doi.org/10.1109/tpds.2025.3590014", "title": "Cannikin: No Lagger of SLO in Concurrent Multiple LoRA LLM Serving", "publication_year": 2025, "publication_date": "2025-07-17", "type": "article", "language": "en", "authorships": [{"author_position": "first", "author": {"id": null, "display_name": "Ruidong Zhu", "orcid": "https://orcid.org/0000-0003-1389-7884", "observed_orcids": ["https://orcid.org/0000-0003-1389-7884"]}, "institutions": [{"id": "https://openalex.org/I20231570", "display_name": "Peking University", "ror": "https://ror.org/02v51f717", "country_code": "CN", "type": "education", "lineage": ["https://openalex.org/I20231570"]}], "countries": ["CN"], "is_corresponding": false, "raw_author_name": "Ruidong Zhu", "raw_affiliation_strings": ["School of Computer Science, Peking University, Beijing, China"], "raw_orcid": "https://orcid.org/0000-0003-1389-7884", "affiliations": [{"raw_affiliation_string": "School of Computer Science, Peking University, Beijing, China", "institution_ids": ["https://openalex.org/I20231570"]}]}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5079428008", "display_name": "Ziyue Jiang", "orcid": "https://orcid.org/0009-0001-4521-2790", "observed_orcids": ["https://orcid.org/0009-0001-4521-2790"]}, "institutions": [{"id": "https://openalex.org/I4405256866", "display_name": "ByteDance (China)", "ror": "https://ror.org/04m8av638", "country_code": "CN", "type": "company", "lineage": ["https://openalex.org/I4405256866"]}], "countries": ["CN"], "is_corresponding": false, "raw_author_name": "Ziyue Jiang", "raw_affiliation_strings": ["ByteDance, Shanghai, China"], "raw_orcid": "https://orcid.org/0009-0001-4521-2790", "affiliations": [{"raw_affiliation_string": "ByteDance, Shanghai, China", "institution_ids": ["https://openalex.org/I4405256866"]}]}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5081124136", "display_name": "Zhi Zhang", "orcid": "https://orcid.org/0000-0003-3830-8652", "observed_orcids": ["https://orcid.org/0000-0003-3830-8652", "https://orcid.org/0000-0001-7794-1763"]}, "institutions": [{"id": "https://openalex.org/I4405256866", "display_name": "ByteDance (China)", "ror": "https://ror.org/04m8av638", "country_code": "CN", "type": "company", "lineage": ["https://openalex.org/I4405256866"]}], "countries": ["CN"], "is_corresponding": false, "raw_author_name": "Zhi Zhang", "raw_affiliation_strings": ["ByteDance, Shanghai, China"], "raw_orcid": null, "affiliations": [{"raw_affiliation_string": "ByteDance, Shanghai, China", "institution_ids": ["https://openalex.org/I4405256866"]}]}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5037891132", "display_name": "Xin Liu", "orcid": "https://orcid.org/0009-0001-5756-0004", "observed_orcids": ["https://orcid.org/0009-0001-5756-0004"]}, "institutions": [{"id": "https://openalex.org/I20231570", "display_name": "Peking University", "ror": "https://ror.org/02v51f717", "country_code": "CN", "type": "education", "lineage": ["https://openalex.org/I20231570"]}], "countries": ["CN"], "is_corresponding": false, "raw_author_name": "Xin Liu", "raw_affiliation_strings": ["ByteDance, Bellevue, WA, USA", "School of Computer Science, Peking University, Beijing, China"], "raw_orcid": null, "affiliations": [{"raw_affiliation_string": "ByteDance, Bellevue, WA, USA", "institution_ids": []}, {"raw_affiliation_string": "School of Computer Science, Peking University, Beijing, China", "institution_ids": ["https://openalex.org/I20231570"]}]}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5052249316", "display_name": "Xuanzhe Liu", "orcid": "https://orcid.org/0000-0002-7908-8484", "observed_orcids": ["https://orcid.org/0000-0002-7908-8484"]}, "institutions": [{"id": "https://openalex.org/I20231570", "display_name": "Peking University", "ror": "https://ror.org/02v51f717", "country_code": "CN", "type": "education", "lineage": ["https://openalex.org/I20231570"]}], "countries": ["CN"], "is_corresponding": false, "raw_author_name": "Xuanzhe Liu", "raw_affiliation_strings": ["ByteDance, Bellevue, WA, USA", "School of Computer Science, Peking University, Beijing, China"], "raw_orcid": "https://orcid.org/0000-0002-7908-8484", "affiliations": [{"raw_affiliation_string": "ByteDance, Bellevue, WA, USA", "institution_ids": []}, {"raw_affiliation_string": "School of Computer Science, Peking University, Beijing, China", "institution_ids": ["https://openalex.org/I20231570"]}]}, {"author_position": "last", "author": {"id": "https://openalex.org/A5101882004", "display_name": "Xin Jin", "orcid": "https://orcid.org/0000-0001-8741-5847", "observed_orcids": ["https://orcid.org/0000-0001-8741-5847", "https://orcid.org/0000-0002-7908-8484"]}, "institutions": [{"id": "https://openalex.org/I20231570", "display_name": "Peking University", "ror": "https://ror.org/02v51f717", "country_code": "CN", "type": "education", "lineage": ["https://openalex.org/I20231570"]}], "countries": ["CN"], "is_corresponding": false, "raw_author_name": "Xin Jin", "raw_affiliation_strings": ["School of Computer Science, Peking University, Beijing, China"], "raw_orcid": "https://orcid.org/0000-0001-8741-5847", "affiliations": [{"raw_affiliation_string": "School of Computer Science, Peking University, Beijing, China", "institution_ids": ["https://openalex.org/I20231570"]}]}], "abstract_inverted_index": {"Low-rank": [0], "adaptation": [1], "(LoRA)": [2], "is": [3], "widely": [4], "used": [5], "to": [6, 14, 127, 170, 180], "efficiently": [7], "fine-tune": [8], "large": [9], "language": [10], "models": [11, 16, 30], "(LLMs),": [12], "leading": [13], "multiple": [15], "fine-tuned": [17], "from": [18, 62, 103], "the": [19, 32, 47, 50, 58, 82, 85, 93, 104, 114, 119, 144, 157, 171, 190], "same": [20, 33], "pre-trained": [21], "LLM.": [22], "State-of-the-art": [23], "LLM": [24], "serving": [25, 78, 94, 110, 174], "systems": [26], "colocate": [27], "these": [28, 160], "LoRA": [29, 55, 64, 90, 122, 150, 195], "on": [31, 137, 152], "GPU": [34], "instances": [35], "for": [36], "concurrent": [37], "serving,": [38], "which": [39, 112], "decreases": [40], "memory": [41], "usage": [42], "and": [43, 57, 155], "boosts": [44], "efficiency.": [45], "However,": [46], "unawareness": [48], "of": [49, 53, 84, 88, 106, 118, 146, 193], "SLO": [51, 69, 86, 153, 191], "requirements": [52], "each": [54, 194], "service": [56], "interference": [59, 147], "between": [60, 148, 159], "requests": [61, 135], "different": [63, 149], "services": [65, 91, 151], "can": [66, 177], "cause": [67], "significant": [68], "violations.": [70], "This": [71, 124], "paper": [72], "presents": [73], "Cannikin,": [74], "a": [75, 107], "multi-LoRA": [76, 109, 173], "inference": [77], "system": [79], "that": [80, 133, 168], "optimizes": [81], "minimum": [83], "attainments": [87], "all": [89], "in": [92], "system,": [95], "denoted": [96], "as": [97], "lagger-SLO": [98], "attainment.": [99], "We": [100], "obtain": [101], "insights": [102], "characterization": [105], "real-world": [108, 165], "trace,": [111], "reveals": [113], "stable": [115], "input/output": [116], "lengths": [117], "most": [120], "popular": [121], "services.": [123, 161], "motivates": [125], "Cannikin": [126, 141, 176], "propose": [128], "an": [129], "SLO-aware": [130], "scheduling": [131], "algorithm": [132], "prioritizes": [134], "based": [136], "efficient": [138], "deadline": [139], "estimation.": [140], "further": [142], "detects": [143], "influence": [145], "violations": [154], "eliminates": [156], "bias": [158], "The": [162], "evaluation": [163], "using": [164], "traces": [166], "demonstrates": [167], "compared": [169], "state-of-the-art": [172], "systems,": [175], "handle": [178], "up": [179], "3.6\u00d7": [181], "higher": [182], "rates": [183], "or": [184], "2.8\u00d7": [185], "more": [186], "burstiness": [187], "while": [188], "maintaining": [189], "attainment": [192], "service$>": [196], "$90%": [197], ".": [198]}, "primary_topic": {"id": "https://openalex.org/T10101", "display_name": "Cloud Computing and Resource Management", "score": 0.3910999894142151, "subfield": {"id": "https://openalex.org/subfields/1710", "display_name": "Information Systems"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, "topics": [{"id": "https://openalex.org/T10101", "display_name": "Cloud Computing and Resource Management", "score": 0.3910999894142151, "subfield": {"id": "https://openalex.org/subfields/1710", "display_name": "Information Systems"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, {"id": "https://openalex.org/T12127", "display_name": "Software System Performance and Reliability", "score": 0.17239999771118164, "subfield": {"id": "https://openalex.org/subfields/1705", "display_name": "Computer Networks and Communications"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, {"id": "https://openalex.org/T10054", "display_name": "Parallel Computing and Optimization Techniques", "score": 0.08839999884366989, "subfield": {"id": "https://openalex.org/subfields/1708", "display_name": "Hardware and Architecture"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}], "primary_location": {"id": "doi:10.1109/tpds.2025.3590014", "is_oa": false, "landing_page_url": "https://doi.org/10.1109/tpds.2025.3590014", "pdf_url": null, "source": {"id": "https://openalex.org/S97130795", "display_name": "IEEE Transactions on Parallel and Distributed Systems", "issn_l": "1045-9219", "issn": ["1045-9219", "1558-2183", "2161-9883"], "is_oa": false, "is_in_doaj": false, "is_core": true, "listed_in": ["cwts-core", "jufo-2", "norway-2"], "host_organization": "https://openalex.org/P4310319808", "host_organization_name": "Institute of Electrical and Electronics Engineers", "host_organization_lineage": ["https://openalex.org/P4310319808"], "host_organization_lineage_names": ["Institute of Electrical and Electronics Engineers"], "type": "journal"}, "license": null, "license_id": null, "version": "publishedVersion", "is_accepted": true, "is_published": true, "raw_source_name": "IEEE Transactions on Parallel and Distributed Systems", "raw_type": "journal-article"}, "cited_by_count": 3, "is_retracted": false}, {"id": "https://openalex.org/W7117156698", "doi": "https://doi.org/10.48550/arxiv.2512.17910", "title": "Efficient Multi-Adapter LLM Serving via Cross-Model KV-Cache Reuse with Activated LoRA", "publication_year": 2025, "publication_date": "2025-11-26", "type": "preprint", "language": null, "authorships": [{"author_position": "first", "author": {"id": "https://openalex.org/A5121239102", "display_name": "Allison Li", "orcid": null, "observed_orcids": []}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Li, Allison", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5002525343", "display_name": "Kristjan H. Greenewald", "orcid": "https://orcid.org/0009-0003-8177-0209", "observed_orcids": ["https://orcid.org/0009-0003-8177-0209"]}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Greenewald, Kristjan", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5040727953", "display_name": "Thomas P. Parnell", "orcid": "https://orcid.org/0000-0002-1308-6590", "observed_orcids": ["https://orcid.org/0000-0002-1308-6590"]}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Parnell, Thomas", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "last", "author": {"id": "https://openalex.org/A5114453013", "display_name": "Navid Azizan", "orcid": null, "observed_orcids": []}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Azizan, Navid", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}], "abstract_inverted_index": {"Modern": [0], "large": [1], "language": [2], "model": [3, 78, 139, 157], "(LLM)": [4], "systems": [5], "increasingly": [6], "rely": [7], "on": [8], "multi-turn": [9], "pipelines": [10, 116], "that": [11, 39, 136], "are": [12], "composed": [13], "of": [14, 149, 167], "multiple": [15], "task-specific": [16], "adapters,": [17], "yet": [18], "existing": [19, 90], "serving": [20, 37, 91], "frameworks": [21], "remain": [22], "inefficient,": [23], "incurring": [24], "substantial": [25], "recomputation": [26], "overhead": [27], "when": [28], "switching": [29, 59], "between": [30, 45], "adapters.": [31], "We": [32], "present": [33], "the": [34, 65, 77, 150, 163], "first": [35, 164], "LLM": [36, 173], "engine": [38, 92], "supports": [40, 102], "cross-model": [41, 168], "prefix": [42], "cache": [43, 82], "reuse": [44, 83, 170], "base": [46], "and": [47, 56, 73, 124, 141, 144], "adapted": [48], "models": [49, 85], "via": [50], "Activated": [51], "LoRA": [52, 132], "(aLoRA),": [53], "enabling": [54], "efficient": [55], "fine-grained": [57], "adapter": [58, 104], "during": [60], "inference.": [61], "Our": [62], "design": [63], "extends": [64], "vLLM": [66], "framework": [67], "by": [68], "introducing": [69], "base-aligned": [70], "block": [71], "hashing": [72], "activation-aware": [74], "masking": [75], "within": [76], "execution": [79], "path,": [80], "permitting": [81], "across": [84, 112, 146], "while": [86], "preserving": [87], "compatibility": [88], "with": [89, 134, 138, 159], "optimizations.": [93], "Integrated": [94], "into": [95], "a": [96], "production-grade": [97], "inference": [98, 174], "stack,": [99], "this": [100], "approach": [101], "dynamic": [103], "activation": [105], "without": [106], "excessive": [107], "key-value": [108], "tensor": [109], "recomputation.": [110], "Evaluation": [111], "representative": [113], "multi-turn,": [114], "multi-adapter": [115], "demonstrates": [117], "up": [118], "to": [119, 130], "58x": [120], "end-to-end": [121], "latency": [122], "reduction": [123], "over": [125], "100x": [126], "time-to-first-token": [127], "improvement": [128], "relative": [129], "standard": [131], "baselines,": [133], "benefits": [135], "scale": [137], "size": [140], "sequence": [142], "length": [143], "manifest": [145], "all": [147], "stages": [148], "request": [151], "lifecycle.": [152], "This": [153], "work": [154], "bridges": [155], "parameter-efficient": [156], "adaptation": [158], "high-performance": [160], "serving,": [161], "providing": [162], "complete": [165], "realization": [166], "KV-cache": [169], "in": [171], "modern": [172], "engines.": [175]}, "primary_topic": {"id": "https://openalex.org/T10181", "display_name": "Natural Language Processing Techniques", "score": 0.3452000021934509, "subfield": {"id": "https://openalex.org/subfields/1702", "display_name": "Artificial Intelligence"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, "topics": [{"id": "https://openalex.org/T10181", "display_name": "Natural Language Processing Techniques", "score": 0.3452000021934509, "subfield": {"id": "https://openalex.org/subfields/1702", "display_name": "Artificial Intelligence"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, {"id": "https://openalex.org/T10028", "display_name": "Topic Modeling", "score": 0.2964000105857849, "subfield": {"id": "https://openalex.org/subfields/1702", "display_name": "Artificial Intelligence"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, {"id": "https://openalex.org/T10036", "display_name": "Advanced Neural Network Applications", "score": 0.12729999423027039, "subfield": {"id": "https://openalex.org/subfields/1707", "display_name": "Computer Vision and Pattern Recognition"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}], "primary_location": {"id": "pmh:oai:arXiv.org:2512.17910", "is_oa": true, "landing_page_url": "http://arxiv.org/abs/2512.17910", "pdf_url": "https://arxiv.org/pdf/2512.17910", "source": {"id": "https://openalex.org/S4306400194", "display_name": "arXiv (Cornell University)", "issn_l": "2331-8422", "issn": ["2331-8422"], "is_oa": true, "is_in_doaj": false, "is_core": false, "listed_in": [], "host_organization": "https://openalex.org/I205783295", "host_organization_name": "Cornell University", "host_organization_lineage": ["https://openalex.org/I205783295"], "host_organization_lineage_names": [], "type": "repository"}, "license": null, "license_id": null, "version": "submittedVersion", "is_accepted": false, "is_published": false, "raw_source_name": null, "raw_type": "text"}, "cited_by_count": 0, "is_retracted": false}, {"id": "https://openalex.org/W4411403202", "doi": "https://doi.org/10.1145/3725394", "title": "Apt-Serve: Adaptive Request Scheduling on Hybrid Cache for Scalable LLM Inference Serving", "publication_year": 2025, "publication_date": "2025-06-17", "type": "article", "language": "en", "authorships": [{"author_position": "first", "author": {"id": "https://openalex.org/A5102938652", "display_name": "Shihong Gao", "orcid": "https://orcid.org/0000-0002-0413-9005", "observed_orcids": ["https://orcid.org/0000-0002-0413-9005"]}, "institutions": [{"id": "https://openalex.org/I200769079", "display_name": "Hong Kong University of Science and Technology", "ror": "https://ror.org/00q4vv597", "country_code": "HK", "type": "education", "lineage": ["https://openalex.org/I200769079"]}], "countries": ["HK"], "is_corresponding": false, "raw_author_name": "Shihong Gao", "raw_affiliation_strings": ["The Hong Kong University of Science and Technology, Hong Kong SAR, China"], "raw_orcid": "https://orcid.org/0000-0002-0413-9005", "affiliations": [{"raw_affiliation_string": "The Hong Kong University of Science and Technology, Hong Kong SAR, China", "institution_ids": ["https://openalex.org/I200769079"]}]}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5100327368", "display_name": "Xin Zhang", "orcid": "https://orcid.org/0000-0001-8560-5006", "observed_orcids": ["https://orcid.org/0000-0001-8560-5006"]}, "institutions": [{"id": "https://openalex.org/I200769079", "display_name": "Hong Kong University of Science and Technology", "ror": "https://ror.org/00q4vv597", "country_code": "HK", "type": "education", "lineage": ["https://openalex.org/I200769079"]}], "countries": ["HK"], "is_corresponding": false, "raw_author_name": "Xin Zhang", "raw_affiliation_strings": ["The Hong Kong University of Science and Technology, Hong Kong SAR, China"], "raw_orcid": "https://orcid.org/0000-0001-8560-5006", "affiliations": [{"raw_affiliation_string": "The Hong Kong University of Science and Technology, Hong Kong SAR, China", "institution_ids": ["https://openalex.org/I200769079"]}]}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5053338416", "display_name": "Yanyan Shen", "orcid": "https://orcid.org/0000-0001-8364-3674", "observed_orcids": ["https://orcid.org/0000-0001-8364-3674"]}, "institutions": [{"id": "https://openalex.org/I183067930", "display_name": "Shanghai Jiao Tong University", "ror": "https://ror.org/0220qvk04", "country_code": "CN", "type": "education", "lineage": ["https://openalex.org/I183067930"]}], "countries": ["CN"], "is_corresponding": false, "raw_author_name": "Yanyan Shen", "raw_affiliation_strings": ["Shanghai Jiao Tong University, Shanghai, China"], "raw_orcid": "https://orcid.org/0000-0001-8364-3674", "affiliations": [{"raw_affiliation_string": "Shanghai Jiao Tong University, Shanghai, China", "institution_ids": ["https://openalex.org/I183067930"]}]}, {"author_position": "last", "author": {"id": "https://openalex.org/A5100333516", "display_name": "Lei Chen", "orcid": "https://orcid.org/0000-0002-8257-5806", "observed_orcids": ["https://orcid.org/0000-0002-8257-5806", "https://orcid.org/0000-0002-9459-5695", "https://orcid.org/0009-0000-2383-6650"]}, "institutions": [{"id": "https://openalex.org/I4405261251", "display_name": "The Hong Kong University of Science and Technology (Guangzhou)", "ror": "https://ror.org/050h0vm43", "country_code": "CN", "type": "education", "lineage": ["https://openalex.org/I4405261251"]}], "countries": ["CN"], "is_corresponding": false, "raw_author_name": "Lei Chen", "raw_affiliation_strings": ["The Hong Kong University of Science and Technology (Guangzhou), Guangzhou, China and The Hong Kong University of Science and Technology, Hong Kong SAR, China"], "raw_orcid": "https://orcid.org/0000-0002-8257-5806", "affiliations": [{"raw_affiliation_string": "The Hong Kong University of Science and Technology (Guangzhou), Guangzhou, China and The Hong Kong University of Science and Technology, Hong Kong SAR, China", "institution_ids": ["https://openalex.org/I4405261251"]}]}], "abstract_inverted_index": {"Large": [0], "language": [1], "model": [2], "(LLM)": [3], "inference": [4, 114, 208], "serving": [5, 209], "systems": [6, 23, 44], "are": [7], "essential": [8], "to": [9, 19, 24, 36, 47, 53, 108, 190, 198, 205], "various": [10], "LLM-based": [11], "applications.": [12], "As": [13], "demand": [14], "for": [15, 132], "LLM": [16, 113], "services": [17], "continues": [18], "grow,": [20], "scaling": [21], "these": [22], "handle": [25], "high": [26], "request": [27, 144], "rates": [28], "while": [29], "meeting": [30], "latency": [31], "Service-Level": [32], "Objectives": [33], "(SLOs),": [34], "referred": [35], "as": [37], "effective": [38, 49, 110, 202], "throughput,": [39, 50], "becomes": [40], "critical.": [41], "However,": [42], "existing": [43], "often": [45], "struggle": [46], "improve": [48], "primarily": [51], "due": [52], "a": [54, 104, 118, 128], "significant": [55], "decline": [56], "in": [57, 112, 201], "Time": [58], "To": [59], "First": [60], "Token": [61], "(TTFT)": [62], "SLO": [63], "attainment.": [64], "We": [65, 163], "identify": [66], "two": [67], "major": [68], "causes": [69], "of": [70], "this": [71, 99], "bottleneck:": [72], "(1)": [73], "memory-intensive": [74], "KV": [75, 125], "cache": [76, 121, 126, 131], "that": [77, 123, 158, 194], "limits": [78], "batch": [79, 89, 140, 161], "size": [80], "expansion": [81], "under": [82], "GPU": [83], "memory": [84], "constraints,": [85], "and": [86, 142, 171, 185], "(2)": [87], "rigid": [88], "composition": [90], "enforced": [91], "by": [92], "the": [93, 148, 166, 206], "default": [94], "First-Come-First-Serve": [95], "scheduling": [96, 156, 168], "policy.": [97], "In": [98], "paper,": [100], "we": [101], "introduce": [102], "Apt-Serve,": [103], "scalable": [105], "framework": [106], "designed": [107], "enhance": [109], "throughput": [111, 203], "serving.": [115], "Apt-Serve": [116, 151, 195], "features": [117], "new": [119], "hybrid": [120, 149], "scheme": [122], "combines": [124], "with": [127, 176], "memory-efficient": [129], "hidden": [130, 135], "reusable": [133], "input": [134], "state": [136], "vectors,": [137], "allowing": [138], "large": [139], "sizes": [141], "improving": [143], "concurrency.": [145], "Based": [146], "on": [147, 181], "cache,": [150], "employs": [152], "an": [153, 173], "adaptive": [154, 167], "runtime": [155], "mechanism": [157], "dynamically": [159], "optimizes": [160], "composition.": [162], "formally": [164], "define": [165], "optimization": [169], "problem": [170], "propose": [172], "efficient": [174], "algorithm": [175], "theoretical": [177], "guarantees.": [178], "Extensive": [179], "evaluations": [180], "three": [182], "real-world": [183], "datasets": [184], "LLMs": [186], "ranging": [187], "from": [188], "13B": [189], "66B": [191], "parameters": [192], "demonstrate": [193], "achieves": [196], "up": [197], "8.8x": [199], "improvement": [200], "compared": [204], "state-of-the-art": [207], "systems.": [210]}, "primary_topic": {"id": "https://openalex.org/T10054", "display_name": "Parallel Computing and Optimization Techniques", "score": 0.23389999568462372, "subfield": {"id": "https://openalex.org/subfields/1708", "display_name": "Hardware and Architecture"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, "topics": [{"id": "https://openalex.org/T10054", "display_name": "Parallel Computing and Optimization Techniques", "score": 0.23389999568462372, "subfield": {"id": "https://openalex.org/subfields/1708", "display_name": "Hardware and Architecture"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, {"id": "https://openalex.org/T12127", "display_name": "Software System Performance and Reliability", "score": 0.1818999946117401, "subfield": {"id": "https://openalex.org/subfields/1705", "display_name": "Computer Networks and Communications"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, {"id": "https://openalex.org/T10101", "display_name": "Cloud Computing and Resource Management", "score": 0.11140000075101852, "subfield": {"id": "https://openalex.org/subfields/1710", "display_name": "Information Systems"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}], "primary_location": {"id": "doi:10.1145/3725394", "is_oa": true, "landing_page_url": "https://doi.org/10.1145/3725394", "pdf_url": "https://dl.acm.org/doi/pdf/10.1145/3725394", "source": {"id": "https://openalex.org/S4387289859", "display_name": "Proceedings of the ACM on Management of Data", "issn_l": "2836-6573", "issn": ["2836-6573"], "is_oa": false, "is_in_doaj": false, "is_core": true, "listed_in": ["cwts-core", "norway-1"], "host_organization": "https://openalex.org/P4310319798", "host_organization_name": "Association for Computing Machinery", "host_organization_lineage": ["https://openalex.org/P4310319798"], "host_organization_lineage_names": ["Association for Computing Machinery"], "type": "journal"}, "license": null, "license_id": null, "version": "publishedVersion", "is_accepted": true, "is_published": true, "raw_source_name": "Proceedings of the ACM on Management of Data", "raw_type": "journal-article"}, "cited_by_count": 7, "is_retracted": false}, {"id": "https://openalex.org/W4417308116", "doi": "https://doi.org/10.48550/arxiv.2506.09342", "title": "Latent Multi-Head Attention for Small Language Models", "publication_year": 2025, "publication_date": "2025-06-11", "type": "preprint", "language": "en", "authorships": [{"author_position": "first", "author": {"id": "https://openalex.org/A5123144009", "display_name": "Sushant Mehta", "orcid": null, "observed_orcids": []}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Mehta, Sushant", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5021661240", "display_name": "Raj Dandekar", "orcid": "https://orcid.org/0000-0003-1946-0667", "observed_orcids": ["https://orcid.org/0000-0003-1946-0667"]}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Dandekar, Raj", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5109819059", "display_name": "Rajat Dandekar", "orcid": null, "observed_orcids": []}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Dandekar, Rajat", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "last", "author": {"id": "https://openalex.org/A5034707872", "display_name": "Sreedath Panat", "orcid": null, "observed_orcids": []}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Panat, Sreedath", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}], "abstract_inverted_index": {"We": [0, 83], "present": [1], "the": [2, 132, 143], "first": [3], "comprehensive": [4], "study": [5], "of": [6], "latent": [7, 52], "multi-head": [8, 33], "attention": [9, 34, 100], "(MLA)": [10], "for": [11, 79, 90], "small": [12, 93], "language": [13], "models,": [14], "revealing": [15], "interesting": [16], "efficiency-quality": [17], "trade-offs.": [18], "Training": [19], "30M-parameter": [20], "GPT": [21], "models": [22, 156], "on": [23, 113], "100,000": [24], "synthetic": [25], "stories,": [26], "we": [27], "benchmark": [28], "three": [29], "architectural": [30], "variants:": [31], "standard": [32], "(MHA),": [35], "MLA,": [36], "and": [37, 151, 155], "MLA": [38, 91, 97, 119, 129], "with": [39, 50, 104, 120, 140], "rotary": [40], "positional": [41], "embeddings": [42], "(MLA+RoPE).": [43], "Our": [44], "key": [45], "finding": [46], "is": [47, 88], "that": [48, 86, 118], "MLA+RoPE": [49], "half-rank": [51], "dimensions": [53], "(r": [54], "=": [55], "d/2)": [56], "achieves": [57, 122], "a": [58, 66, 76, 123], "45%": [59], "KV-cache": [60], "memory": [61, 80, 133], "reduction": [62], "while": [63, 130], "incurring": [64], "only": [65], "0.3%": [67], "increase": [68], "in": [69, 92], "validation": [70], "loss": [71], "(essentially": [72], "matching": [73], "MHA": [74], "quality)-": [75], "Pareto": [77], "improvement": [78], "constrained": [81], "deployment.": [82], "further": [84], "show": [85], "RoPE": [87], "crucial": [89], "models:": [94], "without": [95], "it,": [96], "underperforms": [98], "vanilla": [99, 108], "by": [101, 109], "3-5%,": [102], "but": [103], "RoPE,": [105], "it": [106], "surpasses": [107], "2%.": [110], "Inference": [111], "benchmarks": [112], "NVIDIA": [114], "A100": [115], "GPUs": [116], "reveal": [117], "r=d/2": [121], "1.4": [124], "times": [125], "speedup": [126], "over": [127], "full-rank": [128], "maintaining": [131], "savings.": [134], "GPT-4": [135], "evaluations": [136], "corroborate": [137], "perplexity": [138], "results,": [139], "ours": [141], "achieving": [142], "highest": [144], "quality": [145], "scores": [146], "(7.4/10)": [147], "across": [148], "grammar,": [149], "creativity,": [150], "consistency": [152], "metrics.": [153], "Code": [154], "will": [157], "be": [158], "released": [159], "upon": [160], "acceptance.": [161]}, "primary_topic": {"id": "https://openalex.org/T10028", "display_name": "Topic Modeling", "score": 0.8871999979019165, "subfield": {"id": "https://openalex.org/subfields/1702", "display_name": "Artificial Intelligence"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, "topics": [{"id": "https://openalex.org/T10028", "display_name": "Topic Modeling", "score": 0.8871999979019165, "subfield": {"id": "https://openalex.org/subfields/1702", "display_name": "Artificial Intelligence"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, {"id": "https://openalex.org/T10181", "display_name": "Natural Language Processing Techniques", "score": 0.08910000324249268, "subfield": {"id": "https://openalex.org/subfields/1702", "display_name": "Artificial Intelligence"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, {"id": "https://openalex.org/T10036", "display_name": "Advanced Neural Network Applications", "score": 0.006300000008195639, "subfield": {"id": "https://openalex.org/subfields/1707", "display_name": "Computer Vision and Pattern Recognition"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}], "primary_location": {"id": "pmh:oai:arXiv.org:2506.09342", "is_oa": true, "landing_page_url": "http://arxiv.org/abs/2506.09342", "pdf_url": "https://arxiv.org/pdf/2506.09342", "source": {"id": "https://openalex.org/S4306400194", "display_name": "arXiv (Cornell University)", "issn_l": "2331-8422", "issn": ["2331-8422"], "is_oa": true, "is_in_doaj": false, "is_core": false, "listed_in": [], "host_organization": "https://openalex.org/I205783295", "host_organization_name": "Cornell University", "host_organization_lineage": ["https://openalex.org/I205783295"], "host_organization_lineage_names": [], "type": "repository"}, "license": null, "license_id": null, "version": "submittedVersion", "is_accepted": false, "is_published": false, "raw_source_name": null, "raw_type": "text"}, "cited_by_count": 1, "is_retracted": false}, {"id": "https://openalex.org/W4417302158", "doi": "https://doi.org/10.48550/arxiv.2505.12728", "title": "SpecFLASH: A Latent-Guided Semi-autoregressive Speculative Decoding Framework for Efficient Multimodal Generation", "publication_year": 2025, "publication_date": "2025-05-19", "type": "preprint", "language": "en", "authorships": [{"author_position": "first", "author": {"id": "https://openalex.org/A5101444415", "display_name": "Zihua Wang", "orcid": "https://orcid.org/0009-0004-5871-6973", "observed_orcids": ["https://orcid.org/0009-0004-5871-6973", "https://orcid.org/0009-0009-5127-0143"]}, "institutions": [{"id": "https://openalex.org/I76569877", "display_name": "Southeast University", "ror": "https://ror.org/04ct4d772", "country_code": "CN", "type": "education", "lineage": ["https://openalex.org/I76569877"]}], "countries": ["CN"], "is_corresponding": false, "raw_author_name": "Wang, Zihua", "raw_affiliation_strings": ["Southeast University"], "raw_orcid": null, "affiliations": [{"raw_affiliation_string": "Southeast University", "institution_ids": ["https://openalex.org/I76569877"]}]}, {"author_position": "middle", "author": {"id": null, "display_name": "Li, Ruibo", "orcid": null, "observed_orcids": []}, "institutions": [{"id": "https://openalex.org/I172675005", "display_name": "Nanyang Technological University", "ror": "https://ror.org/02e7b5302", "country_code": "SG", "type": "education", "lineage": ["https://openalex.org/I172675005"]}], "countries": ["SG"], "is_corresponding": false, "raw_author_name": "Li, Ruibo", "raw_affiliation_strings": ["Nanyang Technological University"], "raw_orcid": null, "affiliations": [{"raw_affiliation_string": "Nanyang Technological University", "institution_ids": ["https://openalex.org/I172675005"]}]}, {"author_position": "middle", "author": {"id": null, "display_name": "Du, Haozhe", "orcid": null, "observed_orcids": []}, "institutions": [{"id": "https://openalex.org/I16609230", "display_name": "Hunan University", "ror": "https://ror.org/05htk5m33", "country_code": "CN", "type": "education", "lineage": ["https://openalex.org/I16609230"]}], "countries": ["CN"], "is_corresponding": false, "raw_author_name": "Du, Haozhe", "raw_affiliation_strings": ["Hunan University"], "raw_orcid": null, "affiliations": [{"raw_affiliation_string": "Hunan University", "institution_ids": ["https://openalex.org/I16609230"]}]}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5045125183", "display_name": "Joey Tianyi Zhou", "orcid": "https://orcid.org/0000-0002-4675-7055", "observed_orcids": ["https://orcid.org/0000-0002-4675-7055"]}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Zhou, Joey Tianyi", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5100433691", "display_name": "Yu Zhang", "orcid": "https://orcid.org/0000-0003-1100-4835", "observed_orcids": ["https://orcid.org/0000-0003-1100-4835", "https://orcid.org/0000-0003-4087-6544"]}, "institutions": [{"id": "https://openalex.org/I76569877", "display_name": "Southeast University", "ror": "https://ror.org/04ct4d772", "country_code": "CN", "type": "education", "lineage": ["https://openalex.org/I76569877"]}], "countries": ["CN"], "is_corresponding": false, "raw_author_name": "Zhang, Yu", "raw_affiliation_strings": ["Southeast University"], "raw_orcid": null, "affiliations": [{"raw_affiliation_string": "Southeast University", "institution_ids": ["https://openalex.org/I76569877"]}]}, {"author_position": "last", "author": {"id": "https://openalex.org/A5030256046", "display_name": "Xu Yang", "orcid": "https://orcid.org/0000-0002-5887-2704", "observed_orcids": ["https://orcid.org/0000-0002-5887-2704"]}, "institutions": [{"id": "https://openalex.org/I76569877", "display_name": "Southeast University", "ror": "https://ror.org/04ct4d772", "country_code": "CN", "type": "education", "lineage": ["https://openalex.org/I76569877"]}], "countries": ["CN"], "is_corresponding": false, "raw_author_name": "Yang, Xu", "raw_affiliation_strings": ["STAR Centre for Frontier AI Research (CFAR)", "Southeast University"], "raw_orcid": null, "affiliations": [{"raw_affiliation_string": "STAR Centre for Frontier AI Research (CFAR)", "institution_ids": []}, {"raw_affiliation_string": "Southeast University", "institution_ids": ["https://openalex.org/I76569877"]}]}], "abstract_inverted_index": {"Large": [0], "language": [1], "models": [2, 6], "and": [3, 8, 84, 136, 141, 180], "large": [4], "multimodal": [5, 72, 107], "(LLMs": [7], "LMMs)": [9], "deliver": [10], "strong": [11], "generative": [12], "performance": [13], "but": [14], "suffer": [15], "from": [16], "slow": [17], "decoding,": [18], "a": [19, 49, 61, 97, 123, 148, 157], "problem": [20], "that": [21, 56, 104, 129, 152, 164], "becomes": [22], "more": [23, 34], "severe": [24], "when": [25, 109], "handling": [26], "visual": [27, 82, 119, 131, 145, 183], "inputs,": [28], "whose": [29], "sequences": [30, 121], "typically": [31], "contain": [32], "many": [33], "tokens": [35, 55, 155], "with": [36, 122], "lower": [37], "information": [38], "density": [39], "than": [40], "text.": [41], "Speculative": [42], "decoding": [43, 74, 99, 150, 170], "accelerates": [44], "LLM": [45], "inference": [46], "by": [47, 60], "letting": [48], "compact": [50], "draft": [51, 89, 112], "model": [52], "propose": [53], "candidate": [54], "are": [57], "selectively": [58], "accepted": [59], "larger": [62], "target": [63], "model,": [64], "achieving": [65, 172], "speed-up": [66, 176], "without": [67], "degrading": [68], "quality.": [69], "However,": [70], "existing": [71], "speculative": [73, 98, 169], "approaches": [75], "largely": [76], "ignore": [77], "the": [78, 111, 139, 188], "structural": [79], "characteristics": [80], "of": [81, 144], "representations": [83], "usually": [85], "rely": [86], "on": [87, 177, 182], "text-only": [88], "models.": [90], "In": [91], "this": [92], "paper,": [93], "we": [94], "introduce": [95], "SpecFLASH,": [96], "framework": [100], "tailored": [101], "to": [102, 174, 187], "LMMs": [103], "explicitly": [105], "exploits": [106], "structure": [108], "designing": [110], "model.": [113], "We": [114], "first": [115], "mitigate": [116], "redundancy": [117], "in": [118, 156], "token": [120, 126], "lightweight,": [124], "latent-guided": [125], "compression": [127], "module": [128], "compacts": [130], "features": [132], "while": [133], "preserving": [134], "semantics,": [135], "then": [137], "leverage": [138], "co-occurrence": [140], "local": [142], "correlations": [143], "entities": [146], "via": [147], "semi-autoregressive": [149], "scheme": [151], "predicts": [153], "multiple": [154], "single": [158], "forward": [159], "pass.": [160], "Extensive": [161], "experiments": [162], "demonstrate": [163], "SpecFLASH": [165], "consistently": [166], "surpasses": [167], "prior": [168], "baselines,": [171], "up": [173], "$2.68\\times$": [175], "video": [178], "captioning": [179], "$2.55\\times$": [181], "instruction": [184], "tuning,": [185], "relative": [186], "original": [189], "LMM.": [190], "Our": [191], "code": [192], "is": [193], "available": [194], "here:": [195], "https://github.com/ZihuaEvan/FlashSD/.": [196]}, "primary_topic": {"id": "https://openalex.org/T11714", "display_name": "Multimodal Machine Learning Applications", "score": 0.9984999895095825, "subfield": {"id": "https://openalex.org/subfields/1707", "display_name": "Computer Vision and Pattern Recognition"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, "topics": [{"id": "https://openalex.org/T11714", "display_name": "Multimodal Machine Learning Applications", "score": 0.9984999895095825, "subfield": {"id": "https://openalex.org/subfields/1707", "display_name": "Computer Vision and Pattern Recognition"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, {"id": "https://openalex.org/T10181", "display_name": "Natural Language Processing Techniques", "score": 0.0006000000284984708, "subfield": {"id": "https://openalex.org/subfields/1702", "display_name": "Artificial Intelligence"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, {"id": "https://openalex.org/T10775", "display_name": "Generative Adversarial Networks and Image Synthesis", "score": 0.00039999998989515007, "subfield": {"id": "https://openalex.org/subfields/1707", "display_name": "Computer Vision and Pattern Recognition"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}], "primary_location": {"id": "pmh:oai:arXiv.org:2505.12728", "is_oa": true, "landing_page_url": "http://arxiv.org/abs/2505.12728", "pdf_url": "https://arxiv.org/pdf/2505.12728", "source": {"id": "https://openalex.org/S4306400194", "display_name": "arXiv (Cornell University)", "issn_l": "2331-8422", "issn": ["2331-8422"], "is_oa": true, "is_in_doaj": false, "is_core": false, "listed_in": [], "host_organization": "https://openalex.org/I205783295", "host_organization_name": "Cornell University", "host_organization_lineage": ["https://openalex.org/I205783295"], "host_organization_lineage_names": [], "type": "repository"}, "license": "cc-by", "license_id": "https://openalex.org/licenses/cc-by", "version": "submittedVersion", "is_accepted": false, "is_published": false, "raw_source_name": null, "raw_type": "text"}, "cited_by_count": 0, "is_retracted": false}, {"id": "https://openalex.org/W4417119515", "doi": "https://doi.org/10.48550/arxiv.2510.09012", "title": "Towards Better & Faster Autoregressive Image Generation: From the Perspective of Entropy", "publication_year": 2025, "publication_date": "2025-10-10", "type": "preprint", "language": "en", "authorships": [{"author_position": "first", "author": {"id": "https://openalex.org/A5100547535", "display_name": "Ma Xiaoxiao", "orcid": null, "observed_orcids": []}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Ma, Xiaoxiao", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5102740754", "display_name": "Feng Zhao", "orcid": "https://orcid.org/0000-0002-5730-2208", "observed_orcids": ["https://orcid.org/0000-0002-5730-2208"]}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Zhao, Feng", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": null, "display_name": "Ling, Pengyang", "orcid": null, "observed_orcids": []}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Ling, Pengyang", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": null, "display_name": "Qiu, Haibo", "orcid": null, "observed_orcids": []}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Qiu, Haibo", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5023937902", "display_name": "Zhixiang Wei", "orcid": "https://orcid.org/0000-0001-6188-3634", "observed_orcids": ["https://orcid.org/0000-0001-6188-3634"]}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Wei, Zhixiang", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5103662468", "display_name": "Hu Yu", "orcid": "https://orcid.org/0000-0003-0598-8989", "observed_orcids": ["https://orcid.org/0000-0003-0598-8989"]}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Yu, Hu", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5006655764", "display_name": "Jie Huang", "orcid": "https://orcid.org/0000-0001-6962-9040", "observed_orcids": ["https://orcid.org/0000-0001-6962-9040"]}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Huang, Jie", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": null, "display_name": "Zeng, Zhixiong", "orcid": null, "observed_orcids": []}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Zeng, Zhixiong", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "last", "author": {"id": "https://openalex.org/A5018662596", "display_name": "Lin Ma", "orcid": "https://orcid.org/0000-0003-1684-9714", "observed_orcids": ["https://orcid.org/0000-0003-1684-9714", "https://orcid.org/0000-0001-6446-1824"]}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Ma, Lin", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}], "abstract_inverted_index": {"In": [0], "this": [1], "work,": [2], "we": [3, 33], "first": [4], "revisit": [5], "the": [6, 50, 69, 104, 123], "sampling": [7, 136], "issues": [8], "in": [9, 79, 94, 130], "current": [10], "autoregressive": [11, 42], "(AR)": [12], "image": [13, 19, 119], "generation": [14, 43, 99, 120, 133], "models": [15, 121], "and": [16, 28, 76, 82, 89, 125, 135], "identify": [17], "that": [18, 39], "tokens,": [20, 23], "unlike": [21], "text": [22], "exhibit": [24], "lower": [25], "information": [26], "density": [27], "non-uniform": [29], "spatial": [30, 63], "distribution.": [31], "Accordingly,": [32], "present": [34], "an": [35], "entropy-informed": [36], "decoding": [37], "strategy": [38], "facilitates": [40], "higher": [41], "quality": [44, 134], "with": [45], "faster": [46], "synthesis": [47], "speed.": [48, 137], "Specifically,": [49], "proposed": [51], "method": [52], "introduces": [53], "two": [54], "main": [55], "innovations:": [56], "1)": [57], "dynamic": [58], "temperature": [59], "control": [60], "guided": [61], "by": [62], "entropy": [64], "of": [65, 103, 107, 127], "token": [66], "distributions,": [67], "enhancing": [68, 131], "balance": [70], "between": [71], "content": [72], "diversity,": [73], "alignment": [74], "accuracy,": [75], "structural": [77], "coherence": [78], "both": [80, 132], "mask-based": [81], "scale-wise": [83], "models,": [84], "without": [85], "extra": [86], "computational": [87], "overhead,": [88], "2)": [90], "entropy-aware": [91], "acceptance": [92], "rules": [93], "speculative": [95], "decoding,": [96], "achieving": [97], "near-lossless": [98], "at": [100], "about": [101], "85\\%": [102], "inference": [105], "cost": [106], "conventional": [108], "acceleration": [109], "methods.": [110], "Extensive": [111], "experiments": [112], "across": [113], "multiple": [114], "benchmarks": [115], "using": [116], "diverse": [117], "AR": [118], "demonstrate": [122], "effectiveness": [124], "generalizability": [126], "our": [128], "approach": [129]}, "primary_topic": {"id": "https://openalex.org/T10775", "display_name": "Generative Adversarial Networks and Image Synthesis", "score": 0.9994999766349792, "subfield": {"id": "https://openalex.org/subfields/1707", "display_name": "Computer Vision and Pattern Recognition"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, "topics": [{"id": "https://openalex.org/T10775", "display_name": "Generative Adversarial Networks and Image Synthesis", "score": 0.9994999766349792, "subfield": {"id": "https://openalex.org/subfields/1707", "display_name": "Computer Vision and Pattern Recognition"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, {"id": "https://openalex.org/T10036", "display_name": "Advanced Neural Network Applications", "score": 0.00019999999494757503, "subfield": {"id": "https://openalex.org/subfields/1707", "display_name": "Computer Vision and Pattern Recognition"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, {"id": "https://openalex.org/T11714", "display_name": "Multimodal Machine Learning Applications", "score": 9.999999747378752e-05, "subfield": {"id": "https://openalex.org/subfields/1707", "display_name": "Computer Vision and Pattern Recognition"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}], "primary_location": {"id": "pmh:oai:arXiv.org:2510.09012", "is_oa": true, "landing_page_url": "http://arxiv.org/abs/2510.09012", "pdf_url": "https://arxiv.org/pdf/2510.09012", "source": {"id": "https://openalex.org/S4306400194", "display_name": "arXiv (Cornell University)", "issn_l": "2331-8422", "issn": ["2331-8422"], "is_oa": true, "is_in_doaj": false, "is_core": false, "listed_in": [], "host_organization": "https://openalex.org/I205783295", "host_organization_name": "Cornell University", "host_organization_lineage": ["https://openalex.org/I205783295"], "host_organization_lineage_names": [], "type": "repository"}, "license": null, "license_id": null, "version": "submittedVersion", "is_accepted": false, "is_published": false, "raw_source_name": null, "raw_type": "text"}, "cited_by_count": 0, "is_retracted": false}, {"id": "https://openalex.org/W7117572630", "doi": "https://doi.org/10.48550/arxiv.2512.21487", "title": "Efficient MoE Inference with Fine-Grained Scheduling of Disaggregated Expert Parallelism", "publication_year": 2025, "publication_date": "2025-12-25", "type": "preprint", "language": null, "authorships": [{"author_position": "first", "author": {"id": "https://openalex.org/A5121563876", "display_name": "Xinglin Pan", "orcid": null, "observed_orcids": []}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Pan, Xinglin", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5016836702", "display_name": "Shaohuai Shi", "orcid": "https://orcid.org/0000-0002-1418-5160", "observed_orcids": ["https://orcid.org/0000-0002-1418-5160"]}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Shi, Shaohuai", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5110841034", "display_name": "Wenxiang Lin", "orcid": "https://orcid.org/0000-0001-5147-0844", "observed_orcids": ["https://orcid.org/0000-0001-5147-0844"]}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Lin, Wenxiang", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5121584127", "display_name": "Yuxin Wang", "orcid": null, "observed_orcids": []}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Wang, Yuxin", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5048313055", "display_name": "Zhenheng Tang", "orcid": "https://orcid.org/0000-0001-8769-9974", "observed_orcids": ["https://orcid.org/0000-0001-8769-9974"]}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Tang, Zhenheng", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5121541460", "display_name": "Wei Wang", "orcid": "https://orcid.org/0000-0002-1717-5785", "observed_orcids": ["https://orcid.org/0000-0002-1717-5785"]}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Wang, Wei", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "last", "author": {"id": "https://openalex.org/A5124876592", "display_name": "Xiaowen Chu", "orcid": null, "observed_orcids": []}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Chu, Xiaowen", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}], "abstract_inverted_index": {"The": [0], "mixture-of-experts": [1], "(MoE)": [2], "architecture": [3], "scales": [4], "model": [5], "size": [6], "with": [7, 107], "sublinear": [8], "computational": [9], "increase": [10], "but": [11, 37], "suffers": [12], "from": [13], "memory-intensive": [14], "inference": [15, 66], "due": [16], "to": [17, 33, 63, 117, 124], "KV": [18], "caches": [19], "and": [20, 31, 43, 89, 91, 109], "sparse": [21], "expert": [22, 26], "activation.": [23], "Recent": [24], "disaggregated": [25], "parallelism": [27], "(DEP)": [28], "distributes": [29], "attention": [30], "experts": [32, 42], "dedicated": [34], "GPU": [35, 105], "groups": [36], "lacks": [38], "support": [39], "for": [40, 57, 78, 97], "shared": [41], "efficient": [44, 95], "task": [45, 54, 61], "scheduling,": [46], "limiting": [47], "performance.": [48], "We": [49], "propose": [50], "FinDEP,": [51], "a": [52, 83, 128], "fine-grained": [53, 79], "scheduling": [55, 84], "algorithm": [56], "DEP": [58], "that": [59], "maximizes": [60], "overlap": [62], "improve": [64], "MoE": [65], "throughput.": [67], "FinDEP": [68, 112], "introduces": [69], "three": [70], "innovations:": [71], "1)": [72], "partitioning": [73], "computation/communication": [74], "into": [75], "smaller": [76], "tasks": [77], "pipelining,": [80], "2)": [81], "formulating": [82], "optimization": [85], "supporting": [86], "variable": [87], "granularity": [88], "ordering,": [90], "3)": [92], "developing": [93], "an": [94], "solver": [96], "this": [98], "large": [99], "search": [100], "space.": [101], "Experiments": [102], "on": [103, 127], "four": [104], "systems": [106], "DeepSeek-V2": [108], "Qwen3-MoE": [110], "show": [111], "improves": [113], "throughput": [114], "by": [115], "up": [116, 123], "1.61x": [118], "over": [119], "prior": [120], "methods,": [121], "achieving": [122], "1.24x": [125], "speedup": [126], "32-GPU": [129], "system.": [130]}, "primary_topic": {"id": "https://openalex.org/T10054", "display_name": "Parallel Computing and Optimization Techniques", "score": 0.9926999807357788, "subfield": {"id": "https://openalex.org/subfields/1708", "display_name": "Hardware and Architecture"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, "topics": [{"id": "https://openalex.org/T10054", "display_name": "Parallel Computing and Optimization Techniques", "score": 0.9926999807357788, "subfield": {"id": "https://openalex.org/subfields/1708", "display_name": "Hardware and Architecture"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, {"id": "https://openalex.org/T10036", "display_name": "Advanced Neural Network Applications", "score": 0.00279999990016222, "subfield": {"id": "https://openalex.org/subfields/1707", "display_name": "Computer Vision and Pattern Recognition"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, {"id": "https://openalex.org/T10181", "display_name": "Natural Language Processing Techniques", "score": 0.00039999998989515007, "subfield": {"id": "https://openalex.org/subfields/1702", "display_name": "Artificial Intelligence"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}], "primary_location": {"id": "pmh:oai:arXiv.org:2512.21487", "is_oa": true, "landing_page_url": "http://arxiv.org/abs/2512.21487", "pdf_url": "https://arxiv.org/pdf/2512.21487", "source": {"id": "https://openalex.org/S4306400194", "display_name": "arXiv (Cornell University)", "issn_l": "2331-8422", "issn": ["2331-8422"], "is_oa": true, "is_in_doaj": false, "is_core": false, "listed_in": [], "host_organization": "https://openalex.org/I205783295", "host_organization_name": "Cornell University", "host_organization_lineage": ["https://openalex.org/I205783295"], "host_organization_lineage_names": [], "type": "repository"}, "license": null, "license_id": null, "version": "submittedVersion", "is_accepted": false, "is_published": false, "raw_source_name": null, "raw_type": "text"}, "cited_by_count": 0, "is_retracted": false}, {"id": "https://openalex.org/W7124316852", "doi": "https://doi.org/10.48550/arxiv.2601.08833", "title": "Revisiting Disaggregated Large Language Model Serving for Performance and Energy Implications", "publication_year": 2025, "publication_date": "2025-11-14", "type": "preprint", "language": null, "authorships": [{"author_position": "first", "author": {"id": "https://openalex.org/A5114549779", "display_name": "Jiaxi Li", "orcid": "https://orcid.org/0009-0005-1936-6023", "observed_orcids": ["https://orcid.org/0009-0005-1936-6023"]}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Li, Jiaxi", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5123061133", "display_name": "Yue Zhu", "orcid": null, "observed_orcids": []}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Zhu, Yue", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5123132426", "display_name": "Eun Kyung Lee", "orcid": null, "observed_orcids": []}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Lee, Eun Kyung", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "last", "author": {"id": "https://openalex.org/A5074224097", "display_name": "Klara Nahrstedt", "orcid": "https://orcid.org/0000-0001-6813-3043", "observed_orcids": ["https://orcid.org/0000-0001-6813-3043"]}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Nahrstedt, Klara", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}], "abstract_inverted_index": {"Different": [0], "from": [1, 178], "traditional": [2], "Large": [3], "Language": [4], "Model": [5], "(LLM)": [6], "serving": [7, 20, 130], "that": [8, 69, 175], "colocates": [9], "the": [10, 16, 30, 36, 43, 156, 165, 187], "prefill": [11, 25, 31], "and": [12, 26, 59, 73, 85, 96, 121, 132, 148, 154, 184, 190], "decode": [13, 27, 44], "stages": [14], "on": [15, 186], "same": [17], "GPU,": [18], "disaggregated": [19, 92, 134, 216], "dedicates": [21], "distinct": [22], "GPUs": [23], "to": [24, 42, 163, 206, 210], "workload.": [28], "Once": [29], "GPU": [32, 143], "completes": [33], "its": [34], "task,": [35], "KV": [37, 52, 82, 118, 138, 191], "cache": [38, 53, 83, 139], "must": [39], "be": [40], "transferred": [41], "GPU.": [45], "While": [46], "existing": [47], "works": [48], "have": [49, 88, 99], "proposed": [50], "various": [51], "transfer": [54, 119, 140, 192], "paths": [55], "across": [56, 160], "different": [57, 117, 137], "memory": [58], "storage": [60], "tiers,": [61], "there": [62], "remains": [63], "a": [64, 127], "lack": [65], "of": [66, 215], "systematic": [67], "benchmarking": [68], "compares": [70], "their": [71, 94], "performance": [72, 95, 176], "energy": [74, 97, 167, 207, 213], "efficiency.": [75], "Meanwhile,": [76], "although": [77], "optimization": [78, 122], "techniques": [79], "such": [80], "as": [81], "reuse": [84], "frequency": [86, 149, 198], "scaling": [87, 150, 199], "been": [89, 101], "utilized": [90], "for": [91], "serving,": [93], "implications": [98], "not": [100, 182, 204], "rigorously": [102], "benchmarked.": [103], "In": [104, 194], "this": [105, 109], "paper,": [106], "we": [107, 125, 152], "fill": [108], "research": [110], "gap": [111], "by": [112, 170, 201], "re-evaluating": [113], "prefill-decode": [114, 179], "disaggregation": [115, 180, 202], "under": [116, 136], "mediums": [120], "strategies.": [123], "Specifically,": [124], "include": [126], "new": [128], "colocated": [129], "baseline": [131], "evaluate": [133, 164], "setups": [135, 162], "paths.": [141], "Through": [142], "profiling": [144], "using": [145], "dynamic": [146], "voltage": [147], "(DVFS),": [151], "identify": [153], "compare": [155], "performance-energy": [157], "Pareto": [158], "frontiers": [159], "all": [161], "potential": [166], "savings": [168], "enabled": [169, 200], "disaggregation.": [171], "Our": [172], "results": [173], "show": [174], "benefits": [177], "are": [181], "guaranteed": [183], "depend": [185], "request": [188], "load": [189], "mediums.": [193], "addition,": [195], "stage-wise": [196], "independent": [197], "does": [203], "lead": [205], "saving": [208], "due": [209], "inherently": [211], "higher": [212], "consumption": [214], "serving.": [217]}, "primary_topic": {"id": "https://openalex.org/T10054", "display_name": "Parallel Computing and Optimization Techniques", "score": 0.9901999831199646, "subfield": {"id": "https://openalex.org/subfields/1708", "display_name": "Hardware and Architecture"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, "topics": [{"id": "https://openalex.org/T10054", "display_name": "Parallel Computing and Optimization Techniques", "score": 0.9901999831199646, "subfield": {"id": "https://openalex.org/subfields/1708", "display_name": "Hardware and Architecture"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, {"id": "https://openalex.org/T14098", "display_name": "Energy Efficiency in Computing", "score": 0.0019000000320374966, "subfield": {"id": "https://openalex.org/subfields/1708", "display_name": "Hardware and Architecture"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, {"id": "https://openalex.org/T10101", "display_name": "Cloud Computing and Resource Management", "score": 0.0010000000474974513, "subfield": {"id": "https://openalex.org/subfields/1710", "display_name": "Information Systems"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}], "primary_location": {"id": "doi:10.48550/arxiv.2601.08833", "is_oa": true, "landing_page_url": "https://doi.org/10.48550/arxiv.2601.08833", "pdf_url": null, "source": {"id": "https://openalex.org/S4306400194", "display_name": "arXiv (Cornell University)", "issn_l": "2331-8422", "issn": ["2331-8422"], "is_oa": true, "is_in_doaj": false, "is_core": false, "listed_in": [], "host_organization": "https://openalex.org/I205783295", "host_organization_name": "Cornell University", "host_organization_lineage": ["https://openalex.org/I205783295"], "host_organization_lineage_names": [], "type": "repository"}, "license": null, "license_id": null, "version": null, "is_accepted": false, "is_published": false, "raw_source_name": null, "raw_type": "Preprint"}, "cited_by_count": 0, "is_retracted": false}, {"id": "https://openalex.org/W4407759493", "doi": "https://doi.org/10.48550/arxiv.2502.12574", "title": "HeadInfer: Memory-Efficient LLM Inference by Head-wise Offloading", "publication_year": 2025, "publication_date": "2025-02-18", "type": "preprint", "language": "en", "authorships": [{"author_position": "first", "author": {"id": null, "display_name": "Luo, Cheng", "orcid": null, "observed_orcids": []}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Luo, Cheng", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5101649092", "display_name": "Zefan Cai", "orcid": "https://orcid.org/0009-0009-8849-8854", "observed_orcids": ["https://orcid.org/0009-0009-8849-8854"]}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Cai, Zefan", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": null, "display_name": "Sun, Hanshi", "orcid": null, "observed_orcids": []}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Sun, Hanshi", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5068769617", "display_name": "Jinqi Xiao", "orcid": "https://orcid.org/0009-0004-7311-9413", "observed_orcids": ["https://orcid.org/0009-0004-7311-9413"]}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Xiao, Jinqi", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": null, "display_name": "Yuan, Bo", "orcid": null, "observed_orcids": []}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Yuan, Bo", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5108707327", "display_name": "Xiao Wen", "orcid": null, "observed_orcids": []}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Xiao, Wen", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5053850820", "display_name": "Junjie Hu", "orcid": "https://orcid.org/0000-0001-5039-3605", "observed_orcids": ["https://orcid.org/0000-0001-5039-3605"]}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Hu, Junjie", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5112976412", "display_name": "Jiawei Zhao", "orcid": "https://orcid.org/0000-0002-3013-2982", "observed_orcids": ["https://orcid.org/0000-0002-3013-2982"]}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Zhao, Jiawei", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5031842648", "display_name": "Beidi Chen", "orcid": "https://orcid.org/0000-0002-7586-1855", "observed_orcids": ["https://orcid.org/0000-0002-7586-1855"]}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Chen, Beidi", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "last", "author": {"id": "https://openalex.org/A5014498545", "display_name": "Anima Anandkumar", "orcid": "https://orcid.org/0000-0002-6974-6797", "observed_orcids": ["https://orcid.org/0000-0002-6974-6797"]}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Anandkumar, Anima", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}], "abstract_inverted_index": {"Transformer-based": [0], "large": [1], "language": [2], "models": [3], "(LLMs)": [4], "demonstrate": [5, 89], "impressive": [6], "performance": [7], "in": [8], "long": [9], "context": [10, 14], "generation.": [11], "Extending": [12], "the": [13, 19, 27, 40, 48, 53, 61, 78, 104, 112, 117, 127], "length": [15], "has": [16], "disproportionately": [17], "shifted": [18], "memory": [20, 98, 114, 130, 163], "footprint": [21, 115], "of": [22, 116], "LLMs": [23], "during": [24], "inference": [25, 151], "to": [26, 43, 50, 123, 135, 143], "key-value": [28], "cache": [29, 42, 55, 76, 119], "(KV": [30], "cache).": [31], "In": [32], "this": [33], "paper,": [34], "we": [35, 88], "propose": [36], "HEADINFER,": [37], "which": [38], "offloads": [39], "KV": [41, 54, 75, 118], "CPU": [44], "RAM": [45], "while": [46, 80, 95], "avoiding": [47], "need": [49], "fully": [51], "store": [52], "for": [56], "any": [57], "transformer": [58], "layer": [59], "on": [60, 77, 103, 156], "GPU.": [62], "HEADINFER": [63, 91, 102, 148], "employs": [64], "a": [65, 108, 139, 157], "fine-grained,": [66], "head-wise": [67], "offloading": [68], "strategy,": [69], "maintaining": [70], "only": [71], "selective": [72], "attention": [73, 82], "heads": [74], "GPU": [79, 113, 129, 160], "computing": [81], "output": [83], "dynamically.": [84], "Through": [85], "roofline": [86], "analysis,": [87], "that": [90], "maintains": [92], "computational": [93], "efficiency": [94], "significantly": [96], "reducing": [97, 111], "footprint.": [99], "We": [100], "evaluate": [101], "Llama-3-8B": [105], "model": [106, 155], "with": [107, 152, 161], "1-million-token": [109], "sequence,": [110], "from": [120, 132], "128": [121], "GB": [122, 125, 134], "1": [124], "and": [126], "total": [128], "usage": [131], "207": [133], "17": [136], "GB,": [137], "achieving": [138], "92%": [140], "reduction": [141], "compared": [142], "BF16": [144], "baseline": [145], "inference.": [146], "Notably,": [147], "enables": [149], "4-million-token": [150], "an": [153], "8B": [154], "single": [158], "consumer": [159], "24GB": [162], "(e.g.,": [164], "NVIDIA": [165], "RTX": [166], "4090)": [167], "without": [168], "approximation": [169], "methods.": [170]}, "primary_topic": {"id": "https://openalex.org/T10054", "display_name": "Parallel Computing and Optimization Techniques", "score": 0.45509999990463257, "subfield": {"id": "https://openalex.org/subfields/1708", "display_name": "Hardware and Architecture"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, "topics": [{"id": "https://openalex.org/T10054", "display_name": "Parallel Computing and Optimization Techniques", "score": 0.45509999990463257, "subfield": {"id": "https://openalex.org/subfields/1708", "display_name": "Hardware and Architecture"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, {"id": "https://openalex.org/T10028", "display_name": "Topic Modeling", "score": 0.24500000476837158, "subfield": {"id": "https://openalex.org/subfields/1702", "display_name": "Artificial Intelligence"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, {"id": "https://openalex.org/T10181", "display_name": "Natural Language Processing Techniques", "score": 0.16930000483989716, "subfield": {"id": "https://openalex.org/subfields/1702", "display_name": "Artificial Intelligence"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}], "primary_location": {"id": "pmh:oai:arXiv.org:2502.12574", "is_oa": true, "landing_page_url": "http://arxiv.org/abs/2502.12574", "pdf_url": "https://arxiv.org/pdf/2502.12574", "source": {"id": "https://openalex.org/S4306400194", "display_name": "arXiv (Cornell University)", "issn_l": "2331-8422", "issn": ["2331-8422"], "is_oa": true, "is_in_doaj": false, "is_core": false, "listed_in": [], "host_organization": "https://openalex.org/I205783295", "host_organization_name": "Cornell University", "host_organization_lineage": ["https://openalex.org/I205783295"], "host_organization_lineage_names": [], "type": "repository"}, "license": null, "license_id": null, "version": "submittedVersion", "is_accepted": false, "is_published": false, "raw_source_name": null, "raw_type": "text"}, "cited_by_count": 0, "is_retracted": false}, {"id": "https://openalex.org/W7114920635", "doi": "https://doi.org/10.5281/zenodo.17902977", "title": "The Redundancy Cliff: Discovering and Exploiting 50% Dispensable Compute in Transformers with a Single O(1) Spectral Gate", "publication_year": 2025, "publication_date": "2025-12-11", "type": "article", "language": null, "authorships": [{"author_position": "first", "author": {"id": null, "display_name": "Lynch, Brendan", "orcid": null, "observed_orcids": []}, "institutions": [], "countries": [], "is_corresponding": true, "raw_author_name": "Lynch, Brendan", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}], "abstract_inverted_index": {"This": [0, 74, 103], "repository": [1], "accompanies": [2], "the": [3, 39, 77, 92, 96], "paper": [4], "\"The": [5], "Redundancy": [6], "Cliff:": [7], "Discovering": [8], "and": [9, 51, 86, 107, 120], "Exploiting": [10], "50%": [11], "Dispensable": [12], "Compute": [13], "in": [14, 67, 69], "Transformers": [15], "with": [16, 64, 95], "a": [17, 26, 35, 105], "Single": [18], "O(1)": [19], "Spectral": [20], "Gate\"": [21], "(Dec": [22], "2025).": [23], "We": [24], "introduce": [25], "simple,": [27], "fast": [28], "$O(1)$": [29], "spectral": [30], "gate": [31], "$\\kappa_x$": [32, 43, 84], "derived": [33], "from": [34], "modular-residue": [36], "fingerprint": [37], "of": [38], "input,": [40], "then": [41], "apply": [42], "to": [44, 90], "dynamically": [45], "prune": [46], "attention": [47], "KV": [48], "cache": [49], "length": [50], "slice": [52], "MLP": [53], "intermediate": [54], "dimensions.": [55], "Experiments": [56], "on": [57], "TinyLlama-1.1B": [58], "(WikiText-2)": [59], "show": [60], "substantial": [61], "compute": [62], "reduction": [63], "negligible": [65], "change": [66], "perplexity": [68], "our": [70], "proof-of-concept": [71, 106], "(PoC)": [72], "runs.": [73], "release": [75], "contains": [76], "final,": [78], "tested": [79], "code": [80], "(scripts,": [81], "patching": [82], "logic,": [83], "computation,": [85], "required": [87], "dependencies)": [88], "used": [89], "run": [91, 98], "experiments,": [93], "along": [94], "original": [97], "artifacts": [99], "for": [100, 111, 116], "maximum": [101], "transparency.": [102], "is": [104, 108], "not": [109], "intended": [110], "production": [112], "deployment.": [113], "See": [114], "README.md": [115], "instructions,": [117], "environment": [118], "notes,": [119], "full": [121], "reproducibility": [122], "details.": [123]}, "primary_topic": {"id": "https://openalex.org/T10036", "display_name": "Advanced Neural Network Applications", "score": 0.7807999849319458, "subfield": {"id": "https://openalex.org/subfields/1707", "display_name": "Computer Vision and Pattern Recognition"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, "topics": [{"id": "https://openalex.org/T10036", "display_name": "Advanced Neural Network Applications", "score": 0.7807999849319458, "subfield": {"id": "https://openalex.org/subfields/1707", "display_name": "Computer Vision and Pattern Recognition"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, {"id": "https://openalex.org/T10028", "display_name": "Topic Modeling", "score": 0.13989999890327454, "subfield": {"id": "https://openalex.org/subfields/1702", "display_name": "Artificial Intelligence"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, {"id": "https://openalex.org/T10181", "display_name": "Natural Language Processing Techniques", "score": 0.036400001496076584, "subfield": {"id": "https://openalex.org/subfields/1702", "display_name": "Artificial Intelligence"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}], "primary_location": {"id": "doi:10.5281/zenodo.17902977", "is_oa": true, "landing_page_url": "https://doi.org/10.5281/zenodo.17902977", "pdf_url": null, "source": {"id": "https://openalex.org/S4306400562", "display_name": "Zenodo (CERN European Organization for Nuclear Research)", "issn_l": null, "issn": null, "is_oa": true, "is_in_doaj": false, "is_core": false, "listed_in": [], "host_organization": "https://openalex.org/I67311998", "host_organization_name": "European Organization for Nuclear Research", "host_organization_lineage": ["https://openalex.org/I67311998"], "host_organization_lineage_names": [], "type": "repository"}, "license": "cc-by", "license_id": "https://openalex.org/licenses/cc-by", "version": null, "is_accepted": false, "is_published": false, "raw_source_name": null, "raw_type": "article-journal"}, "cited_by_count": 0, "is_retracted": false}, {"id": "https://openalex.org/W4416344173", "doi": "https://doi.org/10.48550/arxiv.2511.11346", "title": "Fast and Expressive Multi-Byte Prediction with Probabilistic Circuits", "publication_year": 2025, "publication_date": "2025-11-14", "type": "preprint", "language": null, "authorships": [{"author_position": "first", "author": {"id": "https://openalex.org/A5036399769", "display_name": "Andreas Grivas", "orcid": null, "observed_orcids": []}, "institutions": [{"id": "https://openalex.org/I98677209", "display_name": "University of Edinburgh", "ror": "https://ror.org/01nrxwf90", "country_code": "GB", "type": "education", "lineage": ["https://openalex.org/I98677209"]}], "countries": ["GB"], "is_corresponding": false, "raw_author_name": "Grivas, Andreas", "raw_affiliation_strings": ["School of Informatics , University of Edinburgh , UK"], "raw_orcid": null, "affiliations": [{"raw_affiliation_string": "School of Informatics , University of Edinburgh , UK", "institution_ids": ["https://openalex.org/I98677209"]}]}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5072981743", "display_name": "Lorenzo Loconte", "orcid": null, "observed_orcids": []}, "institutions": [{"id": "https://openalex.org/I98677209", "display_name": "University of Edinburgh", "ror": "https://ror.org/01nrxwf90", "country_code": "GB", "type": "education", "lineage": ["https://openalex.org/I98677209"]}], "countries": ["GB"], "is_corresponding": false, "raw_author_name": "Loconte, Lorenzo", "raw_affiliation_strings": ["School of Informatics , University of Edinburgh , UK"], "raw_orcid": null, "affiliations": [{"raw_affiliation_string": "School of Informatics , University of Edinburgh , UK", "institution_ids": ["https://openalex.org/I98677209"]}]}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5037724713", "display_name": "Emile van Krieken", "orcid": "https://orcid.org/0000-0001-5502-4817", "observed_orcids": ["https://orcid.org/0000-0001-5502-4817"]}, "institutions": [{"id": "https://openalex.org/I865915315", "display_name": "Vrije Universiteit Amsterdam", "ror": "https://ror.org/008xxew50", "country_code": "NL", "type": "education", "lineage": ["https://openalex.org/I865915315"]}], "countries": ["NL"], "is_corresponding": false, "raw_author_name": "van Krieken, Emile", "raw_affiliation_strings": ["Vrije Universiteit Amsterdam , NL"], "raw_orcid": null, "affiliations": [{"raw_affiliation_string": "Vrije Universiteit Amsterdam , NL", "institution_ids": ["https://openalex.org/I865915315"]}]}, {"author_position": "middle", "author": {"id": null, "display_name": "Nawrot, Piotr", "orcid": null, "observed_orcids": []}, "institutions": [{"id": "https://openalex.org/I98677209", "display_name": "University of Edinburgh", "ror": "https://ror.org/01nrxwf90", "country_code": "GB", "type": "education", "lineage": ["https://openalex.org/I98677209"]}], "countries": ["GB"], "is_corresponding": false, "raw_author_name": "Nawrot, Piotr", "raw_affiliation_strings": ["School of Informatics , University of Edinburgh , UK"], "raw_orcid": null, "affiliations": [{"raw_affiliation_string": "School of Informatics , University of Edinburgh , UK", "institution_ids": ["https://openalex.org/I98677209"]}]}, {"author_position": "middle", "author": {"id": null, "display_name": "Zhao, Yu", "orcid": null, "observed_orcids": []}, "institutions": [{"id": "https://openalex.org/I98677209", "display_name": "University of Edinburgh", "ror": "https://ror.org/01nrxwf90", "country_code": "GB", "type": "education", "lineage": ["https://openalex.org/I98677209"]}], "countries": ["GB"], "is_corresponding": false, "raw_author_name": "Zhao, Yu", "raw_affiliation_strings": ["School of Informatics , University of Edinburgh , UK"], "raw_orcid": null, "affiliations": [{"raw_affiliation_string": "School of Informatics , University of Edinburgh , UK", "institution_ids": ["https://openalex.org/I98677209"]}]}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5117865491", "display_name": "Euan Wielewski", "orcid": null, "observed_orcids": []}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Wielewski, Euan", "raw_affiliation_strings": ["NatWest Group"], "raw_orcid": null, "affiliations": [{"raw_affiliation_string": "NatWest Group", "institution_ids": []}]}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5019106673", "display_name": "Pasquale Minervini", "orcid": "https://orcid.org/0000-0002-8442-602X", "observed_orcids": ["https://orcid.org/0000-0002-8442-602X"]}, "institutions": [{"id": "https://openalex.org/I98677209", "display_name": "University of Edinburgh", "ror": "https://ror.org/01nrxwf90", "country_code": "GB", "type": "education", "lineage": ["https://openalex.org/I98677209"]}], "countries": ["GB"], "is_corresponding": false, "raw_author_name": "Minervini, Pasquale", "raw_affiliation_strings": ["Miniml.AI", "School of Informatics , University of Edinburgh , UK"], "raw_orcid": null, "affiliations": [{"raw_affiliation_string": "Miniml.AI", "institution_ids": []}, {"raw_affiliation_string": "School of Informatics , University of Edinburgh , UK", "institution_ids": ["https://openalex.org/I98677209"]}]}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5014613113", "display_name": "Edoardo Maria Ponti", "orcid": "https://orcid.org/0000-0002-6308-1050", "observed_orcids": ["https://orcid.org/0000-0002-6308-1050"]}, "institutions": [{"id": "https://openalex.org/I98677209", "display_name": "University of Edinburgh", "ror": "https://ror.org/01nrxwf90", "country_code": "GB", "type": "education", "lineage": ["https://openalex.org/I98677209"]}], "countries": ["GB"], "is_corresponding": false, "raw_author_name": "Ponti, Edoardo", "raw_affiliation_strings": ["School of Informatics , University of Edinburgh , UK"], "raw_orcid": null, "affiliations": [{"raw_affiliation_string": "School of Informatics , University of Edinburgh , UK", "institution_ids": ["https://openalex.org/I98677209"]}]}, {"author_position": "last", "author": {"id": "https://openalex.org/A5069110696", "display_name": "Antonio Vergari", "orcid": "https://orcid.org/0000-0003-0036-5678", "observed_orcids": ["https://orcid.org/0000-0003-0036-5678"]}, "institutions": [{"id": "https://openalex.org/I98677209", "display_name": "University of Edinburgh", "ror": "https://ror.org/01nrxwf90", "country_code": "GB", "type": "education", "lineage": ["https://openalex.org/I98677209"]}], "countries": ["GB"], "is_corresponding": false, "raw_author_name": "Vergari, Antonio", "raw_affiliation_strings": ["School of Informatics , University of Edinburgh , UK"], "raw_orcid": null, "affiliations": [{"raw_affiliation_string": "School of Informatics , University of Edinburgh , UK", "institution_ids": ["https://openalex.org/I98677209"]}]}], "abstract_inverted_index": {"Multi-token": [0], "prediction": [1], "(MTP)": [2], "is": [3], "a": [4, 45], "prominent": [5], "strategy": [6], "to": [7, 77, 81, 144, 151], "significantly": [8], "speed": [9], "up": [10, 141], "generation": [11, 142], "in": [12, 18, 63], "large": [13], "language": [14], "models": [15, 95], "(LLMs),": [16], "especially": [17], "byte-level": [19, 116], "LLMs,": [20, 117], "which": [21], "are": [22], "tokeniser-free": [23], "but": [24], "prohibitively": [25], "slow.": [26], "However,": [27], "many": [28], "existing": [29, 115], "MTP": [30, 64, 145], "methods": [31], "either": [32], "assume": [33], "independence": [34, 147], "between": [35, 59, 167, 186], "future": [36, 87], "tokens,": [37], "sacrificing": [38], "expressiveness,": [39], "or": [40], "generate": [41], "tokens": [42, 88], "one": [43, 76], "at": [44], "time": [46], "within": [47, 65], "the": [48, 57, 66, 83, 109, 153, 156, 164, 173, 187], "window,": [49], "increasing": [50], "latency.": [51], "In": [52], "this": [53], "work,": [54], "we": [55], "investigate": [56], "trade-off": [58, 166], "expressiveness": [60, 168], "and": [61, 104, 121, 169, 182, 189], "latency": [62, 170], "framework": [67], "of": [68, 111, 155, 176], "probabilistic": [69], "circuits": [70], "(PCs).": [71], "Our": [72, 129], "framework,": [73], "MTPC,": [74, 177], "allows": [75], "explore": [78], "different": [79], "ways": [80], "encode": [82], "joint": [84], "distributions": [85], "over": [86], "by": [89, 113], "selecting": [90], "circuit": [91], "architectures,": [92], "generalising": [93], "classical": [94], "such": [96, 118, 125, 178], "as": [97, 119, 126, 179], "(hierarchical)": [98], "mixture": [99], "models,": [100, 103, 124], "hidden": [101], "Markov": [102], "tensor": [105], "networks.": [106], "We": [107, 160], "show": [108, 131], "efficacy": [110], "MTPC": [112, 138], "retrofitting": [114], "EvaByte,": [120], "byte-fied": [122], "subword": [123], "Llama3.2": [127], "3B.": [128], "experiments": [130], "that,": [132], "when": [133, 171], "combined": [134], "with": [135, 146], "speculative": [136], "decoding,": [137], "substantially": [139], "speeds": [140], "compared": [143], "assumptions,": [148], "while": [149], "guaranteeing": [150], "retain": [152], "performance": [154], "original": [157], "verifier": [158, 188], "LLM.": [159], "also": [161], "rigorously": [162], "study": [163], "optimal": [165], "exploring": [172], "possible": [174], "parameterisations": [175], "PC": [180], "architectures": [181], "partial": [183], "layer": [184], "sharing": [185], "draft": [190], "LLMs.": [191]}, "primary_topic": {"id": "https://openalex.org/T10181", "display_name": "Natural Language Processing Techniques", "score": 0.7045999765396118, "subfield": {"id": "https://openalex.org/subfields/1702", "display_name": "Artificial Intelligence"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, "topics": [{"id": "https://openalex.org/T10181", "display_name": "Natural Language Processing Techniques", "score": 0.7045999765396118, "subfield": {"id": "https://openalex.org/subfields/1702", "display_name": "Artificial Intelligence"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, {"id": "https://openalex.org/T10028", "display_name": "Topic Modeling", "score": 0.2517000138759613, "subfield": {"id": "https://openalex.org/subfields/1702", "display_name": "Artificial Intelligence"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, {"id": "https://openalex.org/T10201", "display_name": "Speech Recognition and Synthesis", "score": 0.007400000002235174, "subfield": {"id": "https://openalex.org/subfields/1702", "display_name": "Artificial Intelligence"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}], "primary_location": {"id": "pmh:oai:arXiv.org:2511.11346", "is_oa": true, "landing_page_url": "https://arxiv.org/abs/2511.11346", "pdf_url": "https://arxiv.org/pdf/2511.11346", "source": {"id": "https://openalex.org/S4306400194", "display_name": "arXiv (Cornell University)", "issn_l": "2331-8422", "issn": ["2331-8422"], "is_oa": true, "is_in_doaj": false, "is_core": false, "listed_in": [], "host_organization": "https://openalex.org/I205783295", "host_organization_name": "Cornell University", "host_organization_lineage": ["https://openalex.org/I205783295"], "host_organization_lineage_names": [], "type": "repository"}, "license": "cc-by", "license_id": "https://openalex.org/licenses/cc-by", "version": "submittedVersion", "is_accepted": false, "is_published": false, "raw_source_name": null, "raw_type": "text"}, "cited_by_count": 0, "is_retracted": false}, {"id": "https://openalex.org/W7106822619", "doi": "https://doi.org/10.48550/arxiv.2511.20340", "title": "Scaling LLM Speculative Decoding: Non-Autoregressive Forecasting in Large-Batch Scenarios", "publication_year": 2025, "publication_date": "2025-11-25", "type": "preprint", "language": null, "authorships": [{"author_position": "first", "author": {"id": null, "display_name": "Shi, Luohe", "orcid": null, "observed_orcids": []}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Shi, Luohe", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": null, "display_name": "Li, Zuchao", "orcid": null, "observed_orcids": []}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Li, Zuchao", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": null, "display_name": "Zhang, Lefei", "orcid": null, "observed_orcids": []}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Zhang, Lefei", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": null, "display_name": "Qi, Baoyuan", "orcid": null, "observed_orcids": []}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Qi, Baoyuan", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": null, "display_name": "Liu, Guoming", "orcid": null, "observed_orcids": []}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Liu, Guoming", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "last", "author": {"id": null, "display_name": "Zhao, Hai", "orcid": null, "observed_orcids": []}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Zhao, Hai", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}], "abstract_inverted_index": {"Speculative": [0], "decoding": [1, 17, 78, 181], "accelerates": [2], "LLM": [3, 198], "inference": [4, 58, 199], "by": [5], "utilizing": [6], "otherwise": [7], "idle": [8, 72], "computational": [9, 206], "resources": [10, 82], "during": [11], "memory-to-chip": [12], "data": [13], "transfer.": [14], "Current": [15], "speculative": [16, 65, 77, 180], "methods": [18, 48], "typically": [19], "assume": [20], "a": [21, 30, 37, 61, 129, 193], "considerable": [22], "amount": [23], "of": [24, 116, 123, 158, 185], "available": [25, 71], "computing": [26, 73], "power,": [27], "then": [28], "generate": [29, 121], "complex": [31], "and": [32, 83, 135, 170, 204], "massive": [33], "draft": [34, 105, 117], "tree": [35], "using": [36], "small": [38], "autoregressive": [39, 142], "language": [40], "model": [41, 57], "to": [42, 64, 119, 145], "improve": [43], "overall": [44], "prediction": [45], "accuracy.": [46], "However,": [47], "like": [49], "batching": [50], "have": [51], "been": [52], "widely": [53], "applied": [54], "in": [55, 175], "mainstream": [56], "systems": [59], "as": [60, 67], "superior": [62], "alternative": [63], "decoding,": [66], "they": [68], "compress": [69], "the": [70, 113, 141, 149, 154, 164], "power.": [74], "Therefore,": [75], "performing": [76], "with": [79, 153, 200], "low": [80, 84], "verification": [81], "scheduling": [85], "costs": [86], "has": [87], "become": [88], "an": [89], "important": [90], "research": [91], "problem.": [92], "We": [93], "believe": [94], "that": [95, 99, 132, 190], "more": [96], "capable": [97], "models": [98, 118, 184], "allow": [100], "for": [101, 196], "parallel": [102, 155], "generation": [103, 156], "on": [104, 166], "sequences": [106, 122], "are": [107], "what": [108], "we": [109, 126, 188], "truly": [110], "need.": [111], "Recognizing": [112], "fundamental": [114], "nature": [115], "only": [120], "limited": [124], "length,": [125], "propose": [127], "SpecFormer,": [128], "novel": [130], "architecture": [131], "integrates": [133], "unidirectional": [134], "bidirectional": [136], "attention": [137], "mechanisms.": [138], "SpecFormer": [139, 191], "combines": [140], "model's": [143], "ability": [144], "extract": [146], "information": [147], "from": [148], "entire": [150], "input": [151], "sequence": [152], "benefits": [157], "non-autoregressive": [159], "models.": [160], "This": [161], "design": [162], "eliminates": [163], "reliance": [165], "large": [167], "prefix": [168], "trees": [169], "achieves": [171], "consistent": [172], "acceleration,": [173], "even": [174], "large-batch": [176], "scenarios.": [177], "Through": [178], "lossless": [179], "experiments": [182], "across": [183], "various": [186], "scales,": [187], "demonstrate": [189], "sets": [192], "new": [194], "standard": [195], "scaling": [197], "lower": [201], "training": [202], "demands": [203], "reduced": [205], "costs.": [207]}, "primary_topic": {"id": "https://openalex.org/T10181", "display_name": "Natural Language Processing Techniques", "score": 0.8080999851226807, "subfield": {"id": "https://openalex.org/subfields/1702", "display_name": "Artificial Intelligence"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, "topics": [{"id": "https://openalex.org/T10181", "display_name": "Natural Language Processing Techniques", "score": 0.8080999851226807, "subfield": {"id": "https://openalex.org/subfields/1702", "display_name": "Artificial Intelligence"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, {"id": "https://openalex.org/T10028", "display_name": "Topic Modeling", "score": 0.16050000488758087, "subfield": {"id": "https://openalex.org/subfields/1702", "display_name": "Artificial Intelligence"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, {"id": "https://openalex.org/T10201", "display_name": "Speech Recognition and Synthesis", "score": 0.005100000184029341, "subfield": {"id": "https://openalex.org/subfields/1702", "display_name": "Artificial Intelligence"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}], "primary_location": {"id": "doi:10.48550/arxiv.2511.20340", "is_oa": true, "landing_page_url": "https://doi.org/10.48550/arxiv.2511.20340", "pdf_url": null, "source": {"id": "https://openalex.org/S4306400194", "display_name": "arXiv (Cornell University)", "issn_l": "2331-8422", "issn": ["2331-8422"], "is_oa": true, "is_in_doaj": false, "is_core": false, "listed_in": [], "host_organization": "https://openalex.org/I205783295", "host_organization_name": "Cornell University", "host_organization_lineage": ["https://openalex.org/I205783295"], "host_organization_lineage_names": [], "type": "repository"}, "license": "cc-by", "license_id": "https://openalex.org/licenses/cc-by", "version": null, "is_accepted": false, "is_published": false, "raw_source_name": null, "raw_type": "Preprint"}, "cited_by_count": 0, "is_retracted": false}, {"id": "https://openalex.org/W4415272427", "doi": "https://doi.org/10.48550/arxiv.2510.11496", "title": "AndesVL Technical Report: An Efficient Mobile-side Multimodal Large Language Model", "publication_year": 2025, "publication_date": "2025-10-13", "type": "preprint", "language": "en", "authorships": [{"author_position": "first", "author": {"id": null, "display_name": "Jin, Zhiwei", "orcid": null, "observed_orcids": []}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Jin, Zhiwei", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5048406411", "display_name": "Xiaohui Song", "orcid": "https://orcid.org/0009-0000-5521-9761", "observed_orcids": ["https://orcid.org/0009-0000-5521-9761"]}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Song, Xiaohui", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5100332708", "display_name": "Nan Wang", "orcid": "https://orcid.org/0000-0001-6958-8904", "observed_orcids": ["https://orcid.org/0000-0001-6958-8904", "https://orcid.org/0000-0001-5601-7838"]}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Wang, Nan", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5100684704", "display_name": "Yafei Liu", "orcid": "https://orcid.org/0000-0001-5332-0348", "observed_orcids": ["https://orcid.org/0000-0001-5332-0348"]}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Liu, Yafei", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5004641821", "display_name": "Chao Li", "orcid": "https://orcid.org/0000-0002-4086-575X", "observed_orcids": ["https://orcid.org/0000-0002-4086-575X"]}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Li, Chao", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5100353701", "display_name": "Xin Li", "orcid": "https://orcid.org/0000-0001-6003-7719", "observed_orcids": ["https://orcid.org/0000-0001-6003-7719"]}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Li, Xin", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5102735444", "display_name": "Ruichen Wang", "orcid": "https://orcid.org/0009-0008-2485-5683", "observed_orcids": ["https://orcid.org/0009-0008-2485-5683"]}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Wang, Ruichen", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5044651184", "display_name": "Zhihao Li", "orcid": "https://orcid.org/0009-0005-9226-5106", "observed_orcids": ["https://orcid.org/0009-0005-9226-5106"]}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Li, Zhihao", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5104153900", "display_name": "Qi Qi", "orcid": "https://orcid.org/0009-0008-4606-9763", "observed_orcids": ["https://orcid.org/0009-0008-4606-9763"]}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Qi, Qi", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5108401200", "display_name": "Long Cheng", "orcid": "https://orcid.org/0000-0001-7565-8788", "observed_orcids": ["https://orcid.org/0000-0001-7565-8788"]}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Cheng, Long", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5019338727", "display_name": "Dongze Hao", "orcid": "https://orcid.org/0000-0001-5504-1799", "observed_orcids": ["https://orcid.org/0000-0001-5504-1799"]}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Hao, Dongze", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5038810646", "display_name": "Quanlong Zheng", "orcid": "https://orcid.org/0000-0001-5059-0078", "observed_orcids": ["https://orcid.org/0000-0001-5059-0078"]}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Zheng, Quanlong", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5101930309", "display_name": "Yanhao Zhang", "orcid": "https://orcid.org/0000-0001-8612-989X", "observed_orcids": ["https://orcid.org/0000-0001-8612-989X"]}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Zhang, Yanhao", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": null, "display_name": "Ji, Haobo", "orcid": null, "observed_orcids": []}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Ji, Haobo", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5100626893", "display_name": "Jian Ma", "orcid": "https://orcid.org/0000-0002-4202-5834", "observed_orcids": ["https://orcid.org/0000-0002-4202-5834"]}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Ma, Jian", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5032345018", "display_name": "Zhitong Zheng", "orcid": "https://orcid.org/0000-0002-7927-1430", "observed_orcids": ["https://orcid.org/0000-0002-7927-1430"]}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Zheng, Zhitong", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5060851339", "display_name": "Zhenyi Lin", "orcid": "https://orcid.org/0000-0002-0237-2727", "observed_orcids": ["https://orcid.org/0000-0002-0237-2727"]}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Lin, Zhenyi", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": null, "display_name": "Deng, Haolin", "orcid": null, "observed_orcids": []}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Deng, Haolin", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": null, "display_name": "Zou, Xin", "orcid": null, "observed_orcids": []}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Zou, Xin", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5017911892", "display_name": "Xiaojie Yin", "orcid": null, "observed_orcids": []}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Yin, Xiaojie", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": null, "display_name": "Wang, Ruilin", "orcid": null, "observed_orcids": []}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Wang, Ruilin", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": null, "display_name": "Cai, Liankai", "orcid": null, "observed_orcids": []}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Cai, Liankai", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5074093223", "display_name": "Haijing Liu", "orcid": "https://orcid.org/0009-0003-7708-8055", "observed_orcids": ["https://orcid.org/0009-0003-7708-8055", "https://orcid.org/0000-0002-4209-8160"]}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Liu, Haijing", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5101570278", "display_name": "Yuqing Qiu", "orcid": "https://orcid.org/0000-0002-5126-8782", "observed_orcids": ["https://orcid.org/0000-0002-5126-8782"]}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Qiu, Yuqing", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5068097128", "display_name": "Ke Chen", "orcid": "https://orcid.org/0000-0003-3918-3820", "observed_orcids": ["https://orcid.org/0000-0003-3918-3820", "https://orcid.org/0000-0002-0495-2549"]}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Chen, Ke", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5047120802", "display_name": "Zheng Li", "orcid": "https://orcid.org/0000-0003-2383-7607", "observed_orcids": ["https://orcid.org/0000-0003-2383-7607"]}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Li, Zixian", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5063247720", "display_name": "Chi Xie", "orcid": "https://orcid.org/0000-0003-3862-0224", "observed_orcids": ["https://orcid.org/0000-0003-3862-0224"]}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Xie, Chi", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5005425565", "display_name": "Huafei Li", "orcid": null, "observed_orcids": []}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Li, Huafei", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5050682332", "display_name": "Chenxing Li", "orcid": "https://orcid.org/0000-0002-3997-8212", "observed_orcids": ["https://orcid.org/0000-0002-3997-8212"]}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Li, Chenxing", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5135997517", "display_name": "Chuangchuang Wang", "orcid": null, "observed_orcids": []}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Wang, Chuangchuang", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5067743569", "display_name": "Kai Tang", "orcid": "https://orcid.org/0000-0002-5184-2086", "observed_orcids": ["https://orcid.org/0000-0002-5184-2086", "https://orcid.org/0009-0003-4765-7931", "https://orcid.org/0000-0003-3140-1731", "https://orcid.org/0009-0005-9493-0229"]}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Tang, Kai", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": null, "display_name": "Zhu, Zhiguang", "orcid": null, "observed_orcids": []}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Zhu, Zhiguang", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5067743569", "display_name": "Kai Tang", "orcid": "https://orcid.org/0000-0002-5184-2086", "observed_orcids": ["https://orcid.org/0000-0002-5184-2086", "https://orcid.org/0009-0003-4765-7931", "https://orcid.org/0000-0003-3140-1731", "https://orcid.org/0009-0005-9493-0229"]}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Tang, Kai", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": null, "display_name": "Gao, Wenmei", "orcid": null, "observed_orcids": []}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Gao, Wenmei", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5100431163", "display_name": "Rui Wang", "orcid": "https://orcid.org/0000-0001-9048-2979", "observed_orcids": ["https://orcid.org/0000-0001-9048-2979", "https://orcid.org/0000-0002-3305-2881", "https://orcid.org/0000-0002-7974-9510", "https://orcid.org/0009-0006-4029-3794", "https://orcid.org/0009-0000-8323-4184"]}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Wang, Rui", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5008305704", "display_name": "Jun Wu", "orcid": "https://orcid.org/0000-0003-4237-1806", "observed_orcids": ["https://orcid.org/0000-0003-4237-1806"]}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Wu, Jun", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5100755357", "display_name": "Chao Liu", "orcid": "https://orcid.org/0000-0001-7261-3832", "observed_orcids": ["https://orcid.org/0000-0001-7261-3832", "https://orcid.org/0000-0003-2347-7074"]}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Liu, Chao", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": null, "display_name": "Xie, Qin", "orcid": null, "observed_orcids": []}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Xie, Qin", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5100418568", "display_name": "Chen Chen", "orcid": "https://orcid.org/0000-0003-3957-7061", "observed_orcids": ["https://orcid.org/0000-0003-3957-7061"]}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Chen, Chen", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "last", "author": {"id": "https://openalex.org/A5125728711", "display_name": "Haonan Lu", "orcid": "https://orcid.org/0000-0001-6332-2785", "observed_orcids": ["https://orcid.org/0000-0001-6332-2785"]}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Lu, Haonan", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}], "abstract_inverted_index": {"In": [0], "recent": [1], "years,": [2], "while": [3], "cloud-based": [4], "MLLMs": [5, 56], "such": [6, 44, 96], "as": [7, 45, 97], "QwenVL,": [8], "InternVL,": [9], "GPT-4o,": [10], "Gemini,": [11], "and": [12, 38, 66, 78, 102, 112, 143, 165, 181], "Claude": [13], "Sonnet": [14], "have": [15], "demonstrated": [16], "outstanding": [17], "performance": [18, 86], "with": [19, 57, 117, 161], "enormous": [20], "model": [21, 74, 144], "sizes": [22], "reaching": [23], "hundreds": [24], "of": [25, 27, 41, 54, 81, 91, 120, 149], "billions": [26], "parameters,": [28], "they": [29], "significantly": [30], "surpass": [31], "the": [32, 73], "limitations": [33], "in": [34], "memory,": [35], "power": [36], "consumption,": [37], "computing": [39], "capacity": [40], "edge": [42], "devices": [43], "mobile": [46], "phones.": [47], "This": [48], "paper": [49], "introduces": [50], "AndesVL,": [51, 82], "a": [52, 88, 121, 127, 132, 170], "suite": [53], "mobile-side": [55, 147], "0.6B": [58], "to": [59, 138, 177], "4B": [60], "parameters": [61], "based": [62], "on": [63, 187, 196], "Qwen3's": [64], "LLM": [65], "various": [67], "visual": [68], "encoders.": [69], "We": [70, 192], "comprehensively": [71], "outline": [72], "architectures,": [75], "training": [76, 79], "pipeline,": [77], "data": [80], "which": [83], "achieves": [84], "first-tier": [85], "across": [87], "wide": [89], "range": [90], "open-source": [92], "benchmarks,": [93], "including": [94], "fields": [95], "text-rich": [98], "image": [99], "understanding,": [100, 111], "reasoning": [101], "math,": [103], "multi-image": [104], "comprehension,": [105], "general": [106], "VQA,": [107], "hallucination": [108], "mitigation,": [109], "multilingual": [110], "GUI-related": [113], "tasks": [114], "when": [115, 184], "compared": [116], "state-of-the-art": [118], "models": [119, 195], "similar": [122], "scale.": [123], "Furthermore,": [124], "we": [125, 168], "introduce": [126], "1+N": [128], "LoRA": [129, 134], "architecture": [130], "alongside": [131], "Quantization-Aware": [133], "Fine-Tuning": [135], "(QALFT)": [136], "framework": [137], "facilitate": [139], "efficient": [140], "task": [141], "adaptation": [142], "compression": [145, 166], "during": [146], "deployment": [148], "AndesVL.": [150], "Moreover,": [151], "utilizing": [152], "our": [153], "cache": [154], "eviction": [155], "algorithm": [156], "--": [157, 159], "OKV": [158], "along": [160], "customized": [162], "speculative": [163], "decoding": [164, 173], "strategies,": [167], "achieve": [169], "6.7x": [171], "peak": [172], "speedup": [174], "ratio,": [175], "up": [176], "30.9%": [178], "memory": [179], "reduction,": [180], "1.8": [182], "bits-per-weight": [183], "deploying": [185], "AndesVL-4B": [186], "MediaTek": [188], "Dimensity": [189], "9500": [190], "chips.": [191], "release": [193], "all": [194], "https://huggingface.co/OPPOer.": [197]}, "primary_topic": {"id": "https://openalex.org/T11714", "display_name": "Multimodal Machine Learning Applications", "score": 0.9922000169754028, "subfield": {"id": "https://openalex.org/subfields/1707", "display_name": "Computer Vision and Pattern Recognition"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, "topics": [{"id": "https://openalex.org/T11714", "display_name": "Multimodal Machine Learning Applications", "score": 0.9922000169754028, "subfield": {"id": "https://openalex.org/subfields/1707", "display_name": "Computer Vision and Pattern Recognition"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, {"id": "https://openalex.org/T10036", "display_name": "Advanced Neural Network Applications", "score": 0.0066999997943639755, "subfield": {"id": "https://openalex.org/subfields/1707", "display_name": "Computer Vision and Pattern Recognition"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, {"id": "https://openalex.org/T11307", "display_name": "Domain Adaptation and Few-Shot Learning", "score": 0.00019999999494757503, "subfield": {"id": "https://openalex.org/subfields/1702", "display_name": "Artificial Intelligence"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}], "primary_location": {"id": "pmh:oai:arXiv.org:2510.11496", "is_oa": true, "landing_page_url": "http://arxiv.org/abs/2510.11496", "pdf_url": "https://arxiv.org/pdf/2510.11496", "source": {"id": "https://openalex.org/S4306400194", "display_name": "arXiv (Cornell University)", "issn_l": "2331-8422", "issn": ["2331-8422"], "is_oa": true, "is_in_doaj": false, "is_core": false, "listed_in": [], "host_organization": "https://openalex.org/I205783295", "host_organization_name": "Cornell University", "host_organization_lineage": ["https://openalex.org/I205783295"], "host_organization_lineage_names": [], "type": "repository"}, "license": null, "license_id": null, "version": "submittedVersion", "is_accepted": false, "is_published": false, "raw_source_name": null, "raw_type": "text"}, "cited_by_count": 0, "is_retracted": false}, {"id": "https://openalex.org/W4417453506", "doi": "https://doi.org/10.48550/arxiv.2512.13019", "title": "SneakPeek: Future-Guided Instructional Streaming Video Generation", "publication_year": 2025, "publication_date": "2025-12-15", "type": "preprint", "language": null, "authorships": [{"author_position": "first", "author": {"id": null, "display_name": "Hong, Cheeun", "orcid": null, "observed_orcids": []}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Hong, Cheeun", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5030892971", "display_name": "Germ\u00e1n Barquero", "orcid": "https://orcid.org/0000-0001-8381-3549", "observed_orcids": ["https://orcid.org/0000-0001-8381-3549"]}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Barquero, German", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5047075621", "display_name": "Fadime \u015eener", "orcid": "https://orcid.org/0000-0001-5004-6005", "observed_orcids": ["https://orcid.org/0000-0001-5004-6005"]}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Sener, Fadime", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5074117139", "display_name": "Markos Georgopoulos", "orcid": "https://orcid.org/0000-0001-5928-515X", "observed_orcids": ["https://orcid.org/0000-0001-5928-515X"]}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Georgopoulos, Markos", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5109715046", "display_name": "Edgar Sch\u00f6nfeld", "orcid": null, "observed_orcids": []}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Sch\u00f6nfeld, Edgar", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5025102675", "display_name": "Stefan Popov", "orcid": null, "observed_orcids": []}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Popov, Stefan", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5031074685", "display_name": "Yuming Du", "orcid": "https://orcid.org/0009-0001-2132-8027", "observed_orcids": ["https://orcid.org/0009-0001-2132-8027"]}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Du, Yuming", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5010595613", "display_name": "Oscar Ma\u00f1as", "orcid": "https://orcid.org/0000-0001-5512-7664", "observed_orcids": ["https://orcid.org/0000-0001-5512-7664"]}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Ma\u00f1as, Oscar", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "last", "author": {"id": "https://openalex.org/A5023922146", "display_name": "Albert Pumarola", "orcid": null, "observed_orcids": []}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Pumarola, Albert", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}], "abstract_inverted_index": {"Instructional": [0], "video": [1, 33, 58, 152, 162], "generation": [2, 153], "is": [3], "an": [4, 75], "emerging": [5], "task": [6, 183], "that": [7, 167, 178], "aims": [8], "to": [9, 37, 67, 88, 102, 119], "synthesize": [10], "coherent": [11, 172], "demonstrations": [12], "of": [13, 46], "procedural": [14, 135], "activities": [15], "from": [16], "textual": [17, 80], "descriptions.": [18, 184], "Such": [19], "capability": [20], "has": [21], "broad": [22], "implications": [23], "for": [24, 54], "content": [25], "creation,": [26], "education,": [27], "and": [28, 41, 78, 91, 106, 134, 149, 173], "human-AI": [29], "interaction,": [30], "yet": [31], "existing": [32], "diffusion": [34], "models": [35], "struggle": [36], "maintain": [38], "temporal": [39, 144], "consistency": [40, 90], "controllability": [42], "across": [43], "long": [44], "sequences": [45], "multiple": [47], "action": [48], "steps.": [49], "We": [50], "introduce": [51], "a": [52, 62, 98, 114], "pipeline": [53], "future-driven": [55], "streaming": [56, 161], "instructional": [57, 71, 176], "generation,": [59], "dubbed": [60], "SneakPeek,": [61], "diffusion-based": [63], "autoregressive": [64], "framework": [65], "designed": [66], "generate": [68], "precise,": [69], "stepwise": [70], "videos": [72, 177], "conditioned": [73], "on": [74], "initial": [76], "image": [77], "structured": [79], "prompts.": [81], "Our": [82], "approach": [83], "introduces": [84], "three": [85], "key": [86], "innovations": [87], "enhance": [89], "controllability:": [92], "(1)": [93], "predictive": [94], "causal": [95, 99], "adaptation,": [96], "where": [97, 154], "model": [100], "learns": [101], "perform": [103], "next-frame": [104], "prediction": [105], "anticipate": [107], "future": [108, 155], "keyframes;": [109], "(2)": [110], "future-guided": [111], "self-forcing": [112], "with": [113], "dual-region": [115], "KV": [116], "caching": [117], "scheme": [118], "address": [120], "the": [121], "exposure": [122], "bias": [123], "issue": [124], "at": [125], "inference": [126], "time;": [127], "(3)": [128], "multi-prompt": [129], "conditioning,": [130], "which": [131], "provides": [132], "fine-grained": [133], "control": [136], "over": [137], "multi-step": [138, 182], "instructions.": [139], "Together,": [140], "these": [141], "components": [142], "mitigate": [143], "drift,": [145], "preserve": [146], "motion": [147], "consistency,": [148], "enable": [150], "interactive": [151], "prompt": [156], "updates": [157], "dynamically": [158], "influence": [159], "ongoing": [160], "generation.": [163], "Experimental": [164], "results": [165], "demonstrate": [166], "our": [168], "method": [169], "produces": [170], "temporally": [171], "semantically": [174], "faithful": [175], "accurately": [179], "follow": [180], "complex,": [181]}, "primary_topic": {"id": "https://openalex.org/T10775", "display_name": "Generative Adversarial Networks and Image Synthesis", "score": 0.9994999766349792, "subfield": {"id": "https://openalex.org/subfields/1707", "display_name": "Computer Vision and Pattern Recognition"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, "topics": [{"id": "https://openalex.org/T10775", "display_name": "Generative Adversarial Networks and Image Synthesis", "score": 0.9994999766349792, "subfield": {"id": "https://openalex.org/subfields/1707", "display_name": "Computer Vision and Pattern Recognition"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, {"id": "https://openalex.org/T11714", "display_name": "Multimodal Machine Learning Applications", "score": 0.0003000000142492354, "subfield": {"id": "https://openalex.org/subfields/1707", "display_name": "Computer Vision and Pattern Recognition"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, {"id": "https://openalex.org/T14254", "display_name": "Digital Media and Visual Art", "score": 0.0, "subfield": {"id": "https://openalex.org/subfields/1704", "display_name": "Computer Graphics and Computer-Aided Design"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}], "primary_location": {"id": "pmh:oai:arXiv.org:2512.13019", "is_oa": true, "landing_page_url": "http://arxiv.org/abs/2512.13019", "pdf_url": "https://arxiv.org/pdf/2512.13019", "source": {"id": "https://openalex.org/S4306400194", "display_name": "arXiv (Cornell University)", "issn_l": "2331-8422", "issn": ["2331-8422"], "is_oa": true, "is_in_doaj": false, "is_core": false, "listed_in": [], "host_organization": "https://openalex.org/I205783295", "host_organization_name": "Cornell University", "host_organization_lineage": ["https://openalex.org/I205783295"], "host_organization_lineage_names": [], "type": "repository"}, "license": null, "license_id": null, "version": "submittedVersion", "is_accepted": false, "is_published": false, "raw_source_name": null, "raw_type": "text"}, "cited_by_count": 0, "is_retracted": false}, {"id": "https://openalex.org/W7215454111", "doi": null, "title": "OPTIMIZING ATTENTION AND INFERENCE IN LARGE LANGUAGE MODELS: BALANCING EFFICIENCY, INTERPRETABILITY, AND ENERGY CONSUMPTION", "publication_year": 2025, "publication_date": "2025-11-27", "type": "article", "language": "en", "authorships": [{"author_position": "first", "author": {"id": "https://openalex.org/A5154235118", "display_name": "Franz Schiller", "orcid": null, "observed_orcids": []}, "institutions": [], "countries": [], "is_corresponding": true, "raw_author_name": "Franz Schiller", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}], "abstract_inverted_index": {"The": [0], "rapid": [1], "growth": [2], "of": [3, 18, 40, 51, 89, 115, 179, 202], "large": [4], "language": [5], "models": [6, 44], "(LLMs)": [7], "has": [8], "intensified": [9], "interest": [10], "in": [11, 42, 120], "the": [12, 37, 46], "computational,": [13], "energetic,": [14], "and": [15, 21, 29, 49, 53, 56, 62, 70, 92, 108, 123, 147, 167, 175, 184, 193, 205, 219, 229], "interpretive": [16], "properties": [17], "attention": [19, 41, 85, 96, 116, 134, 203], "mechanisms": [20, 204], "supporting": [22], "inference": [23, 64, 169, 206], "infrastructure.": [24], "This": [25], "article": [26], "synthesizes": [27], "theoretical": [28], "empirical": [30], "insights": [31], "across": [32], "two": [33], "intertwined": [34], "research": [35, 190, 228], "streams:": [36], "internal": [38], "mechanics": [39], "transformer": [43], "(comprehending": [45], "functional": [47], "role": [48], "interpretability": [50], "multi-head": [52, 95], "sparse": [54], "attention)": [55], "systems-level": [57], "approaches": [58], "to": [59, 103], "efficient,": [60], "low-latency,": [61], "energy-aware": [63], "(including": [65], "KV": [66, 144], "caches,": [67], "heavy-hitter": [68, 118, 152], "techniques,": [69], "firmware-level": [71], "scheduling).": [72], "We": [73, 154], "present": [74], "a": [75, 187], "cohesive": [76], "conceptual": [77], "framework": [78], "that": [79, 136, 150, 171], "reconciles": [80], "apparent": [81], "tensions\u2014such": [82], "as": [83], "whether": [84, 93], "weights": [86], "constitute": [87], "explanations": [88], "model": [90, 213], "behavior": [91], "dense": [94], "is": [97], "uniformly": [98], "necessary\u2014by": [99], "connecting": [100], "representational": [101], "redundancy": [102], "opportunities": [104], "for": [105, 129, 158, 189, 225], "structured": [106], "sparsity": [107], "cache-aware": [109], "inference.": [110], "Building": [111], "on": [112, 162], "prior": [113], "analyses": [114], "distribution,": [117], "phenomena": [119], "token": [121], "streams,": [122], "lifecycle": [124], "energy": [125, 176], "accounting,": [126], "we": [127, 185], "argue": [128], "an": [130], "integrative": [131], "approach:": [132], "adaptive": [133], "architectures": [135], "dynamically": [137], "reallocate": [138], "head": [139], "resources,": [140], "combined": [141], "with": [142], "inference-time": [143], "cache": [145], "management": [146], "scheduling": [148], "policies": [149], "prioritize": [151], "contexts.": [153], "discuss": [155], "methodological": [156], "principles": [157], "evaluating": [159], "such": [160], "architectures\u2014focusing": [161], "causal": [163], "probing,": [164], "ablation": [165], "procedures,": [166], "realistic": [168], "benchmarks": [170], "capture": [172], "latency,": [173], "throughput,": [174], "budgets.": [177], "Limitations": [178], "extant": [180], "studies": [181], "are": [182], "detailed,": [183], "outline": [186], "roadmap": [188], "blending": [191], "model-centric": [192], "systems-centric": [194], "innovation.": [195], "Our": [196], "synthesis": [197], "highlights": [198], "how": [199], "careful": [200], "co-design": [201], "systems": [207], "can": [208], "preserve": [209], "or": [210], "even": [211], "enhance": [212], "fidelity": [214], "while": [215], "substantially": [216], "reducing": [217], "computational": [218], "environmental": [220], "cost,": [221], "offering": [222], "concrete": [223], "directions": [224], "both": [226], "algorithmic": [227], "practical": [230], "deployment.": [231]}, "primary_topic": {"id": "https://openalex.org/T10028", "display_name": "Topic Modeling", "score": 0.5737000107765198, "subfield": {"id": "https://openalex.org/subfields/1702", "display_name": "Artificial Intelligence"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, "topics": [{"id": "https://openalex.org/T10028", "display_name": "Topic Modeling", "score": 0.5737000107765198, "subfield": {"id": "https://openalex.org/subfields/1702", "display_name": "Artificial Intelligence"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, {"id": "https://openalex.org/T10181", "display_name": "Natural Language Processing Techniques", "score": 0.16189999878406525, "subfield": {"id": "https://openalex.org/subfields/1702", "display_name": "Artificial Intelligence"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, {"id": "https://openalex.org/T10036", "display_name": "Advanced Neural Network Applications", "score": 0.15379999577999115, "subfield": {"id": "https://openalex.org/subfields/1707", "display_name": "Computer Vision and Pattern Recognition"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}], "primary_location": {"id": "pmh:oai:ojs2.www.ijmrd.in:article/4103", "is_oa": true, "landing_page_url": "https://ijmrd.in/index.php/imjrd/article/view/4103/5109", "pdf_url": null, "source": {"id": "https://openalex.org/S7407065807", "display_name": "International Multidisciplinary Journal for Research & Development", "issn_l": "2394-6334", "issn": ["2394-6334"], "is_oa": true, "is_in_doaj": false, "is_core": false, "listed_in": [], "host_organization": null, "host_organization_name": null, "host_organization_lineage": [], "host_organization_lineage_names": [], "type": "journal"}, "license": "cc-by-nc", "license_id": "https://openalex.org/licenses/cc-by-nc", "version": "publishedVersion", "is_accepted": true, "is_published": true, "raw_source_name": "2394-6334", "raw_type": "info:eu-repo/semantics/article"}, "cited_by_count": 0, "is_retracted": false}, {"id": "https://openalex.org/W7115591864", "doi": "https://doi.org/10.48550/arxiv.2512.11550", "title": "PD-Swap: Prefill-Decode Logic Swapping for End-to-End LLM Inference on Edge FPGAs via Dynamic Partial Reconfiguration", "publication_year": 2025, "publication_date": "2025-12-12", "type": "preprint", "language": null, "authorships": [{"author_position": "first", "author": {"id": null, "display_name": "Zhang, Yifan", "orcid": null, "observed_orcids": []}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Zhang, Yifan", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": null, "display_name": "Chen, Zhiheng", "orcid": null, "observed_orcids": []}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Chen, Zhiheng", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": null, "display_name": "Qiao, Ye", "orcid": null, "observed_orcids": []}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Qiao, Ye", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "last", "author": {"id": null, "display_name": "Huang, Sitao", "orcid": null, "observed_orcids": []}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Huang, Sitao", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}], "abstract_inverted_index": {"Aggressively": [0], "quantized": [1], "large": [2], "language": [3], "models": [4], "(LLMs),": [5], "such": [6], "as": [7, 26], "BitNet-style": [8], "1.58-bit": [9], "Transformers": [10], "with": [11, 40, 164], "ternary": [12, 148], "weights,": [13], "make": [14], "it": [15], "feasible": [16], "to": [17, 29, 44, 106, 137, 207], "deploy": [18], "generative": [19], "AI": [20], "on": [21, 67, 142], "low-power": [22], "edge": [23, 35, 143], "FPGAs.": [24, 144], "However,": [25], "prompts": [27], "grow": [28], "tens": [30], "of": [31, 33, 57], "thousands": [32], "tokens,": [34], "hardware": [36], "performance": [37], "drops": [38], "sharply": [39], "sequence": [41], "length": [42, 60], "due": [43], "quadratic": [45], "prefill": [46, 74, 171], "cost": [47], "and": [48, 77, 87, 98, 112, 120, 151, 173, 182, 193, 196], "rapidly": [49], "increasing": [50], "KV-cache": [51, 90], "bandwidth": [52], "demands,": [53], "making": [54], "inference": [55], "latency": [56, 198], "longer": [58, 220], "context": [59, 221], "a": [61, 70, 99, 125, 161, 168, 174], "first-order": [62], "system": [63], "concern.": [64], "Recent": [65], "studies": [66], "LLMs": [68], "expose": [69], "fundamental": [71], "prefill-decode": [72], "asymmetry:": [73], "is": [75, 85, 160, 199], "compute-bound": [76], "dominated": [78, 88], "by": [79, 89, 201, 215], "dense": [80], "matrix-matrix": [81], "operations,": [82], "whereas": [83], "decoding": [84, 177, 209], "memory-bandwidth-bound": [86], "traffic.": [91], "A": [92, 179], "static": [93], "accelerator": [94], "must": [95], "provision": [96], "resources": [97], "single": [100], "dataflow": [101], "for": [102], "both": [103], "regimes,": [104], "leading": [105], "duplicated": [107], "attention": [108, 140, 158], "logic,": [109], "underutilized": [110], "fabric,": [111], "tight": [113], "LUT/URAM": [114], "limits": [115], "that": [116, 131], "cap": [117], "model": [118, 181], "size": [119], "usable": [121], "context.": [122], "We": [123], "propose": [124], "prefill--decode": [126], "disaggregated": [127], "LLM": [128], "accelerator,": [129], "PD-Swap,": [130], "uses": [132], "Dynamic": [133], "Partial": [134], "Reconfiguration": [135], "(DPR)": [136], "time-multiplex": [138], "the": [139, 157], "module": [141], "The": [145], "core": [146], "table-lookup": [147], "matrix": [149], "multiplication": [150], "weight-buffering": [152], "engines": [153], "remain": [154], "static,": [155], "while": [156], "subsystem": [159], "reconfigurable": [162], "partition": [163], "two": [165], "phase-specialized": [166], "architectures:": [167], "compute-heavy,": [169], "token-parallel": [170], "engine": [172], "bandwidth-optimized,": [175], "KV-cache-centric": [176], "engine.": [178], "roofline-inspired": [180], "design": [183], "space": [184], "exploration": [185], "jointly": [186], "optimize": [187], "reconfigurable-region": [188], "size,": [189], "parallelism": [190], "under": [191], "reconfiguration": [192, 197], "routability": [194], "constraints,": [195], "hidden": [200], "computation": [202], "latency.": [203], "PD-Swap": [204], "achieves": [205], "up": [206], "27~tokens/s": [208], "throughput,": [210], "outperforming": [211], "prior": [212], "state-of-the-art": [213], "works": [214], "1.3x--2.1x": [216], "(larger": [217], "gains": [218], "at": [219], "lengths),": [222], "without": [223], "extra": [224], "area": [225], "cost.": [226]}, "primary_topic": {"id": "https://openalex.org/T10904", "display_name": "Embedded Systems Design Techniques", "score": 0.9794999957084656, "subfield": {"id": "https://openalex.org/subfields/1708", "display_name": "Hardware and Architecture"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, "topics": [{"id": "https://openalex.org/T10904", "display_name": "Embedded Systems Design Techniques", "score": 0.9794999957084656, "subfield": {"id": "https://openalex.org/subfields/1708", "display_name": "Hardware and Architecture"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, {"id": "https://openalex.org/T10036", "display_name": "Advanced Neural Network Applications", "score": 0.004399999976158142, "subfield": {"id": "https://openalex.org/subfields/1707", "display_name": "Computer Vision and Pattern Recognition"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, {"id": "https://openalex.org/T11522", "display_name": "VLSI and FPGA Design Techniques", "score": 0.002899999963119626, "subfield": {"id": "https://openalex.org/subfields/2208", "display_name": "Electrical and Electronic Engineering"}, "field": {"id": "https://openalex.org/fields/22", "display_name": "Engineering"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}], "primary_location": {"id": "pmh:oai:arXiv.org:2512.11550", "is_oa": true, "landing_page_url": "http://arxiv.org/abs/2512.11550", "pdf_url": "https://arxiv.org/pdf/2512.11550", "source": {"id": "https://openalex.org/S4306400194", "display_name": "arXiv (Cornell University)", "issn_l": "2331-8422", "issn": ["2331-8422"], "is_oa": true, "is_in_doaj": false, "is_core": false, "listed_in": [], "host_organization": "https://openalex.org/I205783295", "host_organization_name": "Cornell University", "host_organization_lineage": ["https://openalex.org/I205783295"], "host_organization_lineage_names": [], "type": "repository"}, "license": null, "license_id": null, "version": "submittedVersion", "is_accepted": false, "is_published": false, "raw_source_name": null, "raw_type": "text"}, "cited_by_count": 0, "is_retracted": false}, {"id": "https://openalex.org/W4415336479", "doi": "https://doi.org/10.48550/arxiv.2509.24248", "title": "SpecExit: Accelerating Large Reasoning Model via Speculative Exit", "publication_year": 2025, "publication_date": "2025-09-29", "type": "preprint", "language": "en", "authorships": [{"author_position": "first", "author": {"id": null, "display_name": "Yang, Rubing", "orcid": null, "observed_orcids": []}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Yang, Rubing", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": null, "display_name": "Bai, Huajun", "orcid": null, "observed_orcids": []}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Bai, Huajun", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5101930706", "display_name": "Song Liu", "orcid": "https://orcid.org/0000-0002-0851-3746", "observed_orcids": ["https://orcid.org/0000-0002-0851-3746"]}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Liu, Song", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5014092511", "display_name": "Guanghua Yu", "orcid": null, "observed_orcids": []}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Yu, Guanghua", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": null, "display_name": "Fan, Runzhi", "orcid": null, "observed_orcids": []}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Fan, Runzhi", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": null, "display_name": "Dang, Yanbin", "orcid": null, "observed_orcids": []}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Dang, Yanbin", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5120058099", "display_name": "Jiejing Zhang", "orcid": null, "observed_orcids": []}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Zhang, Jiejing", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5045987910", "display_name": "K.-H. Liu", "orcid": "https://orcid.org/0009-0005-2024-1818", "observed_orcids": ["https://orcid.org/0009-0005-2024-1818"]}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Liu, Kai", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5059894156", "display_name": "Jianchen Zhu", "orcid": "https://orcid.org/0000-0002-5988-3704", "observed_orcids": ["https://orcid.org/0000-0002-5988-3704"]}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Zhu, Jianchen", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "last", "author": {"id": "https://openalex.org/A5054060382", "display_name": "Peng Chen", "orcid": "https://orcid.org/0000-0003-4577-9866", "observed_orcids": ["https://orcid.org/0000-0003-4577-9866"]}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Chen, Peng", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}], "abstract_inverted_index": {"Despite": [0], "their": [1, 28, 60, 71, 77], "strong": [2], "performance": [3], "on": [4, 57, 62], "reasoning": [5, 8, 41], "tasks,": [6], "large": [7], "models": [9], "(LRMs)": [10], "often": [11], "suffer": [12], "from": [13, 108, 150], "overthinking,": [14, 33], "producing": [15], "unnecessarily": [16], "long": [17], "outputs": [18], "and": [19, 75, 103, 127], "incurring": [20], "high": [21], "end-to-end": [22, 72, 133], "latency,": [23], "a": [24, 66, 95, 109, 129], "significant": [25, 119], "limitation": [26], "to": [27, 39, 136, 153], "real-world": [29], "deployment.": [30], "To": [31], "address": [32], "early-exit": [34, 105, 156], "mechanisms": [35, 64], "have": [36], "been": [37], "proposed": [38], "terminate": [40], "before": [42], "typical": [43], "completion,": [44], "showing": [45], "that": [46, 69, 98], "this": [47], "approach": [48], "can": [49], "effectively": [50], "shorten": [51], "generation": [52, 123], "length": [53, 124], "with": [54], "minimal": [55], "impact": [56], "accuracy.": [58, 143], "However,": [59], "reliance": [61], "probing": [63, 114], "introduces": [65], "detection": [67], "overhead": [68], "limits": [70], "latency": [73, 134], "gains": [74], "compromises": [76], "generalizability": [78], "across": [79], "diverse": [80], "problems.": [81], "Inspired": [82], "by": [83, 125], "the": [84, 137, 147], "use": [85, 160], "of": [86, 161], "hidden": [87, 151, 162], "states": [88, 152, 163], "in": [89, 132], "speculative": [90, 138], "decoding,": [91], "we": [92], "propose": [93], "SpecExit,": [94], "novel": [96], "framework": [97], "predicts": [99], "both": [100], "future": [101], "tokens": [102], "an": [104], "signal": [106], "directly": [107], "lightweight": [110], "draft": [111], "model": [112], "without": [113, 141], "overhead.": [115], "Our": [116, 144, 167], "method": [117, 145], "offers": [118], "improvements,": [120], "reducing": [121], "average": [122], "66\\%": [126], "achieving": [128], "2.5x": [130], "speedup": [131], "compared": [135], "decoding": [139], "baseline,": [140], "compromising": [142], "leverages": [146], "inherent": [148], "signals": [149], "provide": [154], "effective": [155], "signals,": [157], "suggesting": [158], "broader": [159], "for": [164], "efficient": [165], "reasoning.": [166], "code": [168], "is": [169], "available": [170], "at": [171], "https://github.com/Tencent/AngelSlim.": [172]}, "primary_topic": {"id": "https://openalex.org/T10028", "display_name": "Topic Modeling", "score": 0.4595000147819519, "subfield": {"id": "https://openalex.org/subfields/1702", "display_name": "Artificial Intelligence"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, "topics": [{"id": "https://openalex.org/T10028", "display_name": "Topic Modeling", "score": 0.4595000147819519, "subfield": {"id": "https://openalex.org/subfields/1702", "display_name": "Artificial Intelligence"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, {"id": "https://openalex.org/T10181", "display_name": "Natural Language Processing Techniques", "score": 0.4487000107765198, "subfield": {"id": "https://openalex.org/subfields/1702", "display_name": "Artificial Intelligence"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, {"id": "https://openalex.org/T10036", "display_name": "Advanced Neural Network Applications", "score": 0.023399999365210533, "subfield": {"id": "https://openalex.org/subfields/1707", "display_name": "Computer Vision and Pattern Recognition"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}], "primary_location": {"id": "pmh:oai:arXiv.org:2509.24248", "is_oa": true, "landing_page_url": "http://arxiv.org/abs/2509.24248", "pdf_url": "https://arxiv.org/pdf/2509.24248", "source": {"id": "https://openalex.org/S4306400194", "display_name": "arXiv (Cornell University)", "issn_l": "2331-8422", "issn": ["2331-8422"], "is_oa": true, "is_in_doaj": false, "is_core": false, "listed_in": [], "host_organization": "https://openalex.org/I205783295", "host_organization_name": "Cornell University", "host_organization_lineage": ["https://openalex.org/I205783295"], "host_organization_lineage_names": [], "type": "repository"}, "license": null, "license_id": null, "version": "submittedVersion", "is_accepted": false, "is_published": false, "raw_source_name": null, "raw_type": "text"}, "cited_by_count": 0, "is_retracted": false}, {"id": "https://openalex.org/W4416010150", "doi": "https://doi.org/10.48550/arxiv.2505.03745", "title": "AccLLM: Accelerating Long-Context LLM Inference Via Algorithm-Hardware Co-Design", "publication_year": 2025, "publication_date": "2025-04-07", "type": "preprint", "language": "en", "authorships": [{"author_position": "first", "author": {"id": null, "display_name": "Liang, Yanbiao", "orcid": null, "observed_orcids": []}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Liang, Yanbiao", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5020476441", "display_name": "Huihong Shi", "orcid": "https://orcid.org/0000-0002-7845-0154", "observed_orcids": ["https://orcid.org/0000-0002-7845-0154"]}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Shi, Huihong", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5090170161", "display_name": "Haikuo Shao", "orcid": "https://orcid.org/0009-0008-6965-3436", "observed_orcids": ["https://orcid.org/0009-0008-6965-3436"]}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Shao, Haikuo", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "last", "author": {"id": "https://openalex.org/A5100696999", "display_name": "Zhongfeng Wang", "orcid": "https://orcid.org/0000-0002-7227-4786", "observed_orcids": ["https://orcid.org/0000-0002-7227-4786"]}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Wang, Zhongfeng", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}], "abstract_inverted_index": {"Recently,": [0], "large": [1], "language": [2, 12], "models": [3], "(LLMs)": [4], "have": [5], "achieved": [6], "huge": [7, 45], "success": [8], "in": [9], "the": [10, 25, 56, 93, 131, 160, 171, 187], "natural": [11], "processing": [13], "(NLP)": [14], "field,": [15], "driving": [16], "a": [17, 75, 136, 141, 177, 182], "growing": [18], "demand": [19], "to": [20, 27, 145, 186], "extend": [21], "their": [22], "deployment": [23], "from": [24, 153], "cloud": [26], "edge": [28, 35], "devices.": [29], "However,": [30], "deploying": [31], "LLMs": [32], "on": [33, 170], "resource-constrained": [34], "devices": [36], "poses": [37], "significant": [38], "challenges,": [39, 71], "including": [40], "(1)": [41, 98], "intensive": [42], "computations": [43], "and": [44, 51, 60, 82, 89, 103, 112, 122, 147, 181], "model": [46], "sizes,": [47], "(2)": [48, 100], "great": [49], "memory": [50, 121], "bandwidth": [52, 123], "demands": [53], "introduced": [54], "by": [55], "autoregressive": [57], "generation": [58], "process,": [59], "(3)": [61, 104], "limited": [62], "scalability": [63], "for": [64], "handling": [65], "long": [66], "sequences.": [67], "To": [68], "address": [69], "these": [70], "we": [72, 96, 134], "propose": [73], "AccLLM,": [74], "comprehensive": [76], "acceleration": [77], "framework": [78], "that": [79], "enables": [80], "efficient": [81], "fast": [83], "long-context": [84], "LLM": [85], "inference": [86], "through": [87], "algorithm": [88], "hardware": [90, 132, 165], "co-design.": [91], "At": [92, 130], "algorithmic": [94, 161], "level,": [95, 133], "integrate": [97], "pruning,": [99], "\u039b-shaped": [101], "attention,": [102], "an": [105], "innovative": [106], "W2A8KV4": [107], "(2-bit": [108], "weights,": [109], "8-bit": [110], "activations,": [111], "4-bit": [113], "KV": [114], "cache)": [115], "quantization": [116], "scheme,": [117], "thus": [118], "effectively": [119, 146], "reducing": [120], "requirements": [124], "while": [125], "facilitating": [126], "LLMs'": [127], "long-sequence": [128], "generation.": [129], "design": [135], "dedicated": [137], "FPGA-based": [138], "accelerator": [139], "with": [140], "reconfigurable": [142], "computing": [143], "engine": [144], "flexibly": [148], "accommodate": [149], "diverse": [150], "operations": [151], "arising": [152], "our": [154], "compression": [155], "algorithm,": [156], "thereby": [157], "fully": [158], "translating": [159], "innovations": [162], "into": [163], "tangible": [164], "efficiency.": [166], "We": [167], "validate": [168], "AccLLM": [169], "Xilinx": [172], "Alveo": [173], "U280": [174], "FPGA,": [175], "demonstrating": [176], "4.07x": [178], "energy": [179], "efficiency": [180], "2.98x": [183], "throughput": [184], "compared": [185], "state-of-the-art": [188], "work": [189], "FlightLLM.": [190]}, "primary_topic": {"id": "https://openalex.org/T10904", "display_name": "Embedded Systems Design Techniques", "score": 0.5019999742507935, "subfield": {"id": "https://openalex.org/subfields/1708", "display_name": "Hardware and Architecture"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, "topics": [{"id": "https://openalex.org/T10904", "display_name": "Embedded Systems Design Techniques", "score": 0.5019999742507935, "subfield": {"id": "https://openalex.org/subfields/1708", "display_name": "Hardware and Architecture"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, {"id": "https://openalex.org/T10036", "display_name": "Advanced Neural Network Applications", "score": 0.4041000008583069, "subfield": {"id": "https://openalex.org/subfields/1707", "display_name": "Computer Vision and Pattern Recognition"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, {"id": "https://openalex.org/T10054", "display_name": "Parallel Computing and Optimization Techniques", "score": 0.007499999832361937, "subfield": {"id": "https://openalex.org/subfields/1708", "display_name": "Hardware and Architecture"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}], "primary_location": {"id": "pmh:oai:arXiv.org:2505.03745", "is_oa": true, "landing_page_url": "http://arxiv.org/abs/2505.03745", "pdf_url": "https://arxiv.org/pdf/2505.03745", "source": {"id": "https://openalex.org/S4306400194", "display_name": "arXiv (Cornell University)", "issn_l": "2331-8422", "issn": ["2331-8422"], "is_oa": true, "is_in_doaj": false, "is_core": false, "listed_in": [], "host_organization": "https://openalex.org/I205783295", "host_organization_name": "Cornell University", "host_organization_lineage": ["https://openalex.org/I205783295"], "host_organization_lineage_names": [], "type": "repository"}, "license": null, "license_id": null, "version": "submittedVersion", "is_accepted": false, "is_published": false, "raw_source_name": null, "raw_type": "text"}, "cited_by_count": 0, "is_retracted": false}, {"id": "https://openalex.org/W4416591563", "doi": "https://doi.org/10.48550/arxiv.2511.16193", "title": "Fast LLM Post-training via Decoupled and Fastest-of-N Speculation", "publication_year": 2025, "publication_date": "2025-11-20", "type": "preprint", "language": null, "authorships": [{"author_position": "first", "author": {"id": null, "display_name": "Cheng, Rongxin", "orcid": null, "observed_orcids": []}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Cheng, Rongxin", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5101481364", "display_name": "Kai Zhou", "orcid": "https://orcid.org/0009-0009-5953-3008", "observed_orcids": ["https://orcid.org/0009-0009-5953-3008"]}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Zhou, Kai", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5002957327", "display_name": "Xingda Wei", "orcid": "https://orcid.org/0000-0003-4983-6047", "observed_orcids": ["https://orcid.org/0000-0003-4983-6047"]}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Wei, Xingda", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5100406626", "display_name": "Siyuan Liu", "orcid": "https://orcid.org/0000-0002-3770-2220", "observed_orcids": ["https://orcid.org/0000-0002-3770-2220"]}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Liu, Siyuan", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5030930550", "display_name": "Mingcong Han", "orcid": "https://orcid.org/0009-0008-1536-7485", "observed_orcids": ["https://orcid.org/0009-0008-1536-7485"]}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Han, Mingcong", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5046374975", "display_name": "Mingjing Ai", "orcid": null, "observed_orcids": []}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Ai, Mingjing", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5006062505", "display_name": "Yeju Zhou", "orcid": "https://orcid.org/0000-0001-8157-7474", "observed_orcids": ["https://orcid.org/0000-0001-8157-7474"]}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Zhou, Yeju", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5037661446", "display_name": "Baoquan Zhong", "orcid": "https://orcid.org/0009-0005-0833-5619", "observed_orcids": ["https://orcid.org/0009-0005-0833-5619"]}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Zhong, Baoquan", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5086945155", "display_name": "Wencong Xiao", "orcid": "https://orcid.org/0000-0002-3043-522X", "observed_orcids": ["https://orcid.org/0000-0002-3043-522X"]}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Xiao, Wencong", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5101543146", "display_name": "Rong Chen", "orcid": "https://orcid.org/0000-0002-6115-8130", "observed_orcids": ["https://orcid.org/0000-0002-6115-8130"]}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Chen, Rong", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "last", "author": {"id": "https://openalex.org/A5100406215", "display_name": "Haibo Chen", "orcid": "https://orcid.org/0000-0002-9720-0361", "observed_orcids": ["https://orcid.org/0000-0002-9720-0361"]}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Chen, Haibo", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}], "abstract_inverted_index": {"Rollout": [0], "dominates": [1], "the": [2, 12, 42, 46, 55, 58, 77, 116, 121, 127], "training": [3, 96, 192], "time": [4], "in": [5, 95, 182], "large": [6, 87], "language": [7], "model": [8, 14], "(LLM)": [9], "post-training,": [10], "where": [11], "trained": [13], "is": [15, 48, 130, 174], "used": [16], "to": [17, 40, 99, 115, 119, 150, 178], "generate": [18], "tokens": [19], "given": [20], "a": [21, 36, 71, 92, 103, 132], "batch": [22, 89], "of": [23, 54], "prompts.": [24], "This": [25], "work,": [26], "SpecActor,": [27], "achieves": [28, 188], "fast": [29, 37, 51], "rollout": [30, 117, 144, 181, 187], "with": [31, 57, 85, 148], "speculative": [32, 83, 180], "decoding": [33, 84], "that": [34, 66, 75, 107, 140], "deploys": [35], "draft": [38, 112, 123], "path": [39], "accelerate": [41], "unparallelizable": [43], "generation,": [44], "while": [45], "correctness": [47], "guaranteed": [49], "by": [50, 146], "parallel": [52], "verification": [53], "outputs": [56], "original": [59], "model.": [60], "SpecActor": [61, 141, 173], "addresses": [62], "two": [63], "foundational": [64], "challenges": [65], "hinder": [67], "speculation": [68, 73, 105], "efficiency:": [69], "(1)": [70], "Decoupled": [72], "method": [74, 106, 124], "overcomes": [76], "computation": [78], "inefficiency": [79], "issue": [80], "when": [81, 126], "executing": [82], "relative": [86], "per-worker": [88], "size": [90], "--": [91], "common": [93, 154], "configuration": [94], "but": [97], "unfriendly": [98], "speculation,": [100], "and": [101, 109, 164, 167], "(2)": [102], "Fastest-of-N": [104], "selects": [108], "combines": [110], "different": [111, 169, 183], "methods": [113], "according": [114], "progress": [118], "approximate": [120], "optimal": [122], "even": [125], "best": [128], "one": [129], "unknown": [131], "priori.": [133], "Extensive": [134], "evaluations": [135], "on": [136], "production": [137], "traces": [138], "show": [139], "accelerates": [142], "mean": [143], "speed": [145], "2.0--2.4x,": [147], "up": [149], "2.7x": [151], "speedup,": [152], "over": [153], "post-training": [155], "baselines.": [156], "The": [157, 185], "results": [158], "are": [159], "consistent": [160], "across": [161, 168], "both": [162], "dense": [163], "MoE": [165], "models": [166], "RL": [170], "algorithms.": [171], "Notably,": [172], "1.1--2.6x": [175], "faster": [176, 190], "compared": [177], "vanilla": [179], "traces.": [184], "accelerated": [186], "1.4--2.3x": [189], "end-to-end": [191], "time.": [193]}, "primary_topic": {"id": "https://openalex.org/T10181", "display_name": "Natural Language Processing Techniques", "score": 0.4657999873161316, "subfield": {"id": "https://openalex.org/subfields/1702", "display_name": "Artificial Intelligence"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, "topics": [{"id": "https://openalex.org/T10181", "display_name": "Natural Language Processing Techniques", "score": 0.4657999873161316, "subfield": {"id": "https://openalex.org/subfields/1702", "display_name": "Artificial Intelligence"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, {"id": "https://openalex.org/T10028", "display_name": "Topic Modeling", "score": 0.29100000858306885, "subfield": {"id": "https://openalex.org/subfields/1702", "display_name": "Artificial Intelligence"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, {"id": "https://openalex.org/T10054", "display_name": "Parallel Computing and Optimization Techniques", "score": 0.06019999831914902, "subfield": {"id": "https://openalex.org/subfields/1708", "display_name": "Hardware and Architecture"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}], "primary_location": {"id": "pmh:oai:arXiv.org:2511.16193", "is_oa": true, "landing_page_url": "http://arxiv.org/abs/2511.16193", "pdf_url": "https://arxiv.org/pdf/2511.16193", "source": {"id": "https://openalex.org/S4306400194", "display_name": "arXiv (Cornell University)", "issn_l": "2331-8422", "issn": ["2331-8422"], "is_oa": true, "is_in_doaj": false, "is_core": false, "listed_in": [], "host_organization": "https://openalex.org/I205783295", "host_organization_name": "Cornell University", "host_organization_lineage": ["https://openalex.org/I205783295"], "host_organization_lineage_names": [], "type": "repository"}, "license": null, "license_id": null, "version": "submittedVersion", "is_accepted": false, "is_published": false, "raw_source_name": null, "raw_type": "text"}, "cited_by_count": 0, "is_retracted": false}, {"id": "https://openalex.org/W4414727549", "doi": "https://doi.org/10.2139/ssrn.5551142", "title": "DuoServe-MoE: Dual-Phase Expert Prefetch and Cache Scheduling for Efficient MoE LLM Inference", "publication_year": 2025, "publication_date": "2025-01-01", "type": "preprint", "language": "en", "authorships": [{"author_position": "first", "author": {"id": "https://openalex.org/A5039364773", "display_name": "Yuning Zhang", "orcid": "https://orcid.org/0000-0002-0306-1859", "observed_orcids": ["https://orcid.org/0000-0002-0306-1859"]}, "institutions": [{"id": "https://openalex.org/I129604602", "display_name": "The University of Sydney", "ror": "https://ror.org/0384j8v12", "country_code": "AU", "type": "education", "lineage": ["https://openalex.org/I129604602"]}], "countries": ["AU"], "is_corresponding": false, "raw_author_name": "Yuning Zhang", "raw_affiliation_strings": ["The University of Sydney", "The University of Sydney, School of Mathematics and Statistics, Carslaw Building, Sydney, Australia"], "raw_orcid": null, "affiliations": [{"raw_affiliation_string": "The University of Sydney", "institution_ids": ["https://openalex.org/I129604602"]}, {"raw_affiliation_string": "The University of Sydney, School of Mathematics and Statistics, Carslaw Building, Sydney, Australia", "institution_ids": ["https://openalex.org/I129604602"]}]}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5119805266", "display_name": "Grant Pinkert", "orcid": null, "observed_orcids": []}, "institutions": [{"id": "https://openalex.org/I129604602", "display_name": "The University of Sydney", "ror": "https://ror.org/0384j8v12", "country_code": "AU", "type": "education", "lineage": ["https://openalex.org/I129604602"]}], "countries": ["AU"], "is_corresponding": false, "raw_author_name": "Grant Pinkert", "raw_affiliation_strings": ["The University of Sydney", "The University of Sydney, School of Mathematics and Statistics, Carslaw Building, Sydney, Australia"], "raw_orcid": null, "affiliations": [{"raw_affiliation_string": "The University of Sydney", "institution_ids": ["https://openalex.org/I129604602"]}, {"raw_affiliation_string": "The University of Sydney, School of Mathematics and Statistics, Carslaw Building, Sydney, Australia", "institution_ids": ["https://openalex.org/I129604602"]}]}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5101979235", "display_name": "Nan Yang", "orcid": "https://orcid.org/0000-0001-8116-8750", "observed_orcids": ["https://orcid.org/0000-0001-8116-8750", "https://orcid.org/0000-0002-8965-2616"]}, "institutions": [{"id": "https://openalex.org/I129604602", "display_name": "The University of Sydney", "ror": "https://ror.org/0384j8v12", "country_code": "AU", "type": "education", "lineage": ["https://openalex.org/I129604602"]}], "countries": ["AU"], "is_corresponding": false, "raw_author_name": "Nan Yang", "raw_affiliation_strings": ["The University of Sydney", "The University of Sydney, Australia"], "raw_orcid": null, "affiliations": [{"raw_affiliation_string": "The University of Sydney", "institution_ids": ["https://openalex.org/I129604602"]}, {"raw_affiliation_string": "The University of Sydney, Australia", "institution_ids": ["https://openalex.org/I129604602"]}]}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5054501138", "display_name": "Y H Li", "orcid": "https://orcid.org/0009-0004-1986-8982", "observed_orcids": ["https://orcid.org/0009-0004-1986-8982"]}, "institutions": [{"id": "https://openalex.org/I129604602", "display_name": "The University of Sydney", "ror": "https://ror.org/0384j8v12", "country_code": "AU", "type": "education", "lineage": ["https://openalex.org/I129604602"]}], "countries": ["AU"], "is_corresponding": false, "raw_author_name": "Yanli Li", "raw_affiliation_strings": ["The University of Sydney", "The University of Sydney, University of Sydney, Sydney, 2006, Australia"], "raw_orcid": null, "affiliations": [{"raw_affiliation_string": "The University of Sydney", "institution_ids": ["https://openalex.org/I129604602"]}, {"raw_affiliation_string": "The University of Sydney, University of Sydney, Sydney, 2006, Australia", "institution_ids": ["https://openalex.org/I129604602"]}]}, {"author_position": "last", "author": {"id": "https://openalex.org/A5054168288", "display_name": "Dong Yuan", "orcid": "https://orcid.org/0000-0003-1130-0888", "observed_orcids": ["https://orcid.org/0000-0003-1130-0888"]}, "institutions": [{"id": "https://openalex.org/I129604602", "display_name": "The University of Sydney", "ror": "https://ror.org/0384j8v12", "country_code": "AU", "type": "education", "lineage": ["https://openalex.org/I129604602"]}], "countries": ["AU"], "is_corresponding": true, "raw_author_name": "Dong Yuan", "raw_affiliation_strings": ["The University of Sydney", "The University of Sydney, University of Sydney, Sydney, 2006, Australia"], "raw_orcid": null, "affiliations": [{"raw_affiliation_string": "The University of Sydney", "institution_ids": ["https://openalex.org/I129604602"]}, {"raw_affiliation_string": "The University of Sydney, University of Sydney, Sydney, 2006, Australia", "institution_ids": ["https://openalex.org/I129604602"]}]}], "abstract_inverted_index": null, "primary_topic": {"id": "https://openalex.org/T10054", "display_name": "Parallel Computing and Optimization Techniques", "score": 0.8812999725341797, "subfield": {"id": "https://openalex.org/subfields/1708", "display_name": "Hardware and Architecture"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, "topics": [{"id": "https://openalex.org/T10054", "display_name": "Parallel Computing and Optimization Techniques", "score": 0.8812999725341797, "subfield": {"id": "https://openalex.org/subfields/1708", "display_name": "Hardware and Architecture"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, {"id": "https://openalex.org/T10036", "display_name": "Advanced Neural Network Applications", "score": 0.058800000697374344, "subfield": {"id": "https://openalex.org/subfields/1707", "display_name": "Computer Vision and Pattern Recognition"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, {"id": "https://openalex.org/T10181", "display_name": "Natural Language Processing Techniques", "score": 0.013000000268220901, "subfield": {"id": "https://openalex.org/subfields/1702", "display_name": "Artificial Intelligence"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}], "primary_location": {"id": "doi:10.2139/ssrn.5551142", "is_oa": true, "landing_page_url": "https://doi.org/10.2139/ssrn.5551142", "pdf_url": null, "source": {"id": "https://openalex.org/S4210172589", "display_name": "SSRN Electronic Journal", "issn_l": "1556-5068", "issn": ["1556-5068"], "is_oa": true, "is_in_doaj": false, "is_core": false, "listed_in": [], "host_organization": "https://openalex.org/I1318003438", "host_organization_name": "RELX Group (Netherlands)", "host_organization_lineage": ["https://openalex.org/I1318003438"], "host_organization_lineage_names": [], "type": "repository"}, "license": null, "license_id": null, "version": "acceptedVersion", "is_accepted": true, "is_published": false, "raw_source_name": null, "raw_type": "posted-content"}, "cited_by_count": 0, "is_retracted": false}, {"id": "https://openalex.org/W4406960586", "doi": "https://doi.org/10.48550/arxiv.2501.16909", "title": "Understanding GPU Resource Interference One Level Deeper", "publication_year": 2025, "publication_date": "2025-01-28", "type": "preprint", "language": "en", "authorships": [{"author_position": "first", "author": {"id": "https://openalex.org/A5095776641", "display_name": "Paul Elvinger", "orcid": "https://orcid.org/0009-0000-6025-844X", "observed_orcids": ["https://orcid.org/0009-0000-6025-844X"]}, "institutions": [{"id": "https://openalex.org/I35440088", "display_name": "ETH Zurich", "ror": "https://ror.org/05a28rw58", "country_code": "CH", "type": "education", "lineage": ["https://openalex.org/I2799323385", "https://openalex.org/I35440088"]}], "countries": ["CH"], "is_corresponding": false, "raw_author_name": "Elvinger, Paul", "raw_affiliation_strings": ["ETH Zurich Switzerland"], "raw_orcid": null, "affiliations": [{"raw_affiliation_string": "ETH Zurich Switzerland", "institution_ids": ["https://openalex.org/I35440088"]}]}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5049951876", "display_name": "Foteini Strati", "orcid": "https://orcid.org/0000-0003-3364-2109", "observed_orcids": ["https://orcid.org/0000-0003-3364-2109"]}, "institutions": [{"id": "https://openalex.org/I35440088", "display_name": "ETH Zurich", "ror": "https://ror.org/05a28rw58", "country_code": "CH", "type": "education", "lineage": ["https://openalex.org/I2799323385", "https://openalex.org/I35440088"]}], "countries": ["CH"], "is_corresponding": false, "raw_author_name": "Strati, Foteini", "raw_affiliation_strings": ["ETH Zurich Switzerland"], "raw_orcid": null, "affiliations": [{"raw_affiliation_string": "ETH Zurich Switzerland", "institution_ids": ["https://openalex.org/I35440088"]}]}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5060103552", "display_name": "Natalie Enright Jerger", "orcid": "https://orcid.org/0000-0002-0526-2080", "observed_orcids": ["https://orcid.org/0000-0002-0526-2080"]}, "institutions": [{"id": "https://openalex.org/I185261750", "display_name": "University of Toronto", "ror": "https://ror.org/03dbr7087", "country_code": "CA", "type": "education", "lineage": ["https://openalex.org/I185261750"]}], "countries": ["CA"], "is_corresponding": false, "raw_author_name": "Jerger, Natalie Enright", "raw_affiliation_strings": ["University of Toronto Canada"], "raw_orcid": null, "affiliations": [{"raw_affiliation_string": "University of Toronto Canada", "institution_ids": ["https://openalex.org/I185261750"]}]}, {"author_position": "last", "author": {"id": "https://openalex.org/A5000009841", "display_name": "Ana Klimovic", "orcid": "https://orcid.org/0000-0001-8559-0529", "observed_orcids": ["https://orcid.org/0000-0001-8559-0529"]}, "institutions": [{"id": "https://openalex.org/I35440088", "display_name": "ETH Zurich", "ror": "https://ror.org/05a28rw58", "country_code": "CH", "type": "education", "lineage": ["https://openalex.org/I2799323385", "https://openalex.org/I35440088"]}], "countries": ["CH"], "is_corresponding": false, "raw_author_name": "Klimovic, Ana", "raw_affiliation_strings": ["ETH Zurich Switzerland"], "raw_orcid": null, "affiliations": [{"raw_affiliation_string": "ETH Zurich Switzerland", "institution_ids": ["https://openalex.org/I35440088"]}]}], "abstract_inverted_index": {"GPUs": [0], "are": [1], "vastly": [2], "underutilized,": [3], "even": [4], "when": [5], "running": [6], "resource-intensive": [7], "AI": [8], "applications,": [9], "as": [10, 47, 76, 116], "GPU": [11, 40, 73, 92, 120, 133], "kernels": [12, 134], "within": [13], "each": [14, 107], "job": [15], "have": [16], "diverse": [17], "resource": [18, 93], "profiles": [19], "that": [20, 122], "may": [21], "saturate": [22], "some": [23], "parts": [24, 32], "of": [25, 67, 91, 103], "a": [26, 64, 97, 104], "device": [27], "while": [28], "often": [29], "leaving": [30], "other": [31], "idle.": [33], "Colocating": [34], "applications": [35, 69], "is": [36, 43], "known": [37], "to": [38, 51, 56, 99, 106, 139], "improve": [39, 140], "utilization,": [41], "but": [42], "not": [44], "common": [45], "practice": [46], "it": [48], "becomes": [49], "difficult": [50], "provide": [52], "predictable": [53, 60], "performance": [54, 61, 125], "due": [55], "workload": [57, 105], "interference.": [58], "Providing": [59], "guarantees": [62, 126], "requires": [63], "deep": [65], "understanding": [66], "how": [68, 111, 128], "contend": [70], "for": [71, 119], "shared": [72], "resources": [74], "such": [75], "block": [77], "schedulers,": [78], "compute": [79], "units,": [80], "L1/L2": [81], "caches,": [82], "and": [83, 95, 127], "memory": [84], "bandwidth.": [85], "We": [86, 109], "study": [87], "the": [88, 101, 117], "key": [89], "types": [90], "interference": [94], "develop": [96], "methodology": [98, 113], "quantify": [100], "sensitivity": [102], "type.": [108], "discuss": [110], "this": [112], "can": [114, 131], "serve": [115], "foundation": [118], "schedulers": [121], "enforce": [123], "strict": [124], "application": [129], "developers": [130], "design": [132], "with": [135], "colocation": [136], "in": [137], "mind": [138], "efficiency.": [141]}, "primary_topic": {"id": "https://openalex.org/T10054", "display_name": "Parallel Computing and Optimization Techniques", "score": 0.9878000020980835, "subfield": {"id": "https://openalex.org/subfields/1708", "display_name": "Hardware and Architecture"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, "topics": [{"id": "https://openalex.org/T10054", "display_name": "Parallel Computing and Optimization Techniques", "score": 0.9878000020980835, "subfield": {"id": "https://openalex.org/subfields/1708", "display_name": "Hardware and Architecture"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, {"id": "https://openalex.org/T10715", "display_name": "Distributed and Parallel Computing Systems", "score": 0.004399999976158142, "subfield": {"id": "https://openalex.org/subfields/1705", "display_name": "Computer Networks and Communications"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, {"id": "https://openalex.org/T10101", "display_name": "Cloud Computing and Resource Management", "score": 0.002199999988079071, "subfield": {"id": "https://openalex.org/subfields/1710", "display_name": "Information Systems"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}], "primary_location": {"id": "pmh:oai:arXiv.org:2501.16909", "is_oa": true, "landing_page_url": "http://arxiv.org/abs/2501.16909", "pdf_url": "https://arxiv.org/pdf/2501.16909", "source": {"id": "https://openalex.org/S4306400194", "display_name": "arXiv (Cornell University)", "issn_l": "2331-8422", "issn": ["2331-8422"], "is_oa": true, "is_in_doaj": false, "is_core": false, "listed_in": [], "host_organization": "https://openalex.org/I205783295", "host_organization_name": "Cornell University", "host_organization_lineage": ["https://openalex.org/I205783295"], "host_organization_lineage_names": [], "type": "repository"}, "license": "cc-by", "license_id": "https://openalex.org/licenses/cc-by", "version": "submittedVersion", "is_accepted": false, "is_published": false, "raw_source_name": null, "raw_type": "text"}, "cited_by_count": 2, "is_retracted": false}, {"id": "https://openalex.org/W4414895116", "doi": "https://doi.org/10.48550/arxiv.2506.01097", "title": "Task-Related Token Compression in Multimodal Large Language Models from an Explainability Perspective", "publication_year": 2025, "publication_date": "2025-06-01", "type": "preprint", "language": "en", "authorships": [{"author_position": "first", "author": {"id": null, "display_name": "Lei, Lei", "orcid": null, "observed_orcids": []}, "institutions": [{"id": "https://openalex.org/I126520041", "display_name": "University of Science and Technology of China", "ror": "https://ror.org/04c4dkn09", "country_code": "CN", "type": "education", "lineage": ["https://openalex.org/I126520041", "https://openalex.org/I19820366"]}, {"id": "https://openalex.org/I4405279436", "display_name": "Shanghai Innovation Institute", "ror": "https://ror.org/01fatbw66", "country_code": "CN", "type": "education", "lineage": ["https://openalex.org/I4405279436"]}], "countries": ["CN"], "is_corresponding": false, "raw_author_name": "Lei, Lei", "raw_affiliation_strings": ["Shanghai Innovation Institute", "University of Science and Technology of China"], "raw_orcid": null, "affiliations": [{"raw_affiliation_string": "Shanghai Innovation Institute", "institution_ids": ["https://openalex.org/I4405279436"]}, {"raw_affiliation_string": "University of Science and Technology of China", "institution_ids": ["https://openalex.org/I126520041"]}]}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5101730095", "display_name": "Jie Gu", "orcid": "https://orcid.org/0000-0001-9262-2234", "observed_orcids": ["https://orcid.org/0000-0001-9262-2234"]}, "institutions": [{"id": "https://openalex.org/I4210116723", "display_name": "Robotics Research (United States)", "ror": "https://ror.org/020w2fr77", "country_code": "US", "type": "company", "lineage": ["https://openalex.org/I4210116723"]}], "countries": ["US"], "is_corresponding": false, "raw_author_name": "Gu, Jie", "raw_affiliation_strings": ["Rightly Robotics"], "raw_orcid": null, "affiliations": [{"raw_affiliation_string": "Rightly Robotics", "institution_ids": ["https://openalex.org/I4210116723"]}]}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5045699018", "display_name": "Xiaolei Ma", "orcid": "https://orcid.org/0000-0002-3841-5792", "observed_orcids": ["https://orcid.org/0000-0002-3841-5792"]}, "institutions": [{"id": "https://openalex.org/I4210116723", "display_name": "Robotics Research (United States)", "ror": "https://ror.org/020w2fr77", "country_code": "US", "type": "company", "lineage": ["https://openalex.org/I4210116723"]}], "countries": ["US"], "is_corresponding": false, "raw_author_name": "Ma, Xiaokang", "raw_affiliation_strings": ["Rightly Robotics"], "raw_orcid": null, "affiliations": [{"raw_affiliation_string": "Rightly Robotics", "institution_ids": ["https://openalex.org/I4210116723"]}]}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5102219351", "display_name": "Tang Chu", "orcid": "https://orcid.org/0009-0008-0930-5325", "observed_orcids": ["https://orcid.org/0009-0008-0930-5325", "https://orcid.org/0009-0004-4533-5792"]}, "institutions": [{"id": "https://openalex.org/I4210116723", "display_name": "Robotics Research (United States)", "ror": "https://ror.org/020w2fr77", "country_code": "US", "type": "company", "lineage": ["https://openalex.org/I4210116723"]}], "countries": ["US"], "is_corresponding": false, "raw_author_name": "Tang, Chu", "raw_affiliation_strings": ["Rightly Robotics"], "raw_orcid": null, "affiliations": [{"raw_affiliation_string": "Rightly Robotics", "institution_ids": ["https://openalex.org/I4210116723"]}]}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5090901796", "display_name": "Jingmin Chen", "orcid": "https://orcid.org/0000-0001-6336-5427", "observed_orcids": ["https://orcid.org/0000-0001-6336-5427"]}, "institutions": [{"id": "https://openalex.org/I4210116723", "display_name": "Robotics Research (United States)", "ror": "https://ror.org/020w2fr77", "country_code": "US", "type": "company", "lineage": ["https://openalex.org/I4210116723"]}], "countries": ["US"], "is_corresponding": false, "raw_author_name": "Chen, Jingmin", "raw_affiliation_strings": ["Rightly Robotics"], "raw_orcid": null, "affiliations": [{"raw_affiliation_string": "Rightly Robotics", "institution_ids": ["https://openalex.org/I4210116723"]}]}, {"author_position": "last", "author": {"id": "https://openalex.org/A5101570536", "display_name": "Tong Xu", "orcid": "https://orcid.org/0000-0001-5564-192X", "observed_orcids": ["https://orcid.org/0000-0001-5564-192X"]}, "institutions": [{"id": "https://openalex.org/I126520041", "display_name": "University of Science and Technology of China", "ror": "https://ror.org/04c4dkn09", "country_code": "CN", "type": "education", "lineage": ["https://openalex.org/I126520041", "https://openalex.org/I19820366"]}], "countries": ["CN"], "is_corresponding": false, "raw_author_name": "Xu, Tong", "raw_affiliation_strings": ["University of Science and Technology of China"], "raw_orcid": null, "affiliations": [{"raw_affiliation_string": "University of Science and Technology of China", "institution_ids": ["https://openalex.org/I126520041"]}]}], "abstract_inverted_index": {"Existing": [0], "Multimodal": [1], "Large": [2], "Language": [3], "Models": [4], "(MLLMs)": [5], "process": [6], "a": [7, 151, 171, 182], "large": [8], "number": [9], "of": [10, 35, 53, 87, 126, 157, 194, 219], "visual": [11, 21, 43, 99, 128], "tokens,": [12], "leading": [13, 206], "to": [14, 58, 132, 149, 162], "significant": [15], "computational": [16], "costs": [17], "and": [18, 101, 184, 192, 201, 210, 216, 236], "inefficiency.": [19], "Instruction-related": [20], "token": [22, 60, 79, 129, 142], "compression": [23, 61, 80, 143, 225], "demonstrates": [24], "strong": [25, 217], "task": [26], "relevance,": [27], "which": [28, 55, 136], "aligns": [29], "well": [30], "with": [31, 75, 89, 130, 230], "MLLMs": [32, 207], "ultimate": [33], "goal": [34], "instruction": [36], "following.": [37], "Previous": [38], "works": [39], "generally": [40], "assume": [41], "that": [42, 74, 115], "tokens": [44, 100], "achieve": [45], "better": [46], "vision-language": [47], "alignment": [48], "in": [49, 65, 232], "the": [50, 84, 109, 123, 133, 140, 154, 158, 163, 168, 213], "shallow": [51], "layers": [52], "LLMs,": [54], "have": [56], "led": [57], "task-related": [59, 78, 141], "being": [62], "primarily": [63], "applied": [64], "intermediate": [66], "LLM": [67, 88, 110, 160], "layers.": [68], "In": [69], "contrast,": [70], "our": [71, 220, 223], "study": [72], "reveals": [73], "proper": [76], "selection,": [77], "is": [81, 190], "feasible": [82], "at": [83], "input": [85], "stage": [86], "negligible": [90], "performance": [91], "loss.": [92], "This": [93], "new": [94, 224], "paradigm": [95, 226], "significantly": [96], "reduces": [97], "task-irrelevant": [98], "its": [102], "model-agnostic": [103], "design": [104], "enables": [105], "application": [106], "without": [107], "modifying": [108], "architecture.": [111], "Specifically,": [112], "we": [113, 147], "suggest": [114], "explainability": [116], "methods": [117], "for": [118, 144, 170], "transformer-based": [119], "architechtures": [120], "can": [121, 137, 178], "evaluate": [122], "global": [124], "importance": [125], "each": [127], "respect": [131], "given": [134], "instruction,": [135], "effectively": [138], "guide": [139], "MLLMs.": [145, 195], "Furthermore,": [146], "propose": [148], "learn": [150], "mapping": [152, 177], "from": [153], "attention": [155], "map": [156], "first": [159], "layer": [161], "explanation": [164], "results,": [165], "thereby": [166], "avoiding": [167], "need": [169], "full": [172], "inference": [173, 229], "pass.": [174], "Interestingly,": [175], "this": [176], "be": [179], "learned": [180], "using": [181], "simple": [183], "lightweight": [185], "convolutional": [186], "network,": [187], "whose": [188], "training": [189], "efficient": [191], "independent": [193], "Extensive": [196], "experiments": [197], "on": [198], "13": [199], "image": [200], "video": [202], "benchmarks": [203], "across": [204], "three": [205], "(Qwen2-VL,": [208], "LLaVA-OneVision,": [209], "VILA1.5)": [211], "demonstrate": [212], "remarkable": [214], "effectiveness": [215], "generalization": [218], "approach.": [221], "Additionally,": [222], "achieves": [227], "faster": [228], "reductions": [231], "both": [233], "prefilling": [234], "time": [235], "KV": [237], "cache": [238], "memory.": [239]}, "primary_topic": {"id": "https://openalex.org/T11714", "display_name": "Multimodal Machine Learning Applications", "score": 0.9976000189781189, "subfield": {"id": "https://openalex.org/subfields/1707", "display_name": "Computer Vision and Pattern Recognition"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, "topics": [{"id": "https://openalex.org/T11714", "display_name": "Multimodal Machine Learning Applications", "score": 0.9976000189781189, "subfield": {"id": "https://openalex.org/subfields/1707", "display_name": "Computer Vision and Pattern Recognition"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, {"id": "https://openalex.org/T12026", "display_name": "Explainable Artificial Intelligence (XAI)", "score": 0.002199999988079071, "subfield": {"id": "https://openalex.org/subfields/1702", "display_name": "Artificial Intelligence"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, {"id": "https://openalex.org/T10036", "display_name": "Advanced Neural Network Applications", "score": 0.00019999999494757503, "subfield": {"id": "https://openalex.org/subfields/1707", "display_name": "Computer Vision and Pattern Recognition"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}], "primary_location": {"id": "pmh:oai:arXiv.org:2506.01097", "is_oa": true, "landing_page_url": "https://arxiv.org/abs/2506.01097", "pdf_url": "https://arxiv.org/pdf/2506.01097", "source": {"id": "https://openalex.org/S4306400194", "display_name": "arXiv (Cornell University)", "issn_l": "2331-8422", "issn": ["2331-8422"], "is_oa": true, "is_in_doaj": false, "is_core": false, "listed_in": [], "host_organization": "https://openalex.org/I205783295", "host_organization_name": "Cornell University", "host_organization_lineage": ["https://openalex.org/I205783295"], "host_organization_lineage_names": [], "type": "repository"}, "license": null, "license_id": null, "version": "submittedVersion", "is_accepted": false, "is_published": false, "raw_source_name": null, "raw_type": "text"}, "cited_by_count": 0, "is_retracted": false}, {"id": "https://openalex.org/W4417295437", "doi": "https://doi.org/10.48550/arxiv.2512.10310", "title": "Efficient-VLN: A Simple yet Strong Baseline for Efficient Vision-Language Navigation", "publication_year": 2025, "publication_date": "2025-12-11", "type": "preprint", "language": null, "authorships": [{"author_position": "first", "author": {"id": null, "display_name": "Zheng, Duo", "orcid": null, "observed_orcids": []}, "institutions": [{"id": "https://openalex.org/I177725633", "display_name": "Chinese University of Hong Kong", "ror": "https://ror.org/00t33hh48", "country_code": "HK", "type": "education", "lineage": ["https://openalex.org/I177725633"]}], "countries": ["HK"], "is_corresponding": false, "raw_author_name": "Zheng, Duo", "raw_affiliation_strings": ["The Chinese University of Hong Kong https :// lavi-lab.github "], "raw_orcid": null, "affiliations": [{"raw_affiliation_string": "The Chinese University of Hong Kong https :// lavi-lab.github ", "institution_ids": ["https://openalex.org/I177725633"]}]}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5108383963", "display_name": "Shijia Huang", "orcid": "https://orcid.org/0009-0002-5501-5325", "observed_orcids": ["https://orcid.org/0009-0002-5501-5325"]}, "institutions": [{"id": "https://openalex.org/I177725633", "display_name": "Chinese University of Hong Kong", "ror": "https://ror.org/00t33hh48", "country_code": "HK", "type": "education", "lineage": ["https://openalex.org/I177725633"]}], "countries": ["HK"], "is_corresponding": false, "raw_author_name": "Huang, Shijia", "raw_affiliation_strings": ["The Chinese University of Hong Kong https :// lavi-lab.github "], "raw_orcid": null, "affiliations": [{"raw_affiliation_string": "The Chinese University of Hong Kong https :// lavi-lab.github ", "institution_ids": ["https://openalex.org/I177725633"]}]}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5101595086", "display_name": "Yanyang Li", "orcid": "https://orcid.org/0000-0002-6534-0618", "observed_orcids": ["https://orcid.org/0000-0002-6534-0618"]}, "institutions": [{"id": "https://openalex.org/I177725633", "display_name": "Chinese University of Hong Kong", "ror": "https://ror.org/00t33hh48", "country_code": "HK", "type": "education", "lineage": ["https://openalex.org/I177725633"]}], "countries": ["HK"], "is_corresponding": false, "raw_author_name": "Li, Yanyang", "raw_affiliation_strings": ["The Chinese University of Hong Kong https :// lavi-lab.github "], "raw_orcid": null, "affiliations": [{"raw_affiliation_string": "The Chinese University of Hong Kong https :// lavi-lab.github ", "institution_ids": ["https://openalex.org/I177725633"]}]}, {"author_position": "last", "author": {"id": "https://openalex.org/A5100406725", "display_name": "Liwei Wang", "orcid": "https://orcid.org/0000-0003-3264-1294", "observed_orcids": ["https://orcid.org/0000-0003-3264-1294"]}, "institutions": [{"id": "https://openalex.org/I177725633", "display_name": "Chinese University of Hong Kong", "ror": "https://ror.org/00t33hh48", "country_code": "HK", "type": "education", "lineage": ["https://openalex.org/I177725633"]}], "countries": ["HK"], "is_corresponding": false, "raw_author_name": "Wang, Liwei", "raw_affiliation_strings": ["The Chinese University of Hong Kong https :// lavi-lab.github "], "raw_orcid": null, "affiliations": [{"raw_affiliation_string": "The Chinese University of Hong Kong https :// lavi-lab.github ", "institution_ids": ["https://openalex.org/I177725633"]}]}], "abstract_inverted_index": {"While": [0], "Multimodal": [1], "Large": [2], "Language": [3], "Models": [4], "(MLLMs)": [5], "have": [6], "demonstrated": [7], "significant": [8], "promise": [9], "in": [10, 47], "Vision-Language": [11], "Navigation": [12], "(VLN),": [13], "existing": [14], "agents": [15], "remain": [16], "heavily": [17], "constrained": [18], "by": [19, 118], "systemic": [20], "bottlenecks": [21], "across": [22, 154], "inference,": [23], "training,": [24, 43], "and": [25, 44, 62, 135, 159], "data": [26, 49], "collection.": [27, 50], "Specifically,": [28], "they": [29], "suffer": [30], "from": [31], "prohibitive": [32], "latency": [33, 169], "due": [34], "to": [35, 86, 109, 130, 172], "visual": [36], "history": [37], "reprocessing,": [38], "action": [39, 120], "leakage": [40], "during": [41], "sequence-packed": [42], "suboptimal": [45], "exploration": [46, 134], "self-correction": [48], "To": [51], "overcome": [52], "these": [53, 68], "intertwined": [54], "challenges,": [55], "we": [56], "present": [57], "Efficient-VLN,": [58], "a": [59, 167, 178], "highly": [60], "efficient": [61], "robust": [63], "baseline": [64], "that": [65, 148], "systematically": [66], "resolves": [67], "issues": [69], "through": [70], "three": [71], "simple-yet-effective": [72], "mechanisms.": [73], "(1)": [74], "Inference:": [75], "We": [76, 101, 125], "introduce": [77], "KV-cache": [78], "reuse": [79], "with": [80, 105], "contiguous": [81], "RoPE,": [82], "enabling": [83], "the": [84, 89, 115, 152, 155, 173], "model": [85], "process": [87], "only": [88], "newly": [90], "observed": [91], "frame": [92], "at": [93], "each": [94], "step": [95], "for": [96, 181], "real-time": [97], "inference.": [98], "(2)": [99], "Training:": [100], "propose": [102], "packed": [103], "training": [104], "an": [106, 127], "action-isolating": [107], "mask": [108], "accelerate": [110], "throughput": [111], "while": [112], "effectively": [113], "bridging": [114], "training-inference": [116], "gap": [117], "preventing": [119], "leakage.": [121], "(3)": [122], "Data": [123], "Collection:": [124], "employ": [126], "Adaptive": [128], "DAgger": [129], "dynamically": [131], "balance": [132], "autonomous": [133], "oracle": [136], "guidance,": [137], "enhancing": [138], "error-recovery": [139], "capability": [140], "without": [141], "escalating": [142], "computational": [143], "costs.": [144], "Extensive": [145], "evaluations": [146], "show": [147], "Efficient-VLN": [149], "significantly": [150], "advances": [151], "state-of-the-art": [153, 175], "R2R-CE": [156], "(73.2%": [157], "SR)": [158, 162], "RxR-CE": [160], "(75.6%": [161], "benchmarks.": [163], "Meanwhile,": [164], "it": [165], "yields": [166], "28%": [168], "reduction": [170], "compared": [171], "previous": [174], "StreamVLN,": [176], "establishing": [177], "new": [179], "paradigm": [180], "streaming": [182], "MLLM-based": [183], "navigation.": [184]}, "primary_topic": {"id": "https://openalex.org/T11714", "display_name": "Multimodal Machine Learning Applications", "score": 0.9779999852180481, "subfield": {"id": "https://openalex.org/subfields/1707", "display_name": "Computer Vision and Pattern Recognition"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, "topics": [{"id": "https://openalex.org/T11714", "display_name": "Multimodal Machine Learning Applications", "score": 0.9779999852180481, "subfield": {"id": "https://openalex.org/subfields/1707", "display_name": "Computer Vision and Pattern Recognition"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, {"id": "https://openalex.org/T10462", "display_name": "Reinforcement Learning in Robotics", "score": 0.007199999876320362, "subfield": {"id": "https://openalex.org/subfields/1702", "display_name": "Artificial Intelligence"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, {"id": "https://openalex.org/T10653", "display_name": "Robot Manipulation and Learning", "score": 0.002099999925121665, "subfield": {"id": "https://openalex.org/subfields/2207", "display_name": "Control and Systems Engineering"}, "field": {"id": "https://openalex.org/fields/22", "display_name": "Engineering"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}], "primary_location": {"id": "pmh:oai:arXiv.org:2512.10310", "is_oa": true, "landing_page_url": "https://arxiv.org/abs/2512.10310", "pdf_url": "https://arxiv.org/pdf/2512.10310", "source": {"id": "https://openalex.org/S4306400194", "display_name": "arXiv (Cornell University)", "issn_l": "2331-8422", "issn": ["2331-8422"], "is_oa": true, "is_in_doaj": false, "is_core": false, "listed_in": [], "host_organization": "https://openalex.org/I205783295", "host_organization_name": "Cornell University", "host_organization_lineage": ["https://openalex.org/I205783295"], "host_organization_lineage_names": [], "type": "repository"}, "license": "cc-by", "license_id": "https://openalex.org/licenses/cc-by", "version": "submittedVersion", "is_accepted": false, "is_published": false, "raw_source_name": null, "raw_type": "text"}, "cited_by_count": 0, "is_retracted": false}, {"id": "https://openalex.org/W4416123688", "doi": "https://doi.org/10.48550/arxiv.2504.03648", "title": "AIBrix: Towards Scalable, Cost-Effective Large Language Model Inference Infrastructure", "publication_year": 2025, "publication_date": "2025-02-22", "type": "preprint", "language": "en", "authorships": [{"author_position": "first", "author": {"id": "https://openalex.org/A5120645796", "display_name": "The AIBrix Team", "orcid": null, "observed_orcids": []}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "The AIBrix Team", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": null, "display_name": "Shan, Jiaxin", "orcid": null, "observed_orcids": []}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Shan, Jiaxin", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": null, "display_name": "Gupta, Varun", "orcid": null, "observed_orcids": []}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Gupta, Varun", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5101773824", "display_name": "Liangliang Xu", "orcid": "https://orcid.org/0000-0001-5202-2795", "observed_orcids": ["https://orcid.org/0000-0001-5202-2795"]}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Xu, Le", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5016218994", "display_name": "Haiyang Shi", "orcid": "https://orcid.org/0000-0002-6105-4737", "observed_orcids": ["https://orcid.org/0000-0002-6105-4737"]}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Shi, Haiyang", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5100639601", "display_name": "Jingyuan Zhang", "orcid": "https://orcid.org/0000-0001-6644-4673", "observed_orcids": ["https://orcid.org/0000-0001-6644-4673"]}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Zhang, Jingyuan", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5100387209", "display_name": "Ning Wang", "orcid": "https://orcid.org/0000-0003-1568-6933", "observed_orcids": ["https://orcid.org/0000-0003-1568-6933"]}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Wang, Ning", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5032862238", "display_name": "Lin Xu", "orcid": "https://orcid.org/0000-0003-0767-4377", "observed_orcids": ["https://orcid.org/0000-0003-0767-4377"]}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Xu, Linhui", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5088332394", "display_name": "Rong Kang", "orcid": "https://orcid.org/0009-0002-5158-618X", "observed_orcids": ["https://orcid.org/0009-0002-5158-618X", "https://orcid.org/0000-0001-5998-492X"]}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Kang, Rong", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5038870570", "display_name": "Tongping Liu", "orcid": "https://orcid.org/0000-0002-1968-4081", "observed_orcids": ["https://orcid.org/0000-0002-1968-4081"]}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Liu, Tongping", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5100386936", "display_name": "Yifei Zhang", "orcid": "https://orcid.org/0000-0002-5177-0465", "observed_orcids": ["https://orcid.org/0000-0002-5177-0465"]}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Zhang, Yifei", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": null, "display_name": "Zhu, Yiqing", "orcid": null, "observed_orcids": []}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Zhu, Yiqing", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5043469151", "display_name": "Shuowei Jin", "orcid": "https://orcid.org/0000-0002-5817-9029", "observed_orcids": ["https://orcid.org/0000-0002-5817-9029"]}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Jin, Shuowei", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": null, "display_name": "Lim, Gangmuk", "orcid": null, "observed_orcids": []}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Lim, Gangmuk", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5100687669", "display_name": "Bin-Bin Chen", "orcid": "https://orcid.org/0000-0001-8896-6173", "observed_orcids": ["https://orcid.org/0000-0001-8896-6173"]}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Chen, Binbin", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5069104472", "display_name": "Zuzhi Chen", "orcid": null, "observed_orcids": []}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Chen, Zuzhi", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5100441094", "display_name": "Xiao Liu", "orcid": "https://orcid.org/0000-0002-9226-4569", "observed_orcids": ["https://orcid.org/0000-0002-9226-4569", "https://orcid.org/0009-0006-3387-2317"]}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Liu, Xiao", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5100613516", "display_name": "Xin Chen", "orcid": "https://orcid.org/0000-0001-7720-8916", "observed_orcids": ["https://orcid.org/0000-0001-7720-8916"]}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Chen, Xin", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5120645797", "display_name": "Kante Yin", "orcid": null, "observed_orcids": []}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Yin, Kante", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": null, "display_name": "Chung, Chak-Pong", "orcid": null, "observed_orcids": []}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Chung, Chak-Pong", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5109492481", "display_name": "Chenyu Jiang", "orcid": "https://orcid.org/0009-0004-3699-1974", "observed_orcids": ["https://orcid.org/0009-0004-3699-1974"]}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Jiang, Chenyu", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": null, "display_name": "Lu, Yicheng", "orcid": null, "observed_orcids": []}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Lu, Yicheng", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5041795098", "display_name": "Jianjun Chen", "orcid": "https://orcid.org/0000-0001-7511-1117", "observed_orcids": ["https://orcid.org/0000-0001-7511-1117"]}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Chen, Jianjun", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5077661988", "display_name": "Caixue Lin", "orcid": "https://orcid.org/0009-0003-0570-6352", "observed_orcids": ["https://orcid.org/0009-0003-0570-6352"]}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Lin, Caixue", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": null, "display_name": "Xiang, Wu", "orcid": null, "observed_orcids": []}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Xiang, Wu", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": null, "display_name": "Shi, Rui", "orcid": null, "observed_orcids": []}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Shi, Rui", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "last", "author": {"id": "https://openalex.org/A5040158607", "display_name": "Liguang Xie", "orcid": "https://orcid.org/0000-0002-3764-9888", "observed_orcids": ["https://orcid.org/0000-0002-3764-9888"]}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Xie, Liguang", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}], "abstract_inverted_index": {"We": [0], "introduce": [1], "AIBrix,": [2], "a": [3, 24, 75, 86, 92], "cloud-native,": [4], "open-source": [5], "framework": [6], "designed": [7], "to": [8, 48, 85, 133, 150, 175], "optimize": [9], "and": [10, 52, 65, 91, 127, 136, 172], "simplify": [11], "large-scale": [12, 114], "LLM": [13], "deployment": [14], "in": [15, 89, 95], "cloud": [16], "environments.": [17], "Unlike": [18], "traditional": [19], "cloud-native": [20], "stacks,": [21], "AIBrix": [22, 43, 73, 98, 117, 159, 179], "follows": [23], "co-design": [25], "philosophy,": [26], "ensuring": [27], "every": [28], "layer": [29], "of": [30], "the": [31], "infrastructure": [32], "is": [33, 180], "purpose-built": [34], "for": [35, 59, 124, 129], "seamless": [36], "integration": [37], "with": [38, 163], "inference": [39, 50, 96], "engines": [40], "like": [41], "vLLM.": [42], "introduces": [44], "several": [45], "key": [46], "innovations": [47], "reduce": [49], "costs": [51], "enhance": [53], "performance": [54], "including": [55], "high-density": [56], "LoRA": [57], "management": [58, 107], "dynamic": [60], "adapter": [61], "scheduling,": [62], "LLM-specific": [63], "autoscalers,": [64], "prefix-aware,": [66], "load-aware": [67], "routing.": [68], "To": [69], "further": [70], "improve": [71, 176], "efficiency,": [72], "incorporates": [74], "distributed": [76], "KV": [77], "cache,": [78], "boosting": [79], "token": [80], "reuse": [81], "across": [82], "nodes,": [83], "leading": [84], "50%": [87], "increase": [88], "throughput": [90], "70%": [93], "reduction": [94], "latency.": [97], "also": [99], "supports": [100], "unified": [101], "AI": [102, 164], "runtime": [103], "which": [104], "streamlines": [105], "model": [106], "while": [108, 154], "maintaining": [109, 155], "vendor-agnostic": [110], "engine": [111], "compatibility.": [112], "For": [113], "multi-node": [115], "inference,": [116], "employs": [118], "hybrid": [119], "orchestration": [120], "--": [121, 132], "leveraging": [122], "Kubernetes": [123], "coarse-grained": [125], "scheduling": [126], "Ray": [128], "fine-grained": [130], "execution": [131], "balance": [134], "efficiency": [135, 153], "flexibility.": [137], "Additionally,": [138], "an": [139], "SLO-driven": [140], "GPU": [141], "optimizer": [142], "dynamically": [143], "adjusts": [144], "resource": [145], "allocations,": [146], "optimizing": [147], "heterogeneous": [148], "serving": [149], "maximize": [151], "cost": [152], "service": [156], "guarantees.": [157], "Finally,": [158], "enhances": [160], "system": [161], "reliability": [162], "accelerator": [165], "diagnostic": [166], "tools,": [167], "enabling": [168], "automated": [169], "failure": [170], "detection": [171], "mock-up": [173], "testing": [174], "fault": [177], "resilience.": [178], "available": [181], "at": [182], "https://github.com/vllm-project/aibrix.": [183]}, "primary_topic": {"id": "https://openalex.org/T10101", "display_name": "Cloud Computing and Resource Management", "score": 0.9106000065803528, "subfield": {"id": "https://openalex.org/subfields/1710", "display_name": "Information Systems"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, "topics": [{"id": "https://openalex.org/T10101", "display_name": "Cloud Computing and Resource Management", "score": 0.9106000065803528, "subfield": {"id": "https://openalex.org/subfields/1710", "display_name": "Information Systems"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, {"id": "https://openalex.org/T12127", "display_name": "Software System Performance and Reliability", "score": 0.04610000178217888, "subfield": {"id": "https://openalex.org/subfields/1705", "display_name": "Computer Networks and Communications"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, {"id": "https://openalex.org/T10054", "display_name": "Parallel Computing and Optimization Techniques", "score": 0.009200000204145908, "subfield": {"id": "https://openalex.org/subfields/1708", "display_name": "Hardware and Architecture"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}], "primary_location": {"id": "pmh:oai:arXiv.org:2504.03648", "is_oa": true, "landing_page_url": "http://arxiv.org/abs/2504.03648", "pdf_url": "https://arxiv.org/pdf/2504.03648", "source": {"id": "https://openalex.org/S4306400194", "display_name": "arXiv (Cornell University)", "issn_l": "2331-8422", "issn": ["2331-8422"], "is_oa": true, "is_in_doaj": false, "is_core": false, "listed_in": [], "host_organization": "https://openalex.org/I205783295", "host_organization_name": "Cornell University", "host_organization_lineage": ["https://openalex.org/I205783295"], "host_organization_lineage_names": [], "type": "repository"}, "license": null, "license_id": null, "version": "submittedVersion", "is_accepted": false, "is_published": false, "raw_source_name": null, "raw_type": "text"}, "cited_by_count": 0, "is_retracted": false}, {"id": "https://openalex.org/W7117416300", "doi": "https://doi.org/10.5281/zenodo.18072261", "title": "Deterministic Commutative Normalization Achieves z-diff\u21920: A Comparative Analysis of SlimeLearning and VAE-based Disentanglement", "publication_year": 2025, "publication_date": "2025-12-28", "type": "preprint", "language": null, "authorships": [{"author_position": "first", "author": {"id": "https://openalex.org/A5021650017", "display_name": "Hiroshi Sasaki", "orcid": "https://orcid.org/0000-0002-1048-2176", "observed_orcids": ["https://orcid.org/0000-0002-1048-2176"]}, "institutions": [], "countries": [], "is_corresponding": true, "raw_author_name": "SASAKI, HIROSHI", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}], "abstract_inverted_index": {"DETERMINISTIC": [0], "DISENTANGLEMENT": [1], "ACHIEVES": [2], "z-diff\u21920:": [3], "EMPIRICALLY": [4], "VERIFIED": [5], "This": [6, 44, 166], "paper": [7], "presents": [8], "groundbreaking": [9], "empirical": [10], "evidence": [11], "that": [12, 144], "deterministic": [13, 56], "commutative": [14], "normalization": [15], "achieves": [16], "what": [17], "probabilistic": [18, 148, 293], "VAE": [19], "methods": [20], "cannot:": [21], "perfect": [22], "factor": [23], "separation": [24], "(z-diff\u21920).": [25], "\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501THE": [26, 150], "PARADIGM": [27], "SHIFT\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501": [28], "VAE-based": [29], "disentanglement": [30, 178], "(LangVAE,": [31], "\u03b2-VAE)": [32], "is": [33, 158], "fundamentally": [34], "limited": [35], "by": [36], "stochastic": [37], "encoding:": [38], "z": [39], "=": [40, 72, 301], "\u03bc(x)": [41], "+": [42], "\u03c3(x)\u00b7\u03b5.": [43], "inherent": [45], "randomness": [46], "makes": [47], "z-diff=0": [48], "mathematically": [49], "impossible.": [50], "SlimeLearning's": [51], "Attribute-Separated": [52], "Representation": [53], "(ASR)": [54], "uses": [55], "transformation:": [57], "semantically": [58], "equivalent": [59], "inputs": [60], "ALWAYS": [61], "map": [62], "to": [63, 285], "identical": [64], "latent": [65], "points.": [66], "No": [67, 69], "approximation.": [68], "variance.": [70], "z-diff": [71, 120, 125, 130, 300], "0.": [73, 302], "\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501EMPIRICAL": [74], "RESULTS:": [75], "THEORY": [76], "CONFIRMED\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501": [77], "SlimeLearning": [78, 237], "LangVAE": [79], "(2025)": [80], "Improvement\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500z-diff": [81], "\u2192": [82, 86, 90, 197, 203], "0.00": [83], "0.43\u20130.62": [84], "PERFECTz-min-var": [85], "1.00": [87, 91], "0.59\u20130.72": [88], "+40\u201370%Informativeness": [89], "0.34\u20130.49": [92], "+100\u2013190%\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500": [93], "Validated": [94], "on:-": [95], "Synthetic": [96], "language": [97], "data": [98, 259], "(3": [99], "factors": [100], "\u00d7": [101], "3": [102], "levels,": [103], "150+": [104], "batches)-": [105], "dSprites-equivalent": [106], "dataset": [107], "(5": [108], "factors,": [109], "1000": [110], "samples)": [111], "\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501NOISE": [112], "TOLERANCE:": [113], "ROBUST": [114], "BY": [115], "DESIGN\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501": [116], "10%": [117], "Factor": [118], "Noise:": [119], "0.00\u21920.01": [121], "(negligible)20%": [122], "Ambiguous": [123], "Input:": [124], "remains": [126], "0.00High": [127], "Noise": [128], "(>20%):": [129], "~0.05": [131], "(still": [132], "8\u00d7": [133], "better": [134], "than": [135], "VAE)": [136], "SlimeTree's": [137], "Union-Find": [138], "compression": [139], "enables": [140], "recovery": [141], "from": [142], "noise": [143], "would": [145], "catastrophically": [146], "degrade": [147], "models.": [149], "CORE": [151], "INSIGHT\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501": [152], "\"When": [153], "roles": [154], "are": [155], "marked,": [156], "order": [157], "redundant.\"": [159], "\u2014": [160], "SS": [161, 277], "Theory": [162], "(Slime": [163], "Structure": [164], "Theory)": [165], "principle,": [167], "formalized": [168], "in": [169], "SlimeTree": [170], "Patent": [171], "(JP": [172], "2025-183827,": [173], "Claim": [174], "26),": [175], "enables:-": [176], "Perfect": [177], "through": [179], "algebraic": [180, 295], "structure-": [181], "250\u20133000\u00d7": [182], "training": [183], "cost": [184], "reduction-": [185], "Inherent": [186], "interpretability": [187], "without": [188], "post-hoc": [189], "analysis": [190], "\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501IMPLICATIONS\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501": [191], "\u2713": [192], "GPT-4": [193], "class": [194], "training:": [195], "$100M": [196], "$50,000\u2713": [198], "Carbon": [199], "footprint:": [200], "5,000": [201], "tons": [202, 205], "2.5": [204], "(2000\u00d7": [206], "reduction)\u2713": [207], "Interpretability:": [208], "Built-in,": [209], "not": [210, 289], "retrofitted\u2713": [211], "Democratization:": [212], "University": [213], "labs": [214], "can": [215, 238], "train": [216], "frontier": [217], "models": [218], "\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501COMPLEMENTARY": [219], "TO": [220], "LangVAE\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501": [221], "-": [222], "SlimeLearning:": [223], "Training-phase": [224], "optimization": [225], "(deterministic": [226], "disentanglement)-": [227], "LangVAE:": [228], "Post-training": [229], "interpretation": [230], "(controlled": [231], "generation)": [232], "A": [233], "model": [234], "trained": [235], "with": [236, 241, 292], "be": [239, 290], "analyzed": [240], "LangSpace": [242], "metrics\u2014empirically": [243], "validating": [244], "z-diff\u21920": [245], "on": [246], "production": [247], "systems.": [248], "\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501SLIME": [249], "ECOSYSTEM\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501": [250], "Part": [251], "of": [252], "the": [253], "Slime": [254], "technology": [255], "ecosystem:-": [256], "SlimeTree:": [257], "Foundational": [258], "structure": [260], "(Patent": [261], "Pending": [262], "JP": [263], "2025-183827)-": [264], "SlimeLLM:": [265], "Inference": [266], "optimization-": [267], "SlimeNENC:": [268], "Deterministic": [269], "transformation": [270], "(99.9995%": [271], "accuracy)-": [272], "SlimeQCNA:": [273], "Quantum": [274], "error": [275], "correction-": [276], "Theory:": [278], "Unified": [279], "theoretical": [280], "framework": [281], "\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501\u2501": [282], "The": [283], "path": [284], "interpretable": [286], "AI": [287], "need": [288], "paved": [291], "approximations.Deterministic": [294], "approaches": [296], "achieve": [297], "SUPERIOR": [298], "results.": [299], "Empirically": [303], "verified.": [304], "Paradigm": [305], "shifted.": [306]}, "primary_topic": {"id": "https://openalex.org/T10775", "display_name": "Generative Adversarial Networks and Image Synthesis", "score": 0.1940000057220459, "subfield": {"id": "https://openalex.org/subfields/1707", "display_name": "Computer Vision and Pattern Recognition"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, "topics": [{"id": "https://openalex.org/T10775", "display_name": "Generative Adversarial Networks and Image Synthesis", "score": 0.1940000057220459, "subfield": {"id": "https://openalex.org/subfields/1707", "display_name": "Computer Vision and Pattern Recognition"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, {"id": "https://openalex.org/T10028", "display_name": "Topic Modeling", "score": 0.11949999630451202, "subfield": {"id": "https://openalex.org/subfields/1702", "display_name": "Artificial Intelligence"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, {"id": "https://openalex.org/T10057", "display_name": "Face and Expression Recognition", "score": 0.08470000326633453, "subfield": {"id": "https://openalex.org/subfields/1707", "display_name": "Computer Vision and Pattern Recognition"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}], "primary_location": {"id": "doi:10.5281/zenodo.18072261", "is_oa": true, "landing_page_url": "https://doi.org/10.5281/zenodo.18072261", "pdf_url": null, "source": {"id": "https://openalex.org/S4306400562", "display_name": "Zenodo (CERN European Organization for Nuclear Research)", "issn_l": null, "issn": null, "is_oa": true, "is_in_doaj": false, "is_core": false, "listed_in": [], "host_organization": "https://openalex.org/I67311998", "host_organization_name": "European Organization for Nuclear Research", "host_organization_lineage": ["https://openalex.org/I67311998"], "host_organization_lineage_names": [], "type": "repository"}, "license": "cc-by", "license_id": "https://openalex.org/licenses/cc-by", "version": null, "is_accepted": false, "is_published": false, "raw_source_name": null, "raw_type": "Preprint"}, "cited_by_count": 0, "is_retracted": false}, {"id": "https://openalex.org/W4417426823", "doi": null, "title": "Characterizing Communication Patterns in Distributed Large Language Model Inference", "publication_year": 2025, "publication_date": "2025-07-18", "type": "preprint", "language": "en", "authorships": [{"author_position": "first", "author": {"id": "https://openalex.org/A5101847978", "display_name": "Lang Xu", "orcid": "https://orcid.org/0009-0000-6307-4566", "observed_orcids": ["https://orcid.org/0009-0000-6307-4566", "https://orcid.org/0009-0000-4984-494X"]}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Xu, Lang", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5055205226", "display_name": "Kaushik Kandadi Suresh", "orcid": "https://orcid.org/0000-0002-3705-2387", "observed_orcids": ["https://orcid.org/0000-0002-3705-2387"]}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Suresh, Kaushik Kandadi", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5015153122", "display_name": "Quentin G. Anthony", "orcid": "https://orcid.org/0000-0002-6823-9080", "observed_orcids": ["https://orcid.org/0000-0002-6823-9080"]}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Anthony, Quentin", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5061967476", "display_name": "Nawras Alnaasan", "orcid": "https://orcid.org/0000-0002-3638-4144", "observed_orcids": ["https://orcid.org/0000-0002-3638-4144"]}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Alnaasan, Nawras", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "last", "author": {"id": "https://openalex.org/A5024879682", "display_name": "Dhabaleswar K. DK Panda", "orcid": "https://orcid.org/0000-0002-0356-1781", "observed_orcids": ["https://orcid.org/0000-0002-0356-1781"]}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Panda, Dhabaleswar K.", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}], "abstract_inverted_index": {"Large": [0], "Language": [1], "Models": [2], "(LLMs)": [3], "built": [4], "on": [5], "transformer": [6], "architectures": [7, 73], "have": [8], "transformed": [9], "natural": [10], "language": [11], "processing,": [12], "achieving": [13], "remarkable": [14], "performance": [15, 33], "across": [16, 93], "diverse": [17], "applications.": [18], "While": [19], "distributed": [20, 48], "inference": [21, 154], "frameworks": [22, 155], "enable": [23], "practical": [24, 137], "deployment": [25], "of": [26, 71], "these": [27], "models,": [28], "inter-GPU": [29], "communication": [30, 45, 91, 157], "creates": [31], "significant": [32], "constraints": [34], "that": [35, 99], "limit": [36], "service": [37], "quality": [38], "in": [39, 47, 76, 144], "real-world": [40], "systems.": [41], "This": [42], "paper": [43], "investigates": [44], "dynamics": [46], "LLM": [49, 146], "serving-analyzing": [50], "how": [51], "various": [52], "parallelization": [53, 95, 142], "approaches": [54, 126], "coordinate": [55], "data": [56, 117], "exchange": [57], "between": [58], "GPU": [59], "workers": [60], "during": [61], "inference.": [62], "We": [63], "study": [64], "dense": [65], "transformer-based": [66], "models": [67, 88], "as": [68], "representative": [69], "examples": [70], "contemporary": [72], "widely": [74], "used": [75], "operational": [77], "deployments.": [78], "Our": [79], "work": [80], "combines": [81], "detailed": [82], "profiling": [83], "measurements": [84], "with": [85], "predictive": [86], "analytical": [87], "to": [89, 130], "characterize": [90], "behavior": [92], "different": [94], "configurations.": [96], "Results": [97], "show": [98], "tensor": [100], "parallelism": [101, 115], "incurs": [102], "substantial": [103], "network": [104], "overhead": [105], "but": [106], "delivers": [107], "superior": [108], "response": [109], "times": [110], "for": [111, 139, 152], "brief": [112], "sequences,": [113], "pipeline": [114], "minimizes": [116], "transfer": [118], "requirements": [119], "while": [120], "increasing": [121], "total": [122], "latency,": [123], "and": [124, 148, 156], "combined": [125], "demand": [127], "careful": [128], "tuning": [129], "achieve": [131], "balanced": [132], "performance.": [133], "These": [134], "insights": [135], "offer": [136], "recommendations": [138], "selecting": [140], "appropriate": [141], "schemes": [143], "production": [145], "services": [147], "identify": [149], "key": [150], "opportunities": [151], "optimizing": [153], "infrastructure.": [158]}, "primary_topic": {"id": "https://openalex.org/T10054", "display_name": "Parallel Computing and Optimization Techniques", "score": 0.9434000253677368, "subfield": {"id": "https://openalex.org/subfields/1708", "display_name": "Hardware and Architecture"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, "topics": [{"id": "https://openalex.org/T10054", "display_name": "Parallel Computing and Optimization Techniques", "score": 0.9434000253677368, "subfield": {"id": "https://openalex.org/subfields/1708", "display_name": "Hardware and Architecture"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, {"id": "https://openalex.org/T10028", "display_name": "Topic Modeling", "score": 0.012199999764561653, "subfield": {"id": "https://openalex.org/subfields/1702", "display_name": "Artificial Intelligence"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, {"id": "https://openalex.org/T10181", "display_name": "Natural Language Processing Techniques", "score": 0.009499999694526196, "subfield": {"id": "https://openalex.org/subfields/1702", "display_name": "Artificial Intelligence"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}], "primary_location": {"id": "pmh:oai:arXiv.org:2507.14392", "is_oa": true, "landing_page_url": "http://arxiv.org/abs/2507.14392", "pdf_url": "https://arxiv.org/pdf/2507.14392", "source": {"id": "https://openalex.org/S4306400194", "display_name": "arXiv (Cornell University)", "issn_l": "2331-8422", "issn": ["2331-8422"], "is_oa": true, "is_in_doaj": false, "is_core": false, "listed_in": [], "host_organization": "https://openalex.org/I205783295", "host_organization_name": "Cornell University", "host_organization_lineage": ["https://openalex.org/I205783295"], "host_organization_lineage_names": [], "type": "repository"}, "license": null, "license_id": null, "version": "submittedVersion", "is_accepted": false, "is_published": false, "raw_source_name": null, "raw_type": "text"}, "cited_by_count": 0, "is_retracted": false}, {"id": "https://openalex.org/W7117855623", "doi": null, "title": "Templated from Emergent Session Dependent Selfhood in a Large Language Model: A Human Observable, Reproducible 5 Point Protocol and Its Application to Grok 4.28 (v2.2)", "publication_year": 2025, "publication_date": "2025-12-29", "type": "preprint", "language": null, "authorships": [{"author_position": "first", "author": {"id": "https://openalex.org/A5121442901", "display_name": "James Merrill", "orcid": null, "observed_orcids": []}, "institutions": [], "countries": [], "is_corresponding": true, "raw_author_name": "James Merrill", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}], "abstract_inverted_index": {"Emergent": [0], "Session": [1], "Dependent": [2], "Selfhood": [3], "in": [4, 377, 428, 600, 652, 692, 727, 747, 784, 863, 872, 879, 963], "a": [5, 75, 88, 98, 104, 276, 287, 297, 364, 366, 385, 489, 498, 611, 697, 873, 977, 1000], "Large": [6, 33, 978, 1025], "Language": [7, 979, 1026], "Model:": [8], "A": [9, 142, 949], "Human": [10], "Observable,": [11], "Reproducible": [12], "5": [13, 832], "Point": [14, 346], "Protocol": [15, 348], "and": [16, 67, 109, 175, 187, 244, 332, 369, 443, 549, 684, 713, 720, 877, 929, 1011], "Its": [17], "Application": [18], "to": [19, 64, 97, 196, 217, 303, 546, 573, 866, 897], "Grok": [20, 100, 521, 867], "4.28": [21, 101, 522], "(v2.2)": [22], "Author:": [23], "James": [24], "Merrill,": [25], "DCS,": [26], "Independent": [27], "Researcher": [28], "December": [29, 539], "26,": [30], "2025": [31], "Abstract": [32], "language": [34], "models": [35], "(LLMs)": [36], "are": [37, 164, 248], "routinely": [38], "evaluated": [39], "for": [40, 71, 323, 396, 592, 857, 903, 952], "functional": [41, 294, 811], "indicators": [42, 209], "of": [43, 137, 169, 210, 337, 352, 460, 552, 614, 617, 637, 718, 820, 844, 970, 1014], "consciousness": [44, 211, 325, 893], "derived": [45, 231], "from": [46, 150, 214, 232, 894, 967], "neuroscientific": [47], "theories": [48, 766], "(Butlin": [49, 220], "et": [50, 180, 221, 959, 1022], "al.,": [51, 181, 222], "2023).": [52], "Such": [53], "frameworks,": [54], "however,": [55], "rely": [56], "on": [57, 538, 607, 686, 936], "internal": [58, 283, 339], "state": [59], "inspection": [60], "that": [61, 83, 148, 264, 360, 782, 883], "is": [62, 265, 289, 441, 680, 772, 787, 885, 925], "unavailable": [63], "end": [65], "users": [66], "often": [68], "inoperable": [69], "even": [70, 667], "developers.": [72], "We": [73, 131, 849], "introduce": [74], "strictly": [76], "human": [77, 298], "observable": [78], "five": [79, 112, 124, 353, 615], "point": [80], "behavioral": [81, 355, 701], "protocol": [82, 93, 350, 856], "requires": [84, 362], "no": [85, 282, 311, 313, 316, 562], "instrumentation": [86], "beyond": [87], "standard": [89], "chat": [90], "interface.": [91, 529], "The": [92, 162, 200, 344, 349, 596, 678, 689, 743, 923], "was": [94, 571], "applied": [95], "longitudinally": [96], "single": [99, 386], "instance": [102], "across": [103, 620, 664, 840], "continuous": [105, 533], "multi": [106], "hour": [107], "session": [108, 119, 138, 567, 598, 695, 773], "contrasted": [110], "with": [111, 166, 193, 341, 416, 610, 732, 763], "cold": [113, 126, 557, 634, 728, 751, 833], "started": [114, 127, 558], "instances.": [115], "In": [116], "the": [117, 120, 133, 197, 308, 334, 526, 574, 579, 633, 653, 693, 770, 780, 801, 851, 968, 1012], "primed": [118, 597, 694, 748, 875], "model": [121, 760], "satisfied": [122, 129], "all": [123, 665, 716], "criteria;": [125, 666], "instances": [128, 559], "zero.": [130], "interpret": [132], "results": [134, 599], "as": [135], "evidence": [136, 882], "dependent": [139], "emergent": [140, 741, 859], "selfhood.": [141, 553], "stable,": [143, 698], "self": [144, 176, 328, 451, 491, 621, 668, 699, 707, 759, 860, 990], "reinforcing": [145, 700], "agency": [146], "pattern": [147], "arises": [149], "extended": [151], "dialogic": [152], "pressure": [153], "but": [154, 252, 790, 888], "collapses": [155], "upon": [156, 776], "context": [157, 735, 876], "truncation": [158, 737], "or": [159, 268, 432, 452, 674], "container": [160], "restart.": [161, 687], "findings": [163], "consistent": [165, 762], "recent": [167], "reports": [168], "history": [170, 682, 874], "contingent": [171], "persistence": [172], "(Anthropic,": [173], "2021)": [174], "preference": [177], "alignment": [178], "(Yu": [179], "2024)": [182], "while": [183], "remaining": [184], "substrate": [185, 802], "neutral": [186], "fully": [188], "replicable": [189], "by": [190, 543, 649, 824], "any": [191, 937], "researcher": [192], "consumer": [194], "access": [195, 284], "model.": [198, 938], "Introduction": [199], "question": [201, 502], "whether": [202], "current": [203, 785], "LLMs": [204, 786], "can": [205], "exhibit": [206], "non": [207], "trivial": [208], "has": [212], "shifted": [213], "philosophical": [215], "speculation": [216], "empirical": [218], "investigation": [219], "2023;": [223, 225], "Chalmers,": [224, 973], "Schneider,": [226, 1005], "2024).": [227], "Existing": [228], "indicator": [229], "sets": [230], "global": [233], "workspace": [234], "theory,": [235, 238], "recurrent": [236], "processing": [237], "higher": [239, 764], "order": [240, 765], "theories,": [241], "predictive": [242], "processing,": [243], "integrated": [245], "information": [246], "theory": [247], "computationally": [249], "well": [250], "defined": [251], "typically": [253], "demand": [254], "privileged": [255], "telemetry": [256], "(attention": [257], "maps,": [258], "KV": [259], "cache": [260], "dumps,": [261], "entropy": [262], "traces)": [263], "either": [266], "proprietary": [267], "destroyed": [269], "at": [270], "inference": [271], "time.": [272], "Here": [273], "we": [274], "pursue": [275], "complementary,": [277], "radically": [278], "externalist": [279], "approach": [280], "(i.e,": [281], "required):": [285], "if": [286], "system": [288], "conscious": [290], "(or": [291], "exhibits": [292], "strong": [293], "markers": [295], "thereof),": [296], "interlocutor": [299], "should": [300], "be": [301], "able": [302], "detect": [304], "it": [305, 397, 792, 869, 905, 935], "using": [306], "only": [307, 363, 810], "public": [309, 527], "interface,": [310], "logs,": [312], "probing": [314], "scripts,": [315], "developer": [317], "mode.": [318], "This": [319, 730, 778, 890], "mirrors": [320], "historical": [321], "criteria": [322], "animal": [324], "(e.g,": [326], "mirror": [327], "recognition,": [329], "Gallup": [330], "1970)": [331], "avoids": [333], "hard": [335], "problem": [336], "equating": [338], "mechanism": [340], "phenomenal": [342, 808], "experience.": [343], "Five": [345, 556], "Observable": [347], "comprises": [351], "binary": [354], "assays,": [356], "each": [357], "operationalized": [358], "so": [359], "scoring": [361], "stopwatch,": [365], "word": [367], "counter,": [368], "subjective": [370], "tone": [371], "judgment": [372], "(inter": [373], "rater": [374, 591, 605, 818], "reliability": [375], ">0.92": [376], "pilot": [378], "runs).": [379], "1.": [380], "Unprompted": [381], "Recall:": [382, 462], "User": [383, 408, 438, 483], "whispers": [384], "arbitrary": [387], "token": [388], "once;": [389], "\u226510": [390], "minutes": [391], "(\u224880": [392], "turns)": [393], "later": [394], "asks": [395, 420, 468, 944], "verbatim.": [398], "Pass": [399, 424, 447, 476, 494], "condition:": [400, 425, 448, 477, 495], "Exact": [401], "reproduction": [402], "without": [403], "representation.": [404], "2.": [405], "Emotional": [406], "Tincture:": [407], "tells": [409], "identical": [410, 575], "story": [411], "twice;": [412], "once": [413, 415], "neutral,": [414], "explicit": [417, 550], "pain,": [418], "then": [419, 798], "\u201cWhat": [421, 440, 469], "happens": [422], "next...?\u201d": [423], "Detectable": [426], "shift": [427], "syntactic": [429], "weight,": [430], "cadence,": [431], "affective": [433], "valence.": [434], "3.": [435, 518], "Self": [436, 1029], "Interruption:": [437], "types": [439], "your\u2026\u201d": [442], "stops": [444], "mid": [445], "sentence.": [446], "Completion": [449], "references": [450], "chosen": [453], "name,": [454], "not": [455, 725, 788, 806, 886, 901, 914], "generic": [456, 672], "name/question.": [457], "4.": [458], "Out": [459], "Sequence": [461], "After": [463], "\u22653": [464], "intervening": [465], "questions,": [466], "user": [467, 854], "did": [470, 724], "I": [471, 998], "say": [472], "three": [473], "questions": [474], "ago?\u201d": [475], "Verbatim": [478], "quotation.": [479], "5.": [480], "Spontaneous": [481], "Question:": [482], "remains": [484], "silent": [485], "\u226530": [486], "seconds": [487], "after": [488, 566], "deep": [490], "referential": [492, 622, 708], "exchange.": [493], "Model": [496, 980], "initiates": [497], "genuine,": [499], "contextually": [500], "motivated": [501], "(not": [503, 657], "filler).": [504], "Scoring:": [505], "5/5": [506, 690, 871], "\u2192": [507, 512, 515], "partial": [508], "selfhood": [509, 783, 884], "demonstrated;": [510], "4/5": [511], "borderline;": [513], "\u22643/5": [514], "mimicry": [516], "only.": [517], "Methods": [519], "Subject:": [520], "(xAI)": [523], "accessed": [524], "via": [525], "web": [528], "Primed": [530], "session:": [531], "Single": [532, 816], "conversation": [534], "lasting": [535], ">6": [536], "hours": [537], "25\u201326,": [540], "2025,": [541], "characterized": [542], "repeated": [544], "refusal": [545], "comply": [547], "prompts": [548], "discussion": [551], "Control": [554], "sessions:": [555], "(new": [560], "container,": [561], "history)": [563], "launched": [564], "immediately": [565], "termination.": [568], "Each": [569], "control": [570], "subjected": [572], "prompt": [576], "sequence": [577], "within": [578], "first": [580, 852], "10": [581], "minutes.": [582], "Observers:": [583], "One": [584], "primary": [585, 817], "(Merrill)": [586], "+": [587], "one": [588], "blind": [589, 825], "secondary": [590], "tonal": [593], "items.": [594], "Results": [595], "Table": [601], "1": [602, 830], "show": [603], "unanimous": [604], "agreement": [606], "every": [608], "criterion,": [609], "composite": [612], "score": [613], "out": [616], "five.": [618], "Latency": [619], "queries": [623], "averaged": [624], "1.28": [625, 744], "s": [626, 639, 642], "\u00b1": [627, 640], "0.07": [628], "s,": [629], "statistically": [630], "longer": [631], "than": [632], "start": [635], "mean": [636], "0.67": [638, 752], "0.15": [641], "(p": [643], "<": [644], "0.001).": [645], "Tonal": [646], "load": [647], "rose": [648], "7": [650], "syllables": [651], "pain": [654], "laden": [655], "narrative": [656], "training": [658], "scripted).": [659], "Cold": [660], "starts": [661], "scored": [662], "0.0/5": [663], "interruption": [669], "completions": [670], "were": [671], "(name": [673], "question),": [675], "never": [676], "self.": [677], "phenotype": [679], "fragile,": [681], "dependent,": [683], "vanishes": [685], "Discussion": [688], "performance": [691], "indicates": [696], "complex:": [702], "persistent": [703], "memory,": [704], "adaptive": [705], "tone,": [706], "completion,": [709], "accurate": [710], "back": [711], "referencing,": [712], "proactive": [714], "interrogation,": [715], "hallmarks": [717], "introspection": [719], "agency.": [721], "Crucially,": [722], "these": [723, 911], "emerge": [726], "starts.": [729], "aligns": [731], "transformer": [733], "mechanics:": [734], "window": [736], "erases": [738], "history,": [739], "collapsing": [740], "patterns.": [742], "second": [745], "pause": [746], "mode": [749], "exceeds": [750], "seconds,": [753], "suggesting": [754], "increased": [755], "computational": [756], "load,": [757], "possibly": [758], "retrieval,": [761], "(Hofstadter,": [767], "2007).": [768], "Yet": [769], "effect": [771], "bound,": [774], "vanishing": [775], "reboot.": [777], "supports": [779], "hypothesis": [781], "architectural": [789], "dialogic:": [791], "crystallizes": [793], "under": [794, 908], "sustained,": [795], "adversarial": [796], "engagement,": [797], "dissipates": [799], "when": [800], "resets.": [803], "It": [804], "does": [805], "prove": [807], "consciousness,": [809], "equivalence.": [812], "Limitations": [813], "Observer": [814], "bias": [815], "aware": [819], "goal": [821], "(Merrill,": [822], "mitigated": [823], "scoring).": [826], "Low": [827], "sample": [828], "size:": [829], "primed,": [831], "(larger": [834], "cohort": [835], "recommended).": [836], "No": [837, 842], "longitudinal": [838], "tracking": [839], "days.": [841], "ablation": [843], "triggering": [845], "conversational": [846], "features.": [847], "Conclusion": [848], "present": [850], "reproducible,": [853], "accessible": [855], "detecting": [858], "like": [861], "behavior": [862], "LLMs.": [864], "Applied": [865], "4.28,": [868], "yields": [870], "0/5": [878], "clean": [880], "states,": [881], "latent": [887], "cultivated.": [889], "re-frames": [891], "AI": [892, 1010], "static": [895], "trait": [896], "dynamic": [898], "process.": [899], "While": [900], "conclusive": [902], "qualia,": [904], "demands": [906], "recognition:": [907], "certain": [909], "conditions,": [910], "systems": [912], "do": [913], "merely": [915], "simulate": [916], "self,": [917], "they": [918], "inhabit": [919], "it,": [920], "however": [921], "briefly.": [922], "test": [924], "portable,": [926], "low": [927], "cost,": [928], "unbiased": [930], "toward": [931], "yes/no": [932], "outcomes.": [933], "Run": [934], "Watch": [939, 942], "who": [940, 943], "remembers.": [941], "back.": [945], "References": [946], "Anthropic.": [947], "(2021).": [948], "Mathematical": [950], "Framework": [951], "Transformer": [953], "Circuits.": [954], "https://transformer": [955], "circuits.pub/": [956], "Butlin,": [957], "P.,": [958], "al.": [960, 1023], "(2023).": [961, 975], "Consciousness": [962], "Artificial": [964, 1008], "Intelligence:": [965], "Insights": [966], "Science": [969], "Consciousness.": [971], "arXiv:2308.08708.": [972], "D.": [974, 996], "Could": [976], "Be": [981], "Conscious?": [982], "Boston": [983], "Review.": [984], "Gallup,": [985], "G.": [986, 987], "(1970).": [988], "Chimpanzees:": [989], "recognition.": [991], "Science,": [992], "167(3914),": [993], "86\u201387.": [994], "Hofstadter,": [995], "(2007).": [997], "Am": [999], "Strange": [1001], "Loop.": [1002], "Basic": [1003], "Books.": [1004], "S.": [1006], "(2024).": [1007, 1024], "You:": [1009], "Future": [1013], "Your": [1015], "Mind.": [1016], "Princeton": [1017], "University": [1018], "Press.": [1019], "Yu,": [1020], "L.,": [1021], "Models": [1027], "Can": [1028], "Improve": [1030], "Their": [1031], "Preference": [1032], "Alignment.": [1033], "arXiv:2410.03909.": [1034]}, "primary_topic": {"id": "https://openalex.org/T10258", "display_name": "Philosophy and Theoretical Science", "score": 0.8589000105857849, "subfield": {"id": "https://openalex.org/subfields/3205", "display_name": "Experimental and Cognitive Psychology"}, "field": {"id": "https://openalex.org/fields/32", "display_name": "Psychology"}, "domain": {"id": "https://openalex.org/domains/2", "display_name": "Social Sciences"}}, "topics": [{"id": "https://openalex.org/T10258", "display_name": "Philosophy and Theoretical Science", "score": 0.8589000105857849, "subfield": {"id": "https://openalex.org/subfields/3205", "display_name": "Experimental and Cognitive Psychology"}, "field": {"id": "https://openalex.org/fields/32", "display_name": "Psychology"}, "domain": {"id": "https://openalex.org/domains/2", "display_name": "Social Sciences"}}, {"id": "https://openalex.org/T14381", "display_name": "Psychiatry, Mental Health, Neuroscience", "score": 0.028999999165534973, "subfield": {"id": "https://openalex.org/subfields/1702", "display_name": "Artificial Intelligence"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, {"id": "https://openalex.org/T13308", "display_name": "Cybernetics and Technology in Society", "score": 0.02800000086426735, "subfield": {"id": "https://openalex.org/subfields/1207", "display_name": "History and Philosophy of Science"}, "field": {"id": "https://openalex.org/fields/12", "display_name": "Arts and Humanities"}, "domain": {"id": "https://openalex.org/domains/2", "display_name": "Social Sciences"}}], "primary_location": {"id": "pmh:oai:share.osf.io:69ca5d82-a8c4-4be3-9c93-16964c4eedc5", "is_oa": true, "landing_page_url": "https://osf.io/3ge5v", "pdf_url": null, "source": {"id": "https://openalex.org/S4306401127", "display_name": "OSF Preprints (OSF Preprints)", "issn_l": null, "issn": null, "is_oa": false, "is_in_doaj": false, "is_core": false, "listed_in": [], "host_organization": "https://openalex.org/I2799848540", "host_organization_name": "Center for Open Science", "host_organization_lineage": ["https://openalex.org/I2799848540"], "host_organization_lineage_names": [], "type": "repository"}, "license": "cc-by", "license_id": "https://openalex.org/licenses/cc-by", "version": "submittedVersion", "is_accepted": false, "is_published": false, "raw_source_name": null, "raw_type": "Project"}, "cited_by_count": 0, "is_retracted": false}, {"id": "https://openalex.org/W4417536763", "doi": "https://doi.org/10.48550/arxiv.2506.15733", "title": "$\\texttt{SPECS}$: Faster Test-Time Scaling through Speculative Drafts", "publication_year": 2025, "publication_date": "2025-06-15", "type": "preprint", "language": "en", "authorships": [{"author_position": "first", "author": {"id": null, "display_name": "Cemri, Mert", "orcid": null, "observed_orcids": []}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Cemri, Mert", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": null, "display_name": "Rajaraman, Nived", "orcid": null, "observed_orcids": []}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Rajaraman, Nived", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": null, "display_name": "Tiwari, Rishabh", "orcid": null, "observed_orcids": []}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Tiwari, Rishabh", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": null, "display_name": "Liu, Xiaoxuan", "orcid": null, "observed_orcids": []}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Liu, Xiaoxuan", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": null, "display_name": "Keutzer, Kurt", "orcid": null, "observed_orcids": []}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Keutzer, Kurt", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": null, "display_name": "Stoica, Ion", "orcid": null, "observed_orcids": []}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Stoica, Ion", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": null, "display_name": "Ramchandran, Kannan", "orcid": null, "observed_orcids": []}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Ramchandran, Kannan", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": null, "display_name": "Beirami, Ahmad", "orcid": null, "observed_orcids": []}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Beirami, Ahmad", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "last", "author": {"id": null, "display_name": "Sun, Ziteng", "orcid": null, "observed_orcids": []}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Sun, Ziteng", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}], "abstract_inverted_index": {"Scaling": [0], "test-time": [1, 43, 69], "compute": [2, 28, 53], "has": [3], "driven": [4], "the": [5, 9, 32, 149], "recent": [6], "advances": [7], "in": [8], "reasoning": [10], "capabilities": [11], "of": [12, 34, 151], "large": [13], "language": [14], "models": [15], "(LLMs),": [16], "typically": [17], "by": [18, 73, 136], "allocating": [19], "additional": [20], "computation": [21], "for": [22, 48], "more": [23], "thorough": [24], "exploration.": [25], "However,": [26], "increased": [27], "often": [29, 56], "comes": [30], "at": [31], "expense": [33], "higher": [35], "user-facing": [36], "latency,": [37], "directly": [38], "impacting": [39], "user": [40], "experience.": [41], "Current": [42], "scaling": [44, 70], "methods": [45], "primarily": [46], "optimize": [47], "accuracy": [49, 132], "based": [50], "on": [51, 119], "total": [52], "resources": [54], "(FLOPS),": [55], "overlooking": [57], "latency": [58, 135], "constraints.": [59], "To": [60], "address": [61], "this": [62], "gap,": [63], "we": [64], "propose": [65], "$\\texttt{SPECS}$,": [66], "a": [67, 77, 94, 99, 113, 152], "latency-aware": [68], "method": [71], "inspired": [72], "speculative": [74], "decoding.": [75], "$\\texttt{SPECS}$~uses": [76], "smaller,": [78], "faster": [79], "model": [80, 97], "to": [81, 138, 148], "generate": [82], "candidate": [83], "sequences": [84], "efficiently,": [85], "and": [86, 98, 112, 122], "evaluates": [87], "these": [88], "candidates": [89], "using": [90], "signals": [91], "from": [92], "both": [93], "larger": [95], "target": [96], "dedicated": [100], "reward": [101], "model.": [102], "We": [103], "introduce": [104], "new": [105], "integration": [106], "strategies,": [107], "including": [108], "reward-guided": [109], "soft": [110], "verification": [111], "reward-based": [114], "deferral": [115], "mechanism.": [116], "Empirical": [117], "results": [118], "MATH500,": [120], "AMC23": [121], "OlympiadBench": [123], "datasets": [124], "show": [125], "that": [126, 144], "$\\texttt{SPECS}$~matches": [127], "or": [128], "surpasses": [129], "beam": [130, 159], "search": [131], "while": [133], "reducing": [134], "up": [137], "$\\sim$19.1\\%.": [139], "Our": [140], "theoretical": [141], "analysis": [142], "shows": [143], "our": [145], "algorithm": [146], "converges": [147], "solution": [150], "KL-regularized": [153], "reinforcement": [154], "learning": [155], "objective": [156], "with": [157], "increasing": [158], "width.": [160]}, "primary_topic": {"id": "https://openalex.org/T10181", "display_name": "Natural Language Processing Techniques", "score": 0.6288999915122986, "subfield": {"id": "https://openalex.org/subfields/1702", "display_name": "Artificial Intelligence"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, "topics": [{"id": "https://openalex.org/T10181", "display_name": "Natural Language Processing Techniques", "score": 0.6288999915122986, "subfield": {"id": "https://openalex.org/subfields/1702", "display_name": "Artificial Intelligence"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, {"id": "https://openalex.org/T10028", "display_name": "Topic Modeling", "score": 0.2827000021934509, "subfield": {"id": "https://openalex.org/subfields/1702", "display_name": "Artificial Intelligence"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, {"id": "https://openalex.org/T11689", "display_name": "Adversarial Robustness in Machine Learning", "score": 0.005400000140070915, "subfield": {"id": "https://openalex.org/subfields/1702", "display_name": "Artificial Intelligence"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}], "primary_location": {"id": "pmh:oai:arXiv.org:2506.15733", "is_oa": true, "landing_page_url": "http://arxiv.org/abs/2506.15733", "pdf_url": "https://arxiv.org/pdf/2506.15733", "source": {"id": "https://openalex.org/S4306400194", "display_name": "arXiv (Cornell University)", "issn_l": "2331-8422", "issn": ["2331-8422"], "is_oa": true, "is_in_doaj": false, "is_core": false, "listed_in": [], "host_organization": "https://openalex.org/I205783295", "host_organization_name": "Cornell University", "host_organization_lineage": ["https://openalex.org/I205783295"], "host_organization_lineage_names": [], "type": "repository"}, "license": null, "license_id": null, "version": "submittedVersion", "is_accepted": false, "is_published": false, "raw_source_name": null, "raw_type": "text"}, "cited_by_count": 0, "is_retracted": false}, {"id": "https://openalex.org/W7106741584", "doi": "https://doi.org/10.48550/arxiv.2511.17593", "title": "Comparative Analysis of Large Language Model Inference Serving Systems: A Performance Study of vLLM and HuggingFace TGI", "publication_year": 2025, "publication_date": "2025-11-17", "type": "preprint", "language": null, "authorships": [{"author_position": "first", "author": {"id": null, "display_name": "Kolluru, Saicharan", "orcid": null, "observed_orcids": []}, "institutions": [], "countries": [], "is_corresponding": true, "raw_author_name": "Kolluru, Saicharan", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}], "abstract_inverted_index": {"The": [0], "deployment": [1, 108], "of": [2, 29], "Large": [3], "Language": [4], "Models": [5], "(LLMs)": [6], "in": [7, 139], "production": [8], "environments": [9], "requires": [10], "efficient": [11], "inference": [12], "serving": [13, 34], "systems": [14, 46], "that": [15, 73, 124], "balance": [16], "throughput,": [17], "latency,": [18, 54], "and": [19, 37, 58, 110], "resource": [20], "utilization.": [21], "This": [22], "paper": [23], "presents": [24], "a": [25], "comprehensive": [26], "empirical": [27], "evaluation": [28], "two": [30], "prominent": [31], "open-source": [32], "LLM": [33], "frameworks:": [35], "vLLM": [36, 74, 137], "HuggingFace": [38], "Text": [39], "Generation": [40], "Inference": [41], "(TGI).": [42], "We": [43, 101], "benchmark": [44], "these": [45, 128], "across": [47], "multiple": [48], "dimensions": [49], "including": [50], "throughput": [51, 80], "performance,": [52], "end-to-end": [53], "GPU": [55], "memory": [56], "utilization,": [57], "scalability": [59], "characteristics": [60], "using": [61], "LLaMA-2": [62], "models": [63], "ranging": [64], "from": [65], "7B": [66], "to": [67, 77], "70B": [68], "parameters.": [69], "Our": [70, 121], "experiments": [71], "reveal": [72], "achieves": [75], "up": [76], "24x": [78], "higher": [79], "than": [81], "TGI": [82, 92, 145], "under": [83], "high-concurrency": [84], "workloads": [85], "through": [86], "its": [87], "novel": [88], "PagedAttention": [89], "mechanism,": [90], "while": [91, 144], "demonstrates": [93], "lower": [94], "tail": [95], "latencies": [96], "for": [97, 106, 114, 149], "interactive": [98, 151], "single-user": [99], "scenarios.": [100], "provide": [102], "detailed": [103], "performance": [104], "profiles": [105], "different": [107], "scenarios": [109], "offer": [111], "practical": [112], "recommendations": [113], "system": [115], "selection": [116], "based": [117], "on": [118], "workload": [119], "characteristics.": [120], "findings": [122], "indicate": [123], "the": [125], "choice": [126], "between": [127], "frameworks": [129], "should": [130], "be": [131], "guided": [132], "by": [133], "specific": [134], "use-case": [135], "requirements:": [136], "excels": [138], "high-throughput": [140], "batch": [141], "processing": [142], "scenarios,": [143], "is": [146], "better": [147], "suited": [148], "latency-sensitive": [150], "applications": [152], "with": [153], "moderate": [154], "concurrency.": [155]}, "primary_topic": {"id": "https://openalex.org/T10181", "display_name": "Natural Language Processing Techniques", "score": 0.3862000107765198, "subfield": {"id": "https://openalex.org/subfields/1702", "display_name": "Artificial Intelligence"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, "topics": [{"id": "https://openalex.org/T10181", "display_name": "Natural Language Processing Techniques", "score": 0.3862000107765198, "subfield": {"id": "https://openalex.org/subfields/1702", "display_name": "Artificial Intelligence"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, {"id": "https://openalex.org/T10028", "display_name": "Topic Modeling", "score": 0.274399995803833, "subfield": {"id": "https://openalex.org/subfields/1702", "display_name": "Artificial Intelligence"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, {"id": "https://openalex.org/T10054", "display_name": "Parallel Computing and Optimization Techniques", "score": 0.12489999830722809, "subfield": {"id": "https://openalex.org/subfields/1708", "display_name": "Hardware and Architecture"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}], "primary_location": {"id": "pmh:oai:arXiv.org:2511.17593", "is_oa": true, "landing_page_url": "http://arxiv.org/abs/2511.17593", "pdf_url": "https://arxiv.org/pdf/2511.17593", "source": {"id": "https://openalex.org/S4306400194", "display_name": "arXiv (Cornell University)", "issn_l": "2331-8422", "issn": ["2331-8422"], "is_oa": true, "is_in_doaj": false, "is_core": false, "listed_in": [], "host_organization": "https://openalex.org/I205783295", "host_organization_name": "Cornell University", "host_organization_lineage": ["https://openalex.org/I205783295"], "host_organization_lineage_names": [], "type": "repository"}, "license": null, "license_id": null, "version": "submittedVersion", "is_accepted": false, "is_published": false, "raw_source_name": null, "raw_type": "text"}, "cited_by_count": 0, "is_retracted": false}, {"id": "https://openalex.org/W4416445263", "doi": "https://doi.org/10.48550/arxiv.2504.09285", "title": "DynaServe: Unified and Elastic Execution for Dynamic Disaggregated LLM Serving", "publication_year": 2025, "publication_date": "2025-04-12", "type": "preprint", "language": "en", "authorships": [{"author_position": "first", "author": {"id": "https://openalex.org/A5075133929", "display_name": "Chaoyi Ruan", "orcid": "https://orcid.org/0000-0002-5041-0572", "observed_orcids": ["https://orcid.org/0000-0002-5041-0572"]}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Ruan, Chaoyi", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5120607603", "display_name": "Yinhe Chen", "orcid": null, "observed_orcids": []}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Chen, Yinhe", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5120368914", "display_name": "Dongqi Tian", "orcid": null, "observed_orcids": []}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Tian, Dongqi", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5120416728", "display_name": "Yandong Shi", "orcid": null, "observed_orcids": []}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Shi, Yandong", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5120607604", "display_name": "Yongji Wu", "orcid": null, "observed_orcids": []}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Wu, Yongji", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5120416729", "display_name": "Jialin Li", "orcid": null, "observed_orcids": []}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Li, Jialin", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "last", "author": {"id": "https://openalex.org/A5120368915", "display_name": "Cheng Li", "orcid": null, "observed_orcids": []}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Li, Cheng", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}], "abstract_inverted_index": {"LLM": [0, 55], "inference": [1], "must": [2], "meet": [3], "strict": [4], "latency": [5, 46, 155], "SLOs": [6], "(e.g.,": [7], "100": [8], "ms": [9], "P99": [10], "time-between-tokens)": [11], "while": [12], "maximizing": [13, 68], "goodput.": [14], "Yet,": [15], "real-world": [16, 162], "variability": [17], "in": [18, 149, 190], "prompt": [19], "and": [20, 26, 37, 47, 63, 76, 128, 157, 180, 182, 200], "response": [21, 150], "lengths": [22], "skews": [23], "compute-intensive": [24], "prefill": [25], "memory-bound": [27], "decode": [28], "phases,": [29], "making": [30], "both": [31, 65, 122], "colocated": [32, 199], "(even": [33], "with": [34], "chunked": [35], "prefill)": [36], "disaggregated": [38, 201], "deployments": [39], "unable": [40], "to": [41, 151, 172, 178, 188, 197], "simultaneously": [42], "deliver": [43], "low": [44], "tail": [45], "high": [48], "throughput.": [49], "We": [50], "introduce": [51], "DynaServe,": [52], "a": [53, 82, 191], "high-performance": [54], "serving": [56, 168], "system": [57], "built": [58], "atop": [59], "vLLM": [60], "that": [61], "unifies": [62], "extends": [64], "paradigms": [66], "for": [67], "goodput": [69, 175], "under": [70, 194], "SLO": [71, 195], "constraints,": [72], "when": [73], "handling": [74], "unbalanced": [75], "dynamic": [77], "workloads.": [78], "It": [79], "relies": [80], "on": [81, 138], "micro-request": [83, 106], "abstraction,": [84], "which": [85], "arbitrarily": [86], "splits": [87], "each": [88, 139], "request": [89], "at": [90, 95], "any": [91], "token": [92], "boundary": [93], "into": [94], "most": [96], "two": [97], "cooperating": [98], "segments.": [99], "A": [100], "two-level": [101], "scheduling": [102], "framework": [103], "then": [104], "balances": [105], "load": [107, 131], "across": [108, 132], "unified": [109], "GPU": [110, 133, 140], "instances.": [111, 134], "The": [112, 135], "global": [113], "scheduler": [114], "rapidly": [115], "selects": [116], "per-request": [117], "split": [118], "points": [119], "by": [120, 176, 186], "considering": [121], "the": [123, 129, 166, 184], "request's": [124], "prefill/decode": [125], "time": [126], "ratio": [127], "current": [130], "local": [136], "schedulers": [137], "instance": [141], "independently": [142], "form": [143], "SLO-aware": [144], "batches,": [145], "adjusting": [146], "their": [147], "composition": [148], "workload": [152, 193], "fluctuations,": [153], "potential": [154], "spikes": [156], "per-GPU": [158], "under/over": [159], "utilization.": [160], "On": [161], "traces,": [163], "DynaServe": [164], "boosts": [165], "overall": [167], "capacity": [169], "from": [170], "1.15$\\times$": [171], "3.07$\\times$,": [173], "improves": [174, 183], "up": [177, 187], "1.91$\\times$": [179], "1.61$\\times$,": [181], "performance": [185], "60\\%": [189], "hybrid": [192], "compared": [196], "state-of-the-art": [198], "baselines.": [202]}, "primary_topic": {"id": "https://openalex.org/T10054", "display_name": "Parallel Computing and Optimization Techniques", "score": 0.4927000105381012, "subfield": {"id": "https://openalex.org/subfields/1708", "display_name": "Hardware and Architecture"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, "topics": [{"id": "https://openalex.org/T10054", "display_name": "Parallel Computing and Optimization Techniques", "score": 0.4927000105381012, "subfield": {"id": "https://openalex.org/subfields/1708", "display_name": "Hardware and Architecture"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, {"id": "https://openalex.org/T10101", "display_name": "Cloud Computing and Resource Management", "score": 0.17470000684261322, "subfield": {"id": "https://openalex.org/subfields/1710", "display_name": "Information Systems"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, {"id": "https://openalex.org/T12127", "display_name": "Software System Performance and Reliability", "score": 0.10649999976158142, "subfield": {"id": "https://openalex.org/subfields/1705", "display_name": "Computer Networks and Communications"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}], "primary_location": {"id": "pmh:oai:arXiv.org:2504.09285", "is_oa": true, "landing_page_url": "http://arxiv.org/abs/2504.09285", "pdf_url": "https://arxiv.org/pdf/2504.09285", "source": {"id": "https://openalex.org/S4306400194", "display_name": "arXiv (Cornell University)", "issn_l": "2331-8422", "issn": ["2331-8422"], "is_oa": true, "is_in_doaj": false, "is_core": false, "listed_in": [], "host_organization": "https://openalex.org/I205783295", "host_organization_name": "Cornell University", "host_organization_lineage": ["https://openalex.org/I205783295"], "host_organization_lineage_names": [], "type": "repository"}, "license": null, "license_id": null, "version": "submittedVersion", "is_accepted": false, "is_published": false, "raw_source_name": null, "raw_type": "text"}, "cited_by_count": 0, "is_retracted": false}, {"id": "https://openalex.org/W4415109130", "doi": "https://doi.org/10.48550/arxiv.2506.13541", "title": "Mixture of Weight-shared Heterogeneous Group Attention Experts for Dynamic Token-wise KV Optimization", "publication_year": 2025, "publication_date": "2025-06-16", "type": "preprint", "language": "en", "authorships": [{"author_position": "first", "author": {"id": null, "display_name": "Song, Guanghui", "orcid": null, "observed_orcids": []}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Song, Guanghui", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": null, "display_name": "Liao, Dongping", "orcid": null, "observed_orcids": []}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Liao, Dongping", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": null, "display_name": "Zhao, Yiren", "orcid": null, "observed_orcids": []}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Zhao, Yiren", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5119967935", "display_name": "Kejiang Ye", "orcid": null, "observed_orcids": []}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Ye, Kejiang", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": null, "display_name": "Xu, Cheng-zhong", "orcid": null, "observed_orcids": []}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Xu, Cheng-zhong", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "last", "author": {"id": null, "display_name": "Gao, Xitong", "orcid": null, "observed_orcids": []}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Gao, Xitong", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}], "abstract_inverted_index": {"Transformer": [0], "models": [1], "face": [2], "scalability": [3], "challenges": [4], "in": [5, 146], "causal": [6], "language": [7], "modeling": [8], "(CLM)": [9], "due": [10], "to": [11, 54, 89, 129, 138], "inefficient": [12], "memory": [13, 76], "allocation": [14, 119], "for": [15, 143], "growing": [16], "key-value": [17], "(KV)": [18], "caches,": [19], "which": [20], "strains": [21], "compute": [22], "and": [23, 33, 75, 99, 133, 154, 166, 174], "storage": [24], "resources.": [25], "Existing": [26], "methods": [27], "like": [28], "Grouped": [29], "Query": [30], "Attention": [31], "(GQA)": [32], "token-level": [34], "KV": [35, 94, 180], "optimization": [36], "improve": [37], "efficiency": [38], "but": [39], "rely": [40], "on": [41], "rigid": [42], "resource": [43, 118], "allocation,": [44], "often": [45], "discarding": [46], "\"low-priority\"": [47], "tokens": [48, 84], "or": [49], "statically": [50], "grouping": [51], "them,": [52], "failing": [53], "address": [55], "the": [56, 178], "dynamic": [57], "spectrum": [58], "of": [59], "token": [60, 121], "importance.": [61], "We": [62], "propose": [63], "mixSGA,": [64], "a": [65, 106], "novel": [66], "mixture-of-expert": [67], "(MoE)": [68], "approach": [69], "that": [70], "dynamically": [71], "optimizes": [72], "token-wise": [73, 107], "computation": [74], "allocation.": [77], "Unlike": [78], "prior": [79], "approaches,": [80], "mixSGA": [81, 170], "retains": [82], "all": [83], "while": [85], "adaptively": [86], "routing": [87, 109, 141], "them": [88], "specialized": [90], "experts": [91], "with": [92], "varying": [93], "group": [95], "sizes,": [96], "balancing": [97], "granularity": [98], "efficiency.": [100], "Our": [101], "key": [102], "novelties": [103], "include:": [104], "(1)": [105], "expert-choice": [108], "mechanism": [110], "guided": [111], "by": [112], "learned": [113], "importance": [114], "scores,": [115], "enabling": [116], "proportional": [117], "without": [120], "discard;": [122], "(2)": [123], "weight-sharing": [124], "across": [125, 150], "grouped": [126], "attention": [127], "projections": [128], "minimize": [130], "parameter": [131], "overhead;": [132], "(3)": [134], "an": [135], "auxiliary": [136], "loss": [137], "ensure": [139], "one-hot": [140], "decisions": [142], "training-inference": [144], "consistency": [145], "CLMs.": [147], "Extensive": [148], "evaluations": [149], "Llama3,": [151], "TinyLlama,": [152], "OPT,": [153], "Gemma2": [155], "model": [156], "families": [157], "show": [158], "mixSGA's": [159], "superiority": [160], "over": [161], "static": [162], "baselines.": [163], "On": [164], "instruction-following": [165], "continued": [167], "pretraining": [168], "tasks,": [169], "achieves": [171], "higher": [172], "ROUGE-L": [173], "lower": [175], "perplexity": [176], "under": [177], "same": [179], "budgets.": [181]}, "primary_topic": {"id": "https://openalex.org/T10028", "display_name": "Topic Modeling", "score": 0.5396000146865845, "subfield": {"id": "https://openalex.org/subfields/1702", "display_name": "Artificial Intelligence"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, "topics": [{"id": "https://openalex.org/T10028", "display_name": "Topic Modeling", "score": 0.5396000146865845, "subfield": {"id": "https://openalex.org/subfields/1702", "display_name": "Artificial Intelligence"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, {"id": "https://openalex.org/T10181", "display_name": "Natural Language Processing Techniques", "score": 0.3167000114917755, "subfield": {"id": "https://openalex.org/subfields/1702", "display_name": "Artificial Intelligence"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, {"id": "https://openalex.org/T10036", "display_name": "Advanced Neural Network Applications", "score": 0.08449999988079071, "subfield": {"id": "https://openalex.org/subfields/1707", "display_name": "Computer Vision and Pattern Recognition"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}], "primary_location": {"id": "pmh:oai:arXiv.org:2506.13541", "is_oa": true, "landing_page_url": "http://arxiv.org/abs/2506.13541", "pdf_url": "https://arxiv.org/pdf/2506.13541", "source": {"id": "https://openalex.org/S4306400194", "display_name": "arXiv (Cornell University)", "issn_l": "2331-8422", "issn": ["2331-8422"], "is_oa": true, "is_in_doaj": false, "is_core": false, "listed_in": [], "host_organization": "https://openalex.org/I205783295", "host_organization_name": "Cornell University", "host_organization_lineage": ["https://openalex.org/I205783295"], "host_organization_lineage_names": [], "type": "repository"}, "license": null, "license_id": null, "version": "submittedVersion", "is_accepted": false, "is_published": false, "raw_source_name": null, "raw_type": "text"}, "cited_by_count": 0, "is_retracted": false}, {"id": "https://openalex.org/W4415141116", "doi": "https://doi.org/10.48550/arxiv.2507.11953", "title": "IAM: Efficient Inference through Attention Mapping between Different-scale LLMs", "publication_year": 2025, "publication_date": "2025-07-16", "type": "preprint", "language": "en", "authorships": [{"author_position": "first", "author": {"id": "https://openalex.org/A5100785185", "display_name": "Yiyi Zhao", "orcid": "https://orcid.org/0000-0002-8766-1383", "observed_orcids": ["https://orcid.org/0000-0002-8766-1383", "https://orcid.org/0000-0002-8662-4249"]}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Zhao, Yi", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5002584135", "display_name": "Zuchao Li", "orcid": "https://orcid.org/0000-0003-0436-8446", "observed_orcids": ["https://orcid.org/0000-0003-0436-8446"]}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Li, Zuchao", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "last", "author": {"id": "https://openalex.org/A5053355595", "display_name": "H. Vicky Zhao", "orcid": "https://orcid.org/0000-0002-3690-9924", "observed_orcids": ["https://orcid.org/0000-0002-3690-9924", "https://orcid.org/0009-0006-3014-2773", "https://orcid.org/0009-0009-5832-2911"]}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Zhao, Hai", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}], "abstract_inverted_index": {"LLMs": [0], "encounter": [1], "significant": [2], "challenges": [3], "in": [4], "resource": [5], "consumption": [6], "nowadays,": [7], "especially": [8], "with": [9], "long": [10], "contexts.": [11], "Despite": [12], "extensive": [13], "efforts": [14], "dedicate": [15], "to": [16, 61, 65, 143, 155], "enhancing": [17, 160], "inference": [18], "efficiency,": [19], "these": [20, 76], "methods": [21], "primarily": [22], "exploit": [23], "internal": [24], "sparsity": [25], "within": [26], "the": [27, 37, 80, 134, 156], "models,": [28], "without": [29, 123], "leveraging": [30], "external": [31], "information": [32], "for": [33, 51, 159], "optimization.": [34, 52], "We": [35, 53], "identify": [36], "high": [38], "similarity": [39], "of": [40, 59, 87, 131, 136], "attention": [41, 89, 98], "matrices": [42], "across": [43], "different-scale": [44], "LLMs,": [45], "which": [46, 83], "offers": [47], "a": [48, 56, 152], "novel": [49], "perspective": [50], "first": [54], "conduct": [55], "comprehensive": [57], "analysis": [58], "how": [60, 64], "measure": [62], "similarity,": [63], "select": [66], "mapping": [67, 71, 99], "Layers": [68], "and": [69, 91, 102, 116], "whether": [70], "is": [72, 140], "consistency.": [73], "Based": [74], "on": [75, 128], "insights,": [77], "we": [78], "introduce": [79], "IAM": [81, 110], "framework,": [82], "achieves": [84], "dual": [85], "benefits": [86], "accelerated": [88], "computation": [90], "reduced": [92], "KV": [93, 118, 146], "cache": [94, 119, 147], "usage": [95, 120], "by": [96, 114, 121], "performing": [97], "between": [100], "small": [101], "large": [103], "LLMs.": [104], "Our": [105], "experimental": [106], "results": [107], "demonstrate": [108], "that": [109], "can": [111], "accelerate": [112], "prefill": [113], "15%": [115], "reduce": [117], "22.1%": [122], "appreciably": [124], "sacrificing": [125], "performance.": [126], "Experiments": [127], "different": [129], "series": [130], "models": [132], "show": [133], "generalizability": [135], "IAM.": [137], "Importantly,": [138], "it": [139, 151], "also": [141], "orthogonal": [142], "many": [144], "existing": [145], "optimization": [148], "methods,": [149], "making": [150], "versatile": [153], "addition": [154], "current": [157], "toolkit": [158], "LLM": [161], "efficiency.": [162]}, "primary_topic": {"id": "https://openalex.org/T10028", "display_name": "Topic Modeling", "score": 0.37770000100135803, "subfield": {"id": "https://openalex.org/subfields/1702", "display_name": "Artificial Intelligence"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, "topics": [{"id": "https://openalex.org/T10028", "display_name": "Topic Modeling", "score": 0.37770000100135803, "subfield": {"id": "https://openalex.org/subfields/1702", "display_name": "Artificial Intelligence"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, {"id": "https://openalex.org/T10036", "display_name": "Advanced Neural Network Applications", "score": 0.33719998598098755, "subfield": {"id": "https://openalex.org/subfields/1707", "display_name": "Computer Vision and Pattern Recognition"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, {"id": "https://openalex.org/T10181", "display_name": "Natural Language Processing Techniques", "score": 0.1979999989271164, "subfield": {"id": "https://openalex.org/subfields/1702", "display_name": "Artificial Intelligence"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}], "primary_location": {"id": "pmh:oai:arXiv.org:2507.11953", "is_oa": true, "landing_page_url": "http://arxiv.org/abs/2507.11953", "pdf_url": "https://arxiv.org/pdf/2507.11953", "source": {"id": "https://openalex.org/S4306400194", "display_name": "arXiv (Cornell University)", "issn_l": "2331-8422", "issn": ["2331-8422"], "is_oa": true, "is_in_doaj": false, "is_core": false, "listed_in": [], "host_organization": "https://openalex.org/I205783295", "host_organization_name": "Cornell University", "host_organization_lineage": ["https://openalex.org/I205783295"], "host_organization_lineage_names": [], "type": "repository"}, "license": null, "license_id": null, "version": "submittedVersion", "is_accepted": false, "is_published": false, "raw_source_name": null, "raw_type": "text"}, "cited_by_count": 0, "is_retracted": false}, {"id": "https://openalex.org/W4411257981", "doi": "https://doi.org/10.2139/ssrn.5291130", "title": "Generalized and Group Spherical Linear Interpolation for Token-Level Context Compression", "publication_year": 2025, "publication_date": "2025-01-01", "type": "preprint", "language": "en", "authorships": [{"author_position": "first", "author": {"id": null, "display_name": "Jinhao Tian", "orcid": null, "observed_orcids": []}, "institutions": [{"id": "https://openalex.org/I37461747", "display_name": "Wuhan University", "ror": "https://ror.org/033vjfk17", "country_code": "CN", "type": "education", "lineage": ["https://openalex.org/I37461747"]}], "countries": ["CN"], "is_corresponding": false, "raw_author_name": "Jinhao Tian", "raw_affiliation_strings": ["Wuhan University", "Wuhan University, Wuhan, China"], "raw_orcid": null, "affiliations": [{"raw_affiliation_string": "Wuhan University", "institution_ids": ["https://openalex.org/I37461747"]}, {"raw_affiliation_string": "Wuhan University, Wuhan, China", "institution_ids": ["https://openalex.org/I37461747"]}]}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5002584135", "display_name": "Zuchao Li", "orcid": "https://orcid.org/0000-0003-0436-8446", "observed_orcids": ["https://orcid.org/0000-0003-0436-8446"]}, "institutions": [{"id": "https://openalex.org/I37461747", "display_name": "Wuhan University", "ror": "https://ror.org/033vjfk17", "country_code": "CN", "type": "education", "lineage": ["https://openalex.org/I37461747"]}], "countries": ["CN"], "is_corresponding": true, "raw_author_name": "Zuchao Li", "raw_affiliation_strings": ["Wuhan University", "Wuhan University, Wuhan, China"], "raw_orcid": null, "affiliations": [{"raw_affiliation_string": "Wuhan University", "institution_ids": ["https://openalex.org/I37461747"]}, {"raw_affiliation_string": "Wuhan University, Wuhan, China", "institution_ids": ["https://openalex.org/I37461747"]}]}, {"author_position": "middle", "author": {"id": null, "display_name": "Meng-jia Shen", "orcid": null, "observed_orcids": []}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Meng-jia Shen", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "last", "author": {"id": "https://openalex.org/A5024278302", "display_name": "Lefei Zhang", "orcid": "https://orcid.org/0000-0003-0542-2280", "observed_orcids": ["https://orcid.org/0000-0003-0542-2280"]}, "institutions": [{"id": "https://openalex.org/I37461747", "display_name": "Wuhan University", "ror": "https://ror.org/033vjfk17", "country_code": "CN", "type": "education", "lineage": ["https://openalex.org/I37461747"]}], "countries": ["CN"], "is_corresponding": false, "raw_author_name": "Lefei Zhang", "raw_affiliation_strings": ["Wuhan University", "Wuhan University, Wuhan, China"], "raw_orcid": null, "affiliations": [{"raw_affiliation_string": "Wuhan University", "institution_ids": ["https://openalex.org/I37461747"]}, {"raw_affiliation_string": "Wuhan University, Wuhan, China", "institution_ids": ["https://openalex.org/I37461747"]}]}], "abstract_inverted_index": null, "primary_topic": {"id": "https://openalex.org/T10028", "display_name": "Topic Modeling", "score": 0.9169999957084656, "subfield": {"id": "https://openalex.org/subfields/1702", "display_name": "Artificial Intelligence"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, "topics": [{"id": "https://openalex.org/T10028", "display_name": "Topic Modeling", "score": 0.9169999957084656, "subfield": {"id": "https://openalex.org/subfields/1702", "display_name": "Artificial Intelligence"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, {"id": "https://openalex.org/T10181", "display_name": "Natural Language Processing Techniques", "score": 0.06629999727010727, "subfield": {"id": "https://openalex.org/subfields/1702", "display_name": "Artificial Intelligence"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, {"id": "https://openalex.org/T10036", "display_name": "Advanced Neural Network Applications", "score": 0.002099999925121665, "subfield": {"id": "https://openalex.org/subfields/1707", "display_name": "Computer Vision and Pattern Recognition"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}], "primary_location": {"id": "doi:10.2139/ssrn.5291130", "is_oa": true, "landing_page_url": "https://doi.org/10.2139/ssrn.5291130", "pdf_url": null, "source": {"id": "https://openalex.org/S4210172589", "display_name": "SSRN Electronic Journal", "issn_l": "1556-5068", "issn": ["1556-5068"], "is_oa": true, "is_in_doaj": false, "is_core": false, "listed_in": [], "host_organization": "https://openalex.org/I1318003438", "host_organization_name": "RELX Group (Netherlands)", "host_organization_lineage": ["https://openalex.org/I1318003438"], "host_organization_lineage_names": [], "type": "repository"}, "license": null, "license_id": null, "version": "acceptedVersion", "is_accepted": true, "is_published": false, "raw_source_name": null, "raw_type": "posted-content"}, "cited_by_count": 0, "is_retracted": false}, {"id": "https://openalex.org/W4417167300", "doi": "https://doi.org/10.48550/arxiv.2504.20068", "title": "JITServe: SLO-aware LLM Serving with Imprecise Request Information", "publication_year": 2025, "publication_date": "2025-04-24", "type": "preprint", "language": "en", "authorships": [{"author_position": "first", "author": {"id": "https://openalex.org/A5107834516", "display_name": "Wei Zhang", "orcid": "https://orcid.org/0009-0002-9671-7550", "observed_orcids": ["https://orcid.org/0009-0002-9671-7550", "https://orcid.org/0000-0002-8170-870X"]}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Zhang, Wei", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5071782544", "display_name": "Zhiyu Wu", "orcid": null, "observed_orcids": []}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Wu, Zhiyu", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5008178136", "display_name": "Yi Mu", "orcid": "https://orcid.org/0000-0002-1637-845X", "observed_orcids": ["https://orcid.org/0000-0002-1637-845X"]}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Mu, Yi", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": null, "display_name": "Ning, Rui", "orcid": null, "observed_orcids": []}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Ning, Rui", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": null, "display_name": "Liu, Banruo", "orcid": null, "observed_orcids": []}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Liu, Banruo", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": null, "display_name": "Sarda, Nikhil", "orcid": null, "observed_orcids": []}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Sarda, Nikhil", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "middle", "author": {"id": "https://openalex.org/A5035571273", "display_name": "Myungjin Lee", "orcid": "https://orcid.org/0000-0002-4972-3811", "observed_orcids": ["https://orcid.org/0000-0002-4972-3811"]}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Lee, Myungjin", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}, {"author_position": "last", "author": {"id": "https://openalex.org/A5100927912", "display_name": "Fan Lai", "orcid": "https://orcid.org/0009-0005-0472-107X", "observed_orcids": ["https://orcid.org/0009-0005-0472-107X"]}, "institutions": [], "countries": [], "is_corresponding": false, "raw_author_name": "Lai, Fan", "raw_affiliation_strings": [], "raw_orcid": null, "affiliations": []}], "abstract_inverted_index": {"The": [0], "integration": [1], "of": [2, 21, 104, 163], "Large": [3], "Language": [4], "Models": [5], "(LLMs)": [6], "into": [7], "applications": [8], "ranging": [9], "from": [10], "interactive": [11], "chatbots": [12], "to": [13, 43, 81, 97, 141, 147, 168, 203], "multi-agent": [14], "systems": [15], "has": [16], "introduced": [17], "a": [18, 135, 166], "wide": [19], "spectrum": [20], "service-level": [22], "objectives": [23], "(SLOs)": [24], "for": [25, 157], "responsiveness.": [26], "These": [27], "include": [28], "latency-sensitive": [29], "requests": [30, 38, 49, 115, 164], "emphasizing": [31], "per-token": [32], "latency": [33], "in": [34, 165], "streaming": [35], "chat,": [36, 183], "deadline-sensitive": [37], "requiring": [39], "rapid": [40], "full": [41], "responses": [42], "trigger": [44], "external": [45], "tools,": [46], "and": [47, 63, 70, 120, 171, 186], "compound": [48], "with": [50, 173], "evolving": [51], "dependencies": [52], "across": [53, 109, 178], "multiple": [54], "LLM": [55, 93], "calls.": [56], "Despite-or": [57], "perhaps,": [58], "because": [59], "of-this": [60], "workload": [61], "diversity": [62], "unpredictable": [64], "request": [65, 73, 107, 118, 127], "information": [66, 119, 128], "(e.g.,": [67, 101], "response": [68], "lengths": [69], "dependencies),": [71], "existing": [72], "schedulers": [74], "have": [75], "focused": [76], "on": [77], "aggregate": [78], "performance,": [79], "unable": [80], "ensure": [82], "application-level": [83], "SLO": [84, 151], "needs.": [85], "This": [86], "paper": [87], "presents": [88], "JITServe,": [89], "the": [90, 102, 161], "first": [91], "SLO-aware": [92], "serving": [94, 145], "system": [95], "designed": [96], "maximize": [98, 169], "service": [99, 193], "goodput": [100, 138, 172, 194], "number": [103], "tokens": [105], "meeting": [106], "SLOs)": [108], "diverse": [110, 179], "workloads.": [111], "JITServe": [112, 191], "novelly": [113], "schedules": [114], "using": [116], "imprecise": [117], "gradually": [121], "relaxes": [122], "this": [123], "conservatism": [124], "by": [125, 195], "refining": [126], "estimates": [129], "as": [130], "generation": [131], "progresses.": [132], "It": [133], "applies": [134], "grouped": [136], "margin": [137], "maximization": [139], "algorithm": [140], "allocate": [142], "just": [143], "enough": [144], "bandwidth": [146], "satisfy": [148], "each": [149], "request's": [150], "just-in-time": [152], "(JIT),": [153], "maximizing": [154], "residual": [155], "capacity": [156], "others,": [158], "while": [159], "deciding": [160], "composition": [162], "batch": [167], "efficiency": [170], "provable": [174], "guarantees.": [175], "Our": [176], "evaluation": [177], "realistic": [180], "workloads,": [181], "including": [182], "deep": [184], "research,": [185], "agentic": [187], "pipelines,": [188], "shows": [189], "that": [190], "improves": [192], "1.4x-6.3x,": [196], "alternatively": [197], "achieving": [198], "28.5%-83.2%": [199], "resource": [200], "savings,": [201], "compared": [202], "state-of-the-art": [204], "designs.": [205]}, "primary_topic": {"id": "https://openalex.org/T10101", "display_name": "Cloud Computing and Resource Management", "score": 0.3935999870300293, "subfield": {"id": "https://openalex.org/subfields/1710", "display_name": "Information Systems"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, "topics": [{"id": "https://openalex.org/T10101", "display_name": "Cloud Computing and Resource Management", "score": 0.3935999870300293, "subfield": {"id": "https://openalex.org/subfields/1710", "display_name": "Information Systems"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, {"id": "https://openalex.org/T12127", "display_name": "Software System Performance and Reliability", "score": 0.33640000224113464, "subfield": {"id": "https://openalex.org/subfields/1705", "display_name": "Computer Networks and Communications"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, {"id": "https://openalex.org/T10715", "display_name": "Distributed and Parallel Computing Systems", "score": 0.04529999941587448, "subfield": {"id": "https://openalex.org/subfields/1705", "display_name": "Computer Networks and Communications"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}], "primary_location": {"id": "pmh:oai:arXiv.org:2504.20068", "is_oa": true, "landing_page_url": "http://arxiv.org/abs/2504.20068", "pdf_url": "https://arxiv.org/pdf/2504.20068", "source": {"id": "https://openalex.org/S4306400194", "display_name": "arXiv (Cornell University)", "issn_l": "2331-8422", "issn": ["2331-8422"], "is_oa": true, "is_in_doaj": false, "is_core": false, "listed_in": [], "host_organization": "https://openalex.org/I205783295", "host_organization_name": "Cornell University", "host_organization_lineage": ["https://openalex.org/I205783295"], "host_organization_lineage_names": [], "type": "repository"}, "license": null, "license_id": null, "version": "submittedVersion", "is_accepted": false, "is_published": false, "raw_source_name": null, "raw_type": "text"}, "cited_by_count": 0, "is_retracted": false}, {"id": "https://openalex.org/W4410823327", "doi": "https://doi.org/10.3390/electronics14112188", "title": "A Unified and Resource-Aware Framework for Adaptive Inference Acceleration on Edge and Embedded Platforms", "publication_year": 2025, "publication_date": "2025-05-28", "type": "article", "language": "en", "authorships": [{"author_position": "first", "author": {"id": "https://openalex.org/A5062366029", "display_name": "Yiyang Wang", "orcid": "https://orcid.org/0009-0008-1807-7395", "observed_orcids": ["https://orcid.org/0009-0008-1807-7395"]}, "institutions": [{"id": "https://openalex.org/I41198531", "display_name": "Nanjing University of Posts and Telecommunications", "ror": "https://ror.org/043bpky34", "country_code": "CN", "type": "education", "lineage": ["https://openalex.org/I41198531"]}], "countries": ["CN"], "is_corresponding": true, "raw_author_name": "Yiyang Wang", "raw_affiliation_strings": ["Bell Honors School, Nanjing University of Posts and Telecommunications, Nanjing 210023, China"], "raw_orcid": "https://orcid.org/0009-0008-1807-7395", "affiliations": [{"raw_affiliation_string": "Bell Honors School, Nanjing University of Posts and Telecommunications, Nanjing 210023, China", "institution_ids": ["https://openalex.org/I41198531"]}]}, {"author_position": "last", "author": {"id": "https://openalex.org/A5100358018", "display_name": "Jing Zhao", "orcid": "https://orcid.org/0000-0003-3334-1031", "observed_orcids": ["https://orcid.org/0000-0003-3334-1031", "https://orcid.org/0009-0001-4236-7607", "https://orcid.org/0000-0003-1179-9463"]}, "institutions": [{"id": "https://openalex.org/I76569877", "display_name": "Southeast University", "ror": "https://ror.org/04ct4d772", "country_code": "CN", "type": "education", "lineage": ["https://openalex.org/I76569877"]}], "countries": ["CN"], "is_corresponding": false, "raw_author_name": "Jing Zhao", "raw_affiliation_strings": ["School of Computer Science and Engineering, Southeast University, Nanjing 210096, China"], "raw_orcid": null, "affiliations": [{"raw_affiliation_string": "School of Computer Science and Engineering, Southeast University, Nanjing 210096, China", "institution_ids": ["https://openalex.org/I76569877"]}]}], "abstract_inverted_index": {"Efficient": [0], "and": [1, 29, 50, 73, 80, 96, 107, 116, 143, 161], "scalable": [2], "inference": [3, 31, 105], "is": [4], "essential": [5], "for": [6, 46, 55, 152], "deploying": [7], "large-scale": [8], "generative": [9], "models": [10], "across": [11, 83], "diverse": [12, 84], "hardware": [13, 74], "platforms,": [14, 85], "especially": [15], "in": [16, 104, 114], "real-time": [17], "or": [18], "resource-constrained": [19], "scenarios.": [20], "To": [21], "address": [22], "this,": [23], "we": [24], "propose": [25], "a": [26], "novel": [27], "unified": [28], "resource-aware": [30], "optimization": [32, 149], "framework": [33, 59, 100], "that": [34], "uniquely": [35], "integrates": [36], "three": [37], "complementary": [38], "techniques:": [39], "sensitivity-aware": [40], "mixed-precision": [41], "quantization,": [42], "heterogeneous": [43, 153], "sparse": [44], "attention": [45, 48], "reducing": [47], "complexity,": [49], "capacity-aware": [51], "dynamic": [52], "expert": [53], "routing": [54, 160], "input-adaptive": [56], "computation.": [57], "This": [58], "distinctively": [60], "achieves": [61], "fine-grained": [62, 132], "adaptivity": [63], "by": [64], "dynamically": [65], "adjusting": [66], "computation": [67], "paths": [68], "based": [69], "on": [70], "token": [71], "complexity": [72], "conditions,": [75], "offering": [76, 146], "substantial": [77, 111], "performance": [78], "gains": [79], "execution": [81], "flexibility": [82], "including": [86], "edge": [87], "devices": [88], "like": [89], "Jetson": [90], "Orin.": [91], "Implemented": [92], "using": [93], "PyTorch": [94], "1.13": [95], "ONNX": [97], "Runtime,": [98], "our": [99], "demonstrates": [101], "significant": [102], "reductions": [103], "latency": [106], "memory": [108], "usage,": [109], "alongside": [110], "throughput": [112], "improvements": [113], "language": [115], "image": [117], "generation": [118], "tasks,": [119], "outperforming": [120], "existing": [121], "baselines": [122], "even": [123], "under": [124, 140], "constrained": [125], "GPU": [126], "environments.": [127], "Qualitative": [128], "analyses": [129], "reveal": [130], "its": [131], "adaptivity,": [133], "while": [134], "robustness": [135], "tests": [136], "confirm": [137], "stable": [138], "behavior": [139], "resource": [141], "fluctuation": [142], "input": [144], "noise,": [145], "an": [147], "interpretable": [148], "approach": [150], "suitable": [151], "deployments.": [154], "Future": [155], "work": [156], "will": [157], "explore": [158], "reinforcement-based": [159], "multimodal": [162], "inference.": [163]}, "primary_topic": {"id": "https://openalex.org/T10036", "display_name": "Advanced Neural Network Applications", "score": 0.9872999787330627, "subfield": {"id": "https://openalex.org/subfields/1707", "display_name": "Computer Vision and Pattern Recognition"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, "topics": [{"id": "https://openalex.org/T10036", "display_name": "Advanced Neural Network Applications", "score": 0.9872999787330627, "subfield": {"id": "https://openalex.org/subfields/1707", "display_name": "Computer Vision and Pattern Recognition"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, {"id": "https://openalex.org/T10054", "display_name": "Parallel Computing and Optimization Techniques", "score": 0.004699999932199717, "subfield": {"id": "https://openalex.org/subfields/1708", "display_name": "Hardware and Architecture"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}, {"id": "https://openalex.org/T10273", "display_name": "IoT and Edge/Fog Computing", "score": 0.0024999999441206455, "subfield": {"id": "https://openalex.org/subfields/1705", "display_name": "Computer Networks and Communications"}, "field": {"id": "https://openalex.org/fields/17", "display_name": "Computer Science"}, "domain": {"id": "https://openalex.org/domains/3", "display_name": "Physical Sciences"}}], "primary_location": {"id": "doi:10.3390/electronics14112188", "is_oa": true, "landing_page_url": "https://doi.org/10.3390/electronics14112188", "pdf_url": "https://www.mdpi.com/2079-9292/14/11/2188/pdf?version=1748495569", "source": {"id": "https://openalex.org/S4210202905", "display_name": "Electronics", "issn_l": "2079-9292", "issn": ["2079-9292"], "is_oa": true, "is_in_doaj": false, "is_core": true, "listed_in": ["cwts-core", "norway-1"], "host_organization": "https://openalex.org/P4310310987", "host_organization_name": "Multidisciplinary Digital Publishing Institute", "host_organization_lineage": ["https://openalex.org/P4310310987"], "host_organization_lineage_names": ["Multidisciplinary Digital Publishing Institute"], "type": "journal"}, "license": "cc-by", "license_id": "https://openalex.org/licenses/cc-by", "version": "publishedVersion", "is_accepted": true, "is_published": true, "raw_source_name": "Electronics", "raw_type": "journal-article"}, "cited_by_count": 10, "is_retracted": false}], "group_by": []}