{"id":"https://openalex.org/W4417536800","doi":"https://doi.org/10.48550/arxiv.2506.15704","title":"Learn from the Past: Fast Sparse Indexing for Large Language Model Decoding","display_name":"Learn from the Past: Fast Sparse Indexing for Large Language Model Decoding","publication_year":2025,"publication_date":"2025-05-30","ids":{"openalex":"https://openalex.org/W4417536800","doi":"https://doi.org/10.48550/arxiv.2506.15704"},"language":"en","primary_location":{"id":"pmh:oai:arXiv.org:2506.15704","doi":null,"is_oa":true,"landing_page_url":"http://arxiv.org/abs/2506.15704","pdf_url":"https://arxiv.org/pdf/2506.15704","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"},"volume":null,"type":"preprint","indexed_in":["arxiv","datacite"],"open_access":{"is_oa":true,"oa_status":"green","oa_url":"https://arxiv.org/pdf/2506.15704","any_repository_has_fulltext":true},"authorships":[{"author_position":"first","author":{"id":null,"display_name":"Yao, Feiyu","orcid":null,"observed_orcids":[]},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Yao, Feiyu","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"last","author":{"id":null,"display_name":"Wang, Qian","orcid":null,"observed_orcids":[]},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Wang, Qian","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]}],"institutions":[],"countries_distinct_count":0,"institutions_distinct_count":0,"corresponding_author_ids":[],"corresponding_institution_ids":[],"apc_list":null,"apc_paid":null,"fwci":null,"has_fulltext":false,"cited_by_count":0,"citation_normalized_percentile":null,"cited_by_percentile_year":null,"biblio":{"volume":null,"issue":null,"first_page":null,"last_page":null},"is_retracted":false,"is_paratext":false,"is_xpac":false,"primary_topic":{"id":"https://openalex.org/T10028","display_name":"Topic Modeling","score":0.5077999830245972,"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":0.5077999830245972,"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":0.31299999356269836,"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":0.0738999992609024,"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"}}],"keywords":[{"id":"https://openalex.org/keywords/sparse-attention","display_name":"sparse attention","score":0.6190000176429749},{"id":"https://openalex.org/keywords/key-value-cache","display_name":"key-value cache","score":0.6290000081062317},{"id":"https://openalex.org/keywords/long-context-inference","display_name":"long-context inference","score":0.781000018119812},{"id":"https://openalex.org/keywords/gpu-memory-efficiency","display_name":"GPU memory efficiency","score":0.5130000114440918},{"id":"https://openalex.org/keywords/llm-inference-acceleration","display_name":"LLM inference acceleration","score":0.5419999957084656}],"concepts":[{"id":"https://openalex.org/C41008148","wikidata":"https://www.wikidata.org/wiki/Q21198","display_name":"Computer science","level":0,"score":0.848800003528595},{"id":"https://openalex.org/C68339613","wikidata":"https://www.wikidata.org/wiki/Q1549489","display_name":"Speedup","level":2,"score":0.784600019454956},{"id":"https://openalex.org/C75165309","wikidata":"https://www.wikidata.org/wiki/Q2258979","display_name":"Search engine indexing","level":2,"score":0.6908000111579895},{"id":"https://openalex.org/C57273362","wikidata":"https://www.wikidata.org/wiki/Q576722","display_name":"Decoding methods","level":2,"score":0.6812999844551086},{"id":"https://openalex.org/C2780513914","wikidata":"https://www.wikidata.org/wiki/Q18210350","display_name":"Bottleneck","level":2,"score":0.5626999735832214},{"id":"https://openalex.org/C137293760","wikidata":"https://www.wikidata.org/wiki/Q3621696","display_name":"Language model","level":2,"score":0.5084999799728394},{"id":"https://openalex.org/C165696696","wikidata":"https://www.wikidata.org/wiki/Q11287","display_name":"Exploit","level":2,"score":0.42879998683929443},{"id":"https://openalex.org/C26517878","wikidata":"https://www.wikidata.org/wiki/Q228039","display_name":"Key (lock)","level":2,"score":0.41499999165534973},{"id":"https://openalex.org/C56372850","wikidata":"https://www.wikidata.org/wiki/Q1050404","display_name":"Sparse matrix","level":3,"score":0.41370001435279846},{"id":"https://openalex.org/C45374587","wikidata":"https://www.wikidata.org/wiki/Q12525525","display_name":"Computation","level":2,"score":0.41269999742507935},{"id":"https://openalex.org/C173608175","wikidata":"https://www.wikidata.org/wiki/Q232661","display_name":"Parallel computing","level":1,"score":0.39820000529289246},{"id":"https://openalex.org/C2780451532","wikidata":"https://www.wikidata.org/wiki/Q759676","display_name":"Task (project management)","level":2,"score":0.3968000113964081},{"id":"https://openalex.org/C154945302","wikidata":"https://www.wikidata.org/wiki/Q11660","display_name":"Artificial intelligence","level":1,"score":0.37540000677108765},{"id":"https://openalex.org/C145108525","wikidata":"https://www.wikidata.org/wiki/Q656154","display_name":"Xeon","level":2,"score":0.3603000044822693},{"id":"https://openalex.org/C193969084","wikidata":"https://www.wikidata.org/wiki/Q7452500","display_name":"Sequential decoding","level":4,"score":0.2890999913215637},{"id":"https://openalex.org/C2781357197","wikidata":"https://www.wikidata.org/wiki/Q5757597","display_name":"High memory","level":2,"score":0.2775999903678894},{"id":"https://openalex.org/C2777317252","wikidata":"https://www.wikidata.org/wiki/Q18393516","display_name":"Rare events","level":2,"score":0.27239999175071716},{"id":"https://openalex.org/C11413529","wikidata":"https://www.wikidata.org/wiki/Q8366","display_name":"Algorithm","level":1,"score":0.27149999141693115},{"id":"https://openalex.org/C80444323","wikidata":"https://www.wikidata.org/wiki/Q2878974","display_name":"Theoretical computer science","level":1,"score":0.26579999923706055},{"id":"https://openalex.org/C150552126","wikidata":"https://www.wikidata.org/wiki/Q339387","display_name":"SIMD","level":2,"score":0.2612999975681305},{"id":"https://openalex.org/C119857082","wikidata":"https://www.wikidata.org/wiki/Q2539","display_name":"Machine learning","level":1,"score":0.25859999656677246},{"id":"https://openalex.org/C106515295","wikidata":"https://www.wikidata.org/wiki/Q26806595","display_name":"Parallel processing","level":2,"score":0.25540000200271606}],"mesh":[],"locations_count":2,"locations":[{"id":"pmh:oai:arXiv.org:2506.15704","doi":null,"is_oa":true,"landing_page_url":"http://arxiv.org/abs/2506.15704","pdf_url":"https://arxiv.org/pdf/2506.15704","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"},{"id":"doi:10.48550/arxiv.2506.15704","doi":"https://doi.org/10.48550/arxiv.2506.15704","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2506.15704","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":null,"raw_source_name":null,"raw_type":"Preprint"}],"best_oa_location":{"id":"pmh:oai:arXiv.org:2506.15704","doi":null,"is_oa":true,"landing_page_url":"http://arxiv.org/abs/2506.15704","pdf_url":"https://arxiv.org/pdf/2506.15704","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"},"sustainable_development_goals":[],"sustainable_development_goals_aurora":[],"study_designs":[],"awards":[],"funders":[],"has_content":{"grobid_xml":false,"pdf":false},"content_urls":null,"referenced_works_count":0,"referenced_works":[],"related_works":[],"abstract_inverted_index":{"As":[0],"large":[1],"language":[2],"models":[3],"(LLMs)":[4],"continue":[5],"to":[6,86,133,140,149,182],"support":[7],"increasingly":[8],"longer":[9],"contexts,":[10],"the":[11,69,88,103,152,156,172],"memory":[12,29],"demand":[13],"for":[14,45,105,155,226],"key-value":[15,47],"(KV)":[16],"caches":[17],"during":[18],"decoding":[19,79,94,227],"grows":[20],"rapidly,":[21],"becoming":[22],"a":[23,145,201,206,221],"critical":[24],"bottleneck":[25],"in":[26,60,92,127,229],"both":[27],"GPU":[28,199],"capacity":[30],"and":[31,63,136,188,200,223],"PCIe":[32],"bandwidth.":[33],"Sparse":[34,106],"attention":[35,42,120,129,187],"mechanisms":[36],"alleviate":[37],"this":[38,97],"issue":[39],"by":[40],"computing":[41],"weights":[43],"only":[44],"selected":[46],"pairs.":[48],"However,":[49],"their":[50],"indexing":[51,115],"computation":[52],"typically":[53],"requires":[54],"traversing":[55],"all":[56],"key":[57],"vectors,":[58],"resulting":[59],"significant":[61],"computational":[62],"data":[64],"transfer":[65],"overhead.":[66],"To":[67,96],"reduce":[68],"cost":[70],"of":[71,205],"index":[72],"retrieval,":[73],"existing":[74],"methods":[75],"often":[76],"treat":[77],"each":[78],"step":[80],"as":[81,167,171],"an":[82,108,196],"independent":[83],"process,":[84],"failing":[85],"exploit":[87],"temporal":[89],"correlations":[90],"embedded":[91],"historical":[93,119],"information.":[95],"end,":[98],"we":[99],"propose":[100],"LFPS(Learn":[101],"From":[102],"Past":[104],"Indexing),":[107],"acceleration":[109],"method":[110],"that":[111,178,218],"dynamically":[112],"constructs":[113],"sparse":[114],"candidates":[116],"based":[117],"on":[118,162,195],"patterns.":[121],"LFPS":[122,161,179,219],"captures":[123],"two":[124],"prevalent":[125],"trends":[126],"decoder":[128],"-vertical":[130],"patterns":[131,138],"(attending":[132,139],"fixed":[134],"positions)":[135,142],"slash":[137],"relative":[141],"-and":[143],"incorporates":[144],"positional":[146],"expansion":[147],"strategy":[148],"effectively":[150],"predict":[151],"Top-k":[153,193],"indices":[154],"current":[157],"step.":[158],"We":[159],"validate":[160],"challenging":[163],"long-context":[164,230],"benchmarks":[165],"such":[166],"LongBench-RULER,":[168],"using":[169],"Llama-3.1-8B-Instruct":[170],"base":[173],"model.":[174],"Experimental":[175],"results":[176,216],"show":[177],"achieves":[180],"up":[181],"22.8$\\times$":[183],"speedup":[184,190],"over":[185,191],"full":[186],"9.6$\\times$":[189],"exact":[192],"retrieval":[194],"RTX":[197],"4090":[198],"single":[202],"CPU":[203],"core":[204],"Xeon":[207],"Gold":[208],"6430,":[209],"respectively,":[210],"while":[211],"preserving":[212],"generation":[213],"accuracy.":[214],"These":[215],"demonstrate":[217],"offers":[220],"practical":[222],"efficient":[224],"solution":[225],"optimization":[228],"LLM":[231],"inference.":[232]},"counts_by_year":[],"updated_date":"2026-08-26T07:47:46.906454","created_date":"2025-10-10T00:00:00"}