{"id":"https://openalex.org/W7166886618","doi":"https://doi.org/10.18653/v1/2026.acl-long.353","title":"HCSpec: Two-Tier Horizontal Cascade Speculative Decoding for High-Efficiency Large Language Model Inference","display_name":"HCSpec: Two-Tier Horizontal Cascade Speculative Decoding for High-Efficiency Large Language Model Inference","publication_year":2026,"publication_date":"2026-01-01","ids":{"openalex":"https://openalex.org/W7166886618","doi":"https://doi.org/10.18653/v1/2026.acl-long.353"},"language":null,"primary_location":{"id":"doi:10.18653/v1/2026.acl-long.353","is_oa":true,"landing_page_url":"https://doi.org/10.18653/v1/2026.acl-long.353","pdf_url":"https://aclanthology.org/2026.acl-long.353.pdf","source":null,"license":"cc-by","license_id":"https://openalex.org/licenses/cc-by","version":"publishedVersion","is_accepted":true,"is_published":true,"raw_source_name":"Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)","raw_type":"proceedings-article"},"type":"conference-paper","indexed_in":["crossref"],"open_access":{"is_oa":true,"oa_status":"gold","oa_url":"https://aclanthology.org/2026.acl-long.353.pdf","any_repository_has_fulltext":null},"authorships":[{"author_position":"first","author":{"id":"https://openalex.org/A5139844370","display_name":"Yizhou Zhang","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Yizhou Zhang","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5139775855","display_name":"Siming Chen","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Siming Chen","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5139758402","display_name":"Hao Ye","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Hao Ye","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"last","author":{"id":"https://openalex.org/A5051943243","display_name":"Erhu Feng","orcid":"https://orcid.org/0009-0006-5957-3024"},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Erhu Feng","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":0.0,"has_fulltext":true,"cited_by_count":0,"citation_normalized_percentile":{"value":0.87480051,"is_in_top_1_percent":false,"is_in_top_10_percent":false},"cited_by_percentile_year":null,"biblio":{"volume":null,"issue":null,"first_page":"7773","last_page":"7783"},"is_retracted":false,"is_paratext":false,"is_xpac":false,"primary_topic":{"id":"https://openalex.org/T10028","display_name":"Topic Modeling","score":0.2702000141143799,"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":0.2702000141143799,"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":0.060600001364946365,"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":0.05889999866485596,"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"}}],"keywords":[{"id":"https://openalex.org/keywords/decoding-methods","display_name":"Decoding methods","score":0.544700026512146},{"id":"https://openalex.org/keywords/language-model","display_name":"Language model","score":0.5250999927520752},{"id":"https://openalex.org/keywords/cascade","display_name":"Cascade","score":0.4697999954223633},{"id":"https://openalex.org/keywords/inference","display_name":"Inference","score":0.4194999933242798},{"id":"https://openalex.org/keywords/encoding","display_name":"Encoding (memory)","score":0.33709999918937683},{"id":"https://openalex.org/keywords/natural-language","display_name":"Natural language","score":0.3165000081062317}],"concepts":[{"id":"https://openalex.org/C41008148","wikidata":"https://www.wikidata.org/wiki/Q21198","display_name":"Computer science","level":0,"score":0.6686999797821045},{"id":"https://openalex.org/C57273362","wikidata":"https://www.wikidata.org/wiki/Q576722","display_name":"Decoding methods","level":2,"score":0.544700026512146},{"id":"https://openalex.org/C137293760","wikidata":"https://www.wikidata.org/wiki/Q3621696","display_name":"Language model","level":2,"score":0.5250999927520752},{"id":"https://openalex.org/C34146451","wikidata":"https://www.wikidata.org/wiki/Q5048094","display_name":"Cascade","level":2,"score":0.4697999954223633},{"id":"https://openalex.org/C154945302","wikidata":"https://www.wikidata.org/wiki/Q11660","display_name":"Artificial intelligence","level":1,"score":0.44130000472068787},{"id":"https://openalex.org/C204321447","wikidata":"https://www.wikidata.org/wiki/Q30642","display_name":"Natural language processing","level":1,"score":0.4341000020503998},{"id":"https://openalex.org/C2776214188","wikidata":"https://www.wikidata.org/wiki/Q408386","display_name":"Inference","level":2,"score":0.4194999933242798},{"id":"https://openalex.org/C11413529","wikidata":"https://www.wikidata.org/wiki/Q8366","display_name":"Algorithm","level":1,"score":0.35010001063346863},{"id":"https://openalex.org/C125411270","wikidata":"https://www.wikidata.org/wiki/Q18653","display_name":"Encoding (memory)","level":2,"score":0.33709999918937683},{"id":"https://openalex.org/C195324797","wikidata":"https://www.wikidata.org/wiki/Q33742","display_name":"Natural language","level":2,"score":0.3165000081062317},{"id":"https://openalex.org/C28490314","wikidata":"https://www.wikidata.org/wiki/Q189436","display_name":"Speech recognition","level":1,"score":0.3084999918937683},{"id":"https://openalex.org/C26517878","wikidata":"https://www.wikidata.org/wiki/Q228039","display_name":"Key (lock)","level":2,"score":0.2856999933719635},{"id":"https://openalex.org/C67186912","wikidata":"https://www.wikidata.org/wiki/Q367664","display_name":"Data modeling","level":2,"score":0.2824000120162964},{"id":"https://openalex.org/C184337299","wikidata":"https://www.wikidata.org/wiki/Q1437428","display_name":"Semantics (computer science)","level":2,"score":0.25859999656677246}],"mesh":[],"locations_count":1,"locations":[{"id":"doi:10.18653/v1/2026.acl-long.353","is_oa":true,"landing_page_url":"https://doi.org/10.18653/v1/2026.acl-long.353","pdf_url":"https://aclanthology.org/2026.acl-long.353.pdf","source":null,"license":"cc-by","license_id":"https://openalex.org/licenses/cc-by","version":"publishedVersion","is_accepted":true,"is_published":true,"raw_source_name":"Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)","raw_type":"proceedings-article"}],"best_oa_location":{"id":"doi:10.18653/v1/2026.acl-long.353","is_oa":true,"landing_page_url":"https://doi.org/10.18653/v1/2026.acl-long.353","pdf_url":"https://aclanthology.org/2026.acl-long.353.pdf","source":null,"license":"cc-by","license_id":"https://openalex.org/licenses/cc-by","version":"publishedVersion","is_accepted":true,"is_published":true,"raw_source_name":"Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)","raw_type":"proceedings-article"},"sustainable_development_goals":[],"awards":[],"funders":[],"has_content":{"grobid_xml":true,"pdf":true},"content_urls":{"pdf":"https://content.openalex.org/works/W7166886618.pdf","grobid_xml":"https://content.openalex.org/works/W7166886618.grobid-xml"},"referenced_works_count":0,"referenced_works":[],"related_works":[],"abstract_inverted_index":{"Speculative":[0,64],"decoding":[1],"accelerates":[2],"large":[3],"language":[4],"model":[5,12],"(LLM)":[6],"inference":[7,131],"by":[8,20,91],"using":[9],"a":[10,30,41,67,77,83,104],"draft":[11,27,74],"to":[13,38,152],"propose":[14,60],"token":[15],"candidates":[16],"for":[17,40,113],"parallel":[18],"verification":[19],"the":[21,45,53,100,138,163],"target":[22],"model.However,":[23],"current":[24],"state-of-the-art":[25,140],"self-distilled":[26],"models":[28,124],"adopt":[29],"homogeneous":[31],"architecture":[32],"across":[33],"all":[34],"drafting":[35,49,118],"positions,":[36],"failing":[37],"account":[39],"critical":[42],"empirical":[43],"observation:":[44],"expected":[46],"utility":[47],"of":[48],"decays":[50],"rapidly":[51],"after":[52],"initial":[54],"positions.To":[55],"exploit":[56],"this":[57],"imbalance,":[58],"we":[59],"Two-tier":[61],"Horizontal":[62],"Cascade":[63],"Decoding":[65],"(HCSpec),":[66],"novel":[68],"framework":[69],"that":[70,87,108,134],"organizes":[71],"heterogeneous,":[72],"positionspecialized":[73],"modules":[75],"into":[76],"horizontal":[78],"cascade.The":[79],"first":[80],"tier":[81,102],"employs":[82],"dual-layer,":[84],"dual-path":[85],"transformer":[86,107],"enhances":[88],"early-step":[89],"fidelity":[90],"decoupling":[92],"token-logit":[93],"prediction":[94],"from":[95],"recurrent":[96],"feature":[97],"propagation,":[98],"while":[99],"second":[101],"adopts":[103],"lightweight":[105],"single-layer":[106],"deliberately":[109],"trades":[110],"marginal":[111],"accuracy":[112],"improved":[114],"efficiency":[115],"at":[116],"later":[117],"steps.Extensive":[119],"experiments":[120],"on":[121],"Qwen":[122],"series":[123],"and":[125,129,149],"Llama3.1-8B-Instruct,across":[126],"multiple":[127],"tasks":[128],"diverse":[130],"configurations,":[132],"demonstrate":[133],"HCSpec":[135],"consistently":[136],"outperforms":[137],"previous":[139],"(EAGLE-3).It":[141],"delivers":[142],"15-30%":[143],"higher":[144],"end-to-end":[145],"speedup":[146],"over":[147,155],"EAGLE-3":[148],"achieves":[150],"up":[151],"3.72x":[153],"acceleration":[154],"vanilla":[156],"autoregressive":[157],"decoding.Our":[158],"code":[159],"is":[160],"provided":[161],"in":[162],"supplementary":[164],"materials.":[165]},"counts_by_year":[],"updated_date":"2026-07-29T14:22:42.915294","created_date":"2026-07-02T00:00:00"}