{"id":"https://openalex.org/W7166809609","doi":"https://doi.org/10.18653/v1/2026.findings-acl.421","title":"MARS: Unleashing the Power of Speculative Decoding via Margin-Aware Verification","display_name":"MARS: Unleashing the Power of Speculative Decoding via Margin-Aware Verification","publication_year":2026,"publication_date":"2026-01-01","ids":{"openalex":"https://openalex.org/W7166809609","doi":"https://doi.org/10.18653/v1/2026.findings-acl.421"},"language":"en","primary_location":{"id":"doi:10.18653/v1/2026.findings-acl.421","doi":"https://doi.org/10.18653/v1/2026.findings-acl.421","is_oa":true,"landing_page_url":"https://doi.org/10.18653/v1/2026.findings-acl.421","pdf_url":"https://aclanthology.org/2026.findings-acl.421.pdf","source":{"id":"https://openalex.org/S4363605144","display_name":"Findings of the Association for Computational Linguistics: ACL (Findings of ACL)","issn_l":null,"issn":null,"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":"conference"},"license":"cc-by","license_id":"https://openalex.org/licenses/cc-by","version":"publishedVersion","is_accepted":true,"is_published":true,"raw_source_name":"Findings of the Association for Computational Linguistics: ACL 2026","raw_type":"proceedings-article"},"volume":null,"type":"conference-paper","indexed_in":["crossref"],"open_access":{"is_oa":true,"oa_status":"hybrid","oa_url":"https://aclanthology.org/2026.findings-acl.421.pdf","any_repository_has_fulltext":false},"authorships":[{"author_position":"first","author":{"id":"https://openalex.org/A5141589341","display_name":"Jingwei Song","orcid":null,"observed_orcids":[]},"institutions":[{"id":"https://openalex.org/I889458895","display_name":"University of Hong Kong","ror":"https://ror.org/02zhqgq86","country_code":"HK","type":"education","lineage":["https://openalex.org/I889458895"]}],"countries":["HK"],"is_corresponding":false,"raw_author_name":"Jingwei Song","raw_affiliation_strings":["The University of Hong Kong ,"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"The University of Hong Kong ,","institution_ids":["https://openalex.org/I889458895"]}]},{"author_position":"middle","author":{"id":"https://openalex.org/A5139844387","display_name":"Xinyu Wang","orcid":null,"observed_orcids":[]},"institutions":[{"id":"https://openalex.org/I5023651","display_name":"McGill University","ror":"https://ror.org/01pxwe438","country_code":"CA","type":"education","lineage":["https://openalex.org/I5023651"]}],"countries":["CA"],"is_corresponding":false,"raw_author_name":"Xinyu Wang","raw_affiliation_strings":["McGill University ,"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"McGill University ,","institution_ids":["https://openalex.org/I5023651"]}]},{"author_position":"middle","author":{"id":"https://openalex.org/A5038231583","display_name":"Hanbin Wang","orcid":"https://orcid.org/0000-0001-6927-2934","observed_orcids":["https://orcid.org/0000-0001-6927-2934"]},"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":"Hanbin Wang","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/A5153447325","display_name":"Xiaoxuan Lei","orcid":null,"observed_orcids":[]},"institutions":[{"id":"https://openalex.org/I5023651","display_name":"McGill University","ror":"https://ror.org/01pxwe438","country_code":"CA","type":"education","lineage":["https://openalex.org/I5023651"]}],"countries":["CA"],"is_corresponding":false,"raw_author_name":"Xiaoxuan Lei","raw_affiliation_strings":["McGill University ,"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"McGill University ,","institution_ids":["https://openalex.org/I5023651"]}]},{"author_position":"middle","author":{"id":"https://openalex.org/A5139762901","display_name":"Tianyu Shi","orcid":null,"observed_orcids":[]},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Tianyu Shi","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5139840792","display_name":"Shixin Han","orcid":null,"observed_orcids":[]},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Shixin Han","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5101576895","display_name":"Eric Yang","orcid":"https://orcid.org/0000-0003-0781-9863","observed_orcids":["https://orcid.org/0000-0003-0781-9863"]},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Eric Yang","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5062087251","display_name":"Xiao-Wen Chang","orcid":"https://orcid.org/0000-0001-9251-2959","observed_orcids":["https://orcid.org/0000-0001-9251-2959"]},"institutions":[{"id":"https://openalex.org/I5023651","display_name":"McGill University","ror":"https://ror.org/01pxwe438","country_code":"CA","type":"education","lineage":["https://openalex.org/I5023651"]}],"countries":["CA"],"is_corresponding":false,"raw_author_name":"Xiao-Wen Chang","raw_affiliation_strings":["McGill University ,"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"McGill University ,","institution_ids":["https://openalex.org/I5023651"]}]},{"author_position":"last","author":{"id":"https://openalex.org/A5119027766","display_name":"Lynn Ai","orcid":null,"observed_orcids":[]},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Lynn Ai","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]}],"institutions":[],"countries_distinct_count":3,"institutions_distinct_count":3,"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":0.44181038,"is_in_top_1_percent":false,"is_in_top_10_percent":false},"cited_by_percentile_year":null,"biblio":{"volume":null,"issue":null,"first_page":"8652","last_page":"8663"},"is_retracted":false,"is_paratext":false,"is_xpac":false,"primary_topic":{"id":"https://openalex.org/T10181","display_name":"Natural Language Processing Techniques","score":0.8574000000953674,"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":0.8574000000953674,"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":0.12049999833106995,"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":0.003100000089034438,"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/speculative-decoding","display_name":"speculative decoding","score":0.9959999918937683},{"id":"https://openalex.org/keywords/large-language-model-inference","display_name":"large language model inference","score":0.7609999775886536},{"id":"https://openalex.org/keywords/decision-stability","display_name":"decision stability","score":0.5370000004768372},{"id":"https://openalex.org/keywords/logit","display_name":"logit","score":0.6150000095367432},{"id":"https://openalex.org/keywords/rejection-sampling","display_name":"rejection sampling","score":0.7210000157356262},{"id":"https://openalex.org/keywords/inference-acceleration","display_name":"inference acceleration","score":0.628000020980835},{"id":"https://openalex.org/keywords/autoregressive-generation","display_name":"autoregressive generation","score":0.6959999799728394}],"concepts":[{"id":"https://openalex.org/C41008148","wikidata":"https://www.wikidata.org/wiki/Q21198","display_name":"Computer science","level":0,"score":0.6635000109672546},{"id":"https://openalex.org/C163258240","wikidata":"https://www.wikidata.org/wiki/Q25342","display_name":"Power (physics)","level":2,"score":0.5712000131607056},{"id":"https://openalex.org/C57273362","wikidata":"https://www.wikidata.org/wiki/Q576722","display_name":"Decoding methods","level":2,"score":0.5221999883651733},{"id":"https://openalex.org/C125411270","wikidata":"https://www.wikidata.org/wiki/Q18653","display_name":"Encoding (memory)","level":2,"score":0.375},{"id":"https://openalex.org/C26517878","wikidata":"https://www.wikidata.org/wiki/Q228039","display_name":"Key (lock)","level":2,"score":0.3303999900817871},{"id":"https://openalex.org/C79403827","wikidata":"https://www.wikidata.org/wiki/Q3988","display_name":"Real-time computing","level":1,"score":0.2736999988555908},{"id":"https://openalex.org/C168167062","wikidata":"https://www.wikidata.org/wiki/Q1117970","display_name":"Component (thermodynamics)","level":2,"score":0.25679999589920044},{"id":"https://openalex.org/C149635348","wikidata":"https://www.wikidata.org/wiki/Q193040","display_name":"Embedded system","level":1,"score":0.25429999828338623},{"id":"https://openalex.org/C77618280","wikidata":"https://www.wikidata.org/wiki/Q1155772","display_name":"Scheme (mathematics)","level":2,"score":0.24410000443458557},{"id":"https://openalex.org/C99498987","wikidata":"https://www.wikidata.org/wiki/Q2210247","display_name":"Noise (video)","level":3,"score":0.22179999947547913}],"mesh":[],"locations_count":1,"locations":[{"id":"doi:10.18653/v1/2026.findings-acl.421","doi":"https://doi.org/10.18653/v1/2026.findings-acl.421","is_oa":true,"landing_page_url":"https://doi.org/10.18653/v1/2026.findings-acl.421","pdf_url":"https://aclanthology.org/2026.findings-acl.421.pdf","source":{"id":"https://openalex.org/S4363605144","display_name":"Findings of the Association for Computational Linguistics: ACL (Findings of ACL)","issn_l":null,"issn":null,"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":"conference"},"license":"cc-by","license_id":"https://openalex.org/licenses/cc-by","version":"publishedVersion","is_accepted":true,"is_published":true,"raw_source_name":"Findings of the Association for Computational Linguistics: ACL 2026","raw_type":"proceedings-article"}],"best_oa_location":{"id":"doi:10.18653/v1/2026.findings-acl.421","doi":"https://doi.org/10.18653/v1/2026.findings-acl.421","is_oa":true,"landing_page_url":"https://doi.org/10.18653/v1/2026.findings-acl.421","pdf_url":"https://aclanthology.org/2026.findings-acl.421.pdf","source":{"id":"https://openalex.org/S4363605144","display_name":"Findings of the Association for Computational Linguistics: ACL (Findings of ACL)","issn_l":null,"issn":null,"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":"conference"},"license":"cc-by","license_id":"https://openalex.org/licenses/cc-by","version":"publishedVersion","is_accepted":true,"is_published":true,"raw_source_name":"Findings of the Association for Computational Linguistics: ACL 2026","raw_type":"proceedings-article"},"sustainable_development_goals":[],"sustainable_development_goals_aurora":[{"display_name":"Reduced inequalities","score":0.40035223960876465,"id":"https://metadata.un.org/sdg/10"}],"study_designs":[],"awards":[],"funders":[],"has_content":{"pdf":true,"grobid_xml":true},"content_urls":{"pdf":"https://content.openalex.org/works/W7166809609.pdf","grobid_xml":"https://content.openalex.org/works/W7166809609.grobid-xml"},"referenced_works_count":0,"referenced_works":[],"related_works":["https://openalex.org/W7166839924","https://openalex.org/W7083826980","https://openalex.org/W4414581587","https://openalex.org/W4407308985","https://openalex.org/W7204895662","https://openalex.org/W7164366554","https://openalex.org/W7127086585","https://openalex.org/W4412887986","https://openalex.org/W7171842053","https://openalex.org/W4407091762"],"abstract_inverted_index":{"Speculative":[0,84],"Decoding":[1],"(SD)":[2],"accelerates":[3],"autoregressive":[4],"large":[5],"language":[6],"model":[7,53,141],"(LLM)":[8],"inference":[9,156],"by":[10,20],"decoupling":[11],"generation":[12,163],"and":[13,88,112,129,154],"verification.While":[14],"recent":[15],"methods":[16],"improve":[17],"draft":[18],"quality":[19,164],"tightly":[21],"coupling":[22],"the":[23,26,29,51,95,109,122,126],"drafter":[24],"with":[25,133],"target":[27,52,96,110],"model,":[28],"verification":[30,90,102,118,127],"mechanism":[31],"itself":[32],"remains":[33],"largely":[34],"unchanged,":[35],"relying":[36],"on":[37,103],"strict":[38,117],"token-level":[39],"rejection":[40,114],"sampling.In":[41],"practice,":[42],"modern":[43],"LLMs":[44],"frequently":[45],"operate":[46],"in":[47,80],"lowmargin":[48],"regimes":[49],"where":[50],"exhibits":[54],"weak":[55],"preference":[56],"among":[57],"top":[58],"candidates.In":[59],"such":[60],"cases,":[61],"rejecting":[62],"plausible":[63],"runner-up":[64],"tokens":[65],"yields":[66],"negligible":[67],"information":[68],"gain":[69],"while":[70,161],"incurring":[71],"substantial":[72],"rollback":[73],"cost,":[74],"leading":[75],"to":[76,94,146],"a":[77,86],"fundamental":[78],"inefficiency":[79],"verification.We":[81],"propose":[82],"Margin-Aware":[83],"Verification,":[85],"training-free":[87],"domain-agnostic":[89],"strategy":[91],"that":[92,149],"adapts":[93],"model's":[97],"local":[98],"decisiveness.Our":[99],"method":[100,151],"conditions":[101],"decision":[104],"stability":[105],"measured":[106],"directly":[107],"from":[108,144],"logits":[111],"relaxes":[113],"only":[115,125],"when":[116],"provides":[119],"minimal":[120],"benefit.Importantly,":[121],"approach":[123],"modifies":[124],"rule":[128],"is":[130,169],"fully":[131],"compatible":[132],"existing":[134],"targetcoupled":[135],"speculative":[136],"decoding":[137],"frameworks.Extensive":[138],"experiments":[139],"across":[140,165],"scales":[142],"ranging":[143],"8B":[145],"235B":[147],"demonstrate":[148],"our":[150],"delivers":[152],"consistent":[153],"significant":[155],"speedups":[157],"over":[158],"state-of-the-art":[159],"baselines":[160],"preserving":[162],"diverse":[166],"benchmarks.The":[167],"code":[168],"available":[170],"at":[171],"https://github.com/5SSjw/MARS.":[172]},"counts_by_year":[],"updated_date":"2026-10-28T10:33:24.094632","created_date":"2026-07-02T00:00:00"}