{"id":"https://openalex.org/W7133306977","doi":"https://doi.org/10.48550/arxiv.2603.01639","title":"Learning to Draft: Adaptive Speculative Decoding with Reinforcement Learning","display_name":"Learning to Draft: Adaptive Speculative Decoding with Reinforcement Learning","publication_year":2026,"publication_date":"2026-03-02","ids":{"openalex":"https://openalex.org/W7133306977","doi":"https://doi.org/10.48550/arxiv.2603.01639"},"language":null,"primary_location":{"id":"doi:10.48550/arxiv.2603.01639","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2603.01639","pdf_url":null,"source":{"id":"https://openalex.org/S4306400194","display_name":"arXiv (Cornell University)","issn_l":null,"issn":null,"is_oa":true,"is_in_doaj":false,"is_core":false,"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":"article"},"type":"preprint","indexed_in":["datacite"],"open_access":{"is_oa":true,"oa_status":"green","oa_url":"https://doi.org/10.48550/arxiv.2603.01639","any_repository_has_fulltext":true},"authorships":[{"author_position":"first","author":{"id":"https://openalex.org/A5127907326","display_name":"Jiebin Zhang","orcid":null},"institutions":[],"countries":[],"is_corresponding":true,"raw_author_name":"Zhang, Jiebin","raw_affiliation_strings":[],"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5127910150","display_name":"Zhenghan Yu","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Yu, Zhenghan","raw_affiliation_strings":[],"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5127894660","display_name":"Liang Wang","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Wang, Liang","raw_affiliation_strings":[],"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5127909203","display_name":"Nan Yang","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Yang, Nan","raw_affiliation_strings":[],"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5113131182","display_name":"Eugene J. Yu","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Yu, Eugene J.","raw_affiliation_strings":[],"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5127898716","display_name":"Zheng Li (26302)","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Li, Zheng","raw_affiliation_strings":[],"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5127886760","display_name":"Yifan Song","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Song, Yifan","raw_affiliation_strings":[],"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5127916119","display_name":"Dawei Zhu","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Zhu, Dawei","raw_affiliation_strings":[],"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5127991853","display_name":"Xingxing Zhang","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Zhang, Xingxing","raw_affiliation_strings":[],"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5127983391","display_name":"Furu Wei","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Wei, Furu","raw_affiliation_strings":[],"affiliations":[]},{"author_position":"last","author":{"id":"https://openalex.org/A5106951666","display_name":"Sujian Li","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Li, Sujian","raw_affiliation_strings":[],"affiliations":[]}],"institutions":[],"countries_distinct_count":0,"institutions_distinct_count":11,"corresponding_author_ids":["https://openalex.org/A5127907326"],"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":0.28209999203681946,"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":0.28209999203681946,"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":0.13179999589920044,"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/T13910","display_name":"Computational and Text Analysis Methods","score":0.08820000290870667,"subfield":{"id":"https://openalex.org/subfields/3300","display_name":"General Social Sciences"},"field":{"id":"https://openalex.org/fields/33","display_name":"Social Sciences"},"domain":{"id":"https://openalex.org/domains/2","display_name":"Social Sciences"}}],"keywords":[{"id":"https://openalex.org/keywords/reinforcement-learning","display_name":"Reinforcement learning","score":0.7251999974250793},{"id":"https://openalex.org/keywords/decoding-methods","display_name":"Decoding methods","score":0.633400022983551},{"id":"https://openalex.org/keywords/speedup","display_name":"Speedup","score":0.6072999835014343},{"id":"https://openalex.org/keywords/inference","display_name":"Inference","score":0.5554999709129333},{"id":"https://openalex.org/keywords/security-token","display_name":"Security token","score":0.505299985408783},{"id":"https://openalex.org/keywords/throughput","display_name":"Throughput","score":0.43389999866485596},{"id":"https://openalex.org/keywords/proxy","display_name":"Proxy (statistics)","score":0.4034999907016754}],"concepts":[{"id":"https://openalex.org/C41008148","wikidata":"https://www.wikidata.org/wiki/Q21198","display_name":"Computer science","level":0,"score":0.8159999847412109},{"id":"https://openalex.org/C97541855","wikidata":"https://www.wikidata.org/wiki/Q830687","display_name":"Reinforcement learning","level":2,"score":0.7251999974250793},{"id":"https://openalex.org/C57273362","wikidata":"https://www.wikidata.org/wiki/Q576722","display_name":"Decoding methods","level":2,"score":0.633400022983551},{"id":"https://openalex.org/C68339613","wikidata":"https://www.wikidata.org/wiki/Q1549489","display_name":"Speedup","level":2,"score":0.6072999835014343},{"id":"https://openalex.org/C2776214188","wikidata":"https://www.wikidata.org/wiki/Q408386","display_name":"Inference","level":2,"score":0.5554999709129333},{"id":"https://openalex.org/C154945302","wikidata":"https://www.wikidata.org/wiki/Q11660","display_name":"Artificial intelligence","level":1,"score":0.513700008392334},{"id":"https://openalex.org/C48145219","wikidata":"https://www.wikidata.org/wiki/Q1335365","display_name":"Security token","level":2,"score":0.505299985408783},{"id":"https://openalex.org/C119857082","wikidata":"https://www.wikidata.org/wiki/Q2539","display_name":"Machine learning","level":1,"score":0.4659999907016754},{"id":"https://openalex.org/C157764524","wikidata":"https://www.wikidata.org/wiki/Q1383412","display_name":"Throughput","level":3,"score":0.43389999866485596},{"id":"https://openalex.org/C2780148112","wikidata":"https://www.wikidata.org/wiki/Q1432581","display_name":"Proxy (statistics)","level":2,"score":0.4034999907016754},{"id":"https://openalex.org/C115051666","wikidata":"https://www.wikidata.org/wiki/Q6522493","display_name":"Ranging","level":2,"score":0.322299987077713},{"id":"https://openalex.org/C22367795","wikidata":"https://www.wikidata.org/wiki/Q7625208","display_name":"Structured prediction","level":2,"score":0.30329999327659607},{"id":"https://openalex.org/C2780451532","wikidata":"https://www.wikidata.org/wiki/Q759676","display_name":"Task (project management)","level":2,"score":0.2897999882698059},{"id":"https://openalex.org/C175154964","wikidata":"https://www.wikidata.org/wiki/Q380077","display_name":"Task analysis","level":3,"score":0.28700000047683716},{"id":"https://openalex.org/C137293760","wikidata":"https://www.wikidata.org/wiki/Q3621696","display_name":"Language model","level":2,"score":0.2856999933719635},{"id":"https://openalex.org/C177264268","wikidata":"https://www.wikidata.org/wiki/Q1514741","display_name":"Set (abstract data type)","level":2,"score":0.28189998865127563},{"id":"https://openalex.org/C125014702","wikidata":"https://www.wikidata.org/wiki/Q4680749","display_name":"Adaptive learning","level":2,"score":0.2551000118255615},{"id":"https://openalex.org/C81917197","wikidata":"https://www.wikidata.org/wiki/Q628760","display_name":"Selection (genetic algorithm)","level":2,"score":0.25459998846054077}],"mesh":[],"locations_count":1,"locations":[{"id":"doi:10.48550/arxiv.2603.01639","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2603.01639","pdf_url":null,"source":{"id":"https://openalex.org/S4306400194","display_name":"arXiv (Cornell University)","issn_l":null,"issn":null,"is_oa":true,"is_in_doaj":false,"is_core":false,"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":null,"raw_source_name":null,"raw_type":"article"}],"best_oa_location":{"id":"doi:10.48550/arxiv.2603.01639","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2603.01639","pdf_url":null,"source":{"id":"https://openalex.org/S4306400194","display_name":"arXiv (Cornell University)","issn_l":null,"issn":null,"is_oa":true,"is_in_doaj":false,"is_core":false,"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":"article"},"sustainable_development_goals":[],"awards":[],"funders":[],"has_content":{"grobid_xml":false,"pdf":false},"content_urls":null,"referenced_works_count":0,"referenced_works":[],"related_works":[],"abstract_inverted_index":{"Speculative":[0],"decoding":[1,136],"accelerates":[2],"large":[3],"language":[4],"model":[5,13,22],"(LLM)":[6],"inference":[7],"by":[8],"using":[9],"a":[10,19,50,90,107],"small":[11],"draft":[12,120],"to":[14,23,87,116,128,130,161,169],"generate":[15],"candidate":[16],"tokens":[17],"for":[18,59,96],"larger":[20],"target":[21],"verify.":[24],"The":[25],"efficacy":[26],"of":[27,98],"this":[28],"technique":[29],"hinges":[30],"on":[31,38,49,142],"the":[32,35,67,73,104,119,126,164],"trade-off":[33],"between":[34],"time":[36,52,69],"spent":[37],"drafting":[39,74],"candidates":[40],"and":[41,71,75,111,121,133,146],"verifying":[42],"them.":[43],"However,":[44],"current":[45],"state-of-the-art":[46,165],"methods":[47],"rely":[48],"static":[51],"allocation,":[53],"while":[54],"recent":[55],"dynamic":[56],"approaches":[57],"optimize":[58],"proxy":[60],"metrics":[61],"like":[62],"acceptance":[63],"length,":[64],"often":[65],"neglecting":[66],"true":[68],"cost":[70],"treating":[72],"verification":[76,122],"phases":[77],"in":[78],"isolation.":[79],"To":[80],"address":[81],"these":[82],"limitations,":[83],"we":[84],"introduce":[85],"Learning":[86],"Draft":[88],"(LTD),":[89],"novel":[91],"method":[92,166],"that":[93,153],"directly":[94],"optimizes":[95],"throughput":[97],"each":[99,131],"draft-and-verify":[100],"cycle.":[101],"We":[102,138],"formulate":[103],"problem":[105],"as":[106],"reinforcement":[108],"learning":[109],"environment":[110],"train":[112],"two":[113],"co-adaptive":[114],"policies":[115,127],"dynamically":[117],"coordinate":[118],"phases.":[123],"This":[124],"encourages":[125],"adapt":[129],"other":[132],"explicitly":[134],"maximize":[135],"efficiency.":[137],"conducted":[139],"extensive":[140],"evaluations":[141],"five":[143],"diverse":[144],"LLMs":[145],"four":[147],"distinct":[148],"tasks.":[149],"Our":[150],"results":[151],"show":[152],"LTD":[154],"achieves":[155],"speedup":[156],"ratios":[157],"ranging":[158],"from":[159],"2.24x":[160],"4.32x,":[162],"outperforming":[163],"Eagle3":[167],"up":[168],"36.4%.":[170]},"counts_by_year":[],"updated_date":"2026-03-04T07:09:34.246503","created_date":"2026-03-04T00:00:00"}