{"id":"https://openalex.org/W4415274198","doi":"https://doi.org/10.48550/arxiv.2510.12966","title":"3-Model Speculative Decoding","display_name":"3-Model Speculative Decoding","publication_year":2025,"publication_date":"2025-10-14","ids":{"openalex":"https://openalex.org/W4415274198","doi":"https://doi.org/10.48550/arxiv.2510.12966"},"language":"en","primary_location":{"id":"pmh:oai:arXiv.org:2510.12966","doi":null,"is_oa":true,"landing_page_url":"http://arxiv.org/abs/2510.12966","pdf_url":"https://arxiv.org/pdf/2510.12966","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"},"volume":null,"type":"preprint","is_traditional":true,"indexed_in":["arxiv","datacite"],"open_access":{"is_oa":true,"oa_status":"green","oa_url":"https://arxiv.org/pdf/2510.12966","any_repository_has_fulltext":true},"authorships":[{"author_position":"first","author":{"id":"https://openalex.org/A5030744163","display_name":"Sanghyun Byun","orcid":"https://orcid.org/0009-0002-2149-5770","observed_orcids":["https://orcid.org/0009-0002-2149-5770"]},"institutions":[{"id":"https://openalex.org/I2818286","display_name":"LG (United States)","ror":"https://ror.org/02b948n83","country_code":"US","type":"company","lineage":["https://openalex.org/I2818286","https://openalex.org/I4210131320"]}],"countries":["US"],"is_corresponding":false,"raw_author_name":"Byun, Sanghyun","raw_affiliation_strings":["LG Electronics USA"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"LG Electronics USA","institution_ids":["https://openalex.org/I2818286"]}]},{"author_position":"middle","author":{"id":"https://openalex.org/A5045322212","display_name":"Mohanad Odema","orcid":"https://orcid.org/0000-0002-0828-949X","observed_orcids":["https://orcid.org/0000-0002-0828-949X"]},"institutions":[{"id":"https://openalex.org/I2818286","display_name":"LG (United States)","ror":"https://ror.org/02b948n83","country_code":"US","type":"company","lineage":["https://openalex.org/I2818286","https://openalex.org/I4210131320"]}],"countries":["US"],"is_corresponding":false,"raw_author_name":"Odema, Mohanad","raw_affiliation_strings":["LG Electronics USA"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"LG Electronics USA","institution_ids":["https://openalex.org/I2818286"]}]},{"author_position":"middle","author":{"id":"https://openalex.org/A5120033447","display_name":"Jung Ick Guack","orcid":null,"observed_orcids":[]},"institutions":[{"id":"https://openalex.org/I2818286","display_name":"LG (United States)","ror":"https://ror.org/02b948n83","country_code":"US","type":"company","lineage":["https://openalex.org/I2818286","https://openalex.org/I4210131320"]}],"countries":["US"],"is_corresponding":false,"raw_author_name":"Guack, Jung Ick","raw_affiliation_strings":["LG Electronics USA"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"LG Electronics USA","institution_ids":["https://openalex.org/I2818286"]}]},{"author_position":"middle","author":{"id":"https://openalex.org/A5058186544","display_name":"B.S. Lee","orcid":null,"observed_orcids":[]},"institutions":[{"id":"https://openalex.org/I2818286","display_name":"LG (United States)","ror":"https://ror.org/02b948n83","country_code":"US","type":"company","lineage":["https://openalex.org/I2818286","https://openalex.org/I4210131320"]}],"countries":["US"],"is_corresponding":false,"raw_author_name":"Lee, Baisub","raw_affiliation_strings":["LG Electronics USA"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"LG Electronics USA","institution_ids":["https://openalex.org/I2818286"]}]},{"author_position":"middle","author":{"id":"https://openalex.org/A5110418450","display_name":"Jilai Song","orcid":"https://orcid.org/0009-0009-4069-7897","observed_orcids":["https://orcid.org/0009-0009-4069-7897"]},"institutions":[{"id":"https://openalex.org/I2818286","display_name":"LG (United States)","ror":"https://ror.org/02b948n83","country_code":"US","type":"company","lineage":["https://openalex.org/I2818286","https://openalex.org/I4210131320"]}],"countries":["US"],"is_corresponding":false,"raw_author_name":"Song, Jacob","raw_affiliation_strings":["LG Electronics USA"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"LG Electronics USA","institution_ids":["https://openalex.org/I2818286"]}]},{"author_position":"last","author":{"id":null,"display_name":"Chung, Woo Seong","orcid":null,"observed_orcids":[]},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Chung, Woo Seong","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]}],"institutions":[],"countries_distinct_count":1,"institutions_distinct_count":1,"corresponding_author_ids":[],"corresponding_institution_ids":[],"apc_list":null,"apc_paid":null,"fwci":null,"has_fulltext":true,"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/T10181","display_name":"Natural Language Processing Techniques","score":0.8320000171661377,"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":0.8320000171661377,"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":0.1378999948501587,"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/T10036","display_name":"Advanced Neural Network Applications","score":0.004800000227987766,"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"}}],"keywords":[{"id":"https://openalex.org/keywords/speculative-decoding","display_name":"speculative decoding","score":0.9990000128746033},{"id":"https://openalex.org/keywords/large-language-model-inference","display_name":"large language model inference","score":0.7599999904632568},{"id":"https://openalex.org/keywords/draft-model","display_name":"draft model","score":0.5849999785423279},{"id":"https://openalex.org/keywords/hierarchical-decoding","display_name":"hierarchical decoding","score":0.6759999990463257},{"id":"https://openalex.org/keywords/inference-throughput","display_name":"inference throughput","score":0.8119999766349792},{"id":"https://openalex.org/keywords/distributional-gap","display_name":"distributional gap","score":0.6200000047683716},{"id":"https://openalex.org/keywords/inference-acceleration","display_name":"inference acceleration","score":0.3050000071525574}],"concepts":[{"id":"https://openalex.org/C57273362","wikidata":"https://www.wikidata.org/wiki/Q576722","display_name":"Decoding methods","level":2,"score":0.8446000218391418},{"id":"https://openalex.org/C41008148","wikidata":"https://www.wikidata.org/wiki/Q21198","display_name":"Computer science","level":0,"score":0.7164999842643738},{"id":"https://openalex.org/C48145219","wikidata":"https://www.wikidata.org/wiki/Q1335365","display_name":"Security token","level":2,"score":0.6682999730110168},{"id":"https://openalex.org/C207390915","wikidata":"https://www.wikidata.org/wiki/Q1230525","display_name":"Divergence (linguistics)","level":2,"score":0.491100013256073},{"id":"https://openalex.org/C157764524","wikidata":"https://www.wikidata.org/wiki/Q1383412","display_name":"Throughput","level":3,"score":0.4551999866962433},{"id":"https://openalex.org/C11413529","wikidata":"https://www.wikidata.org/wiki/Q8366","display_name":"Algorithm","level":1,"score":0.44119998812675476},{"id":"https://openalex.org/C2776214188","wikidata":"https://www.wikidata.org/wiki/Q408386","display_name":"Inference","level":2,"score":0.4375999867916107},{"id":"https://openalex.org/C193969084","wikidata":"https://www.wikidata.org/wiki/Q7452500","display_name":"Sequential decoding","level":4,"score":0.4020000100135803},{"id":"https://openalex.org/C2779530757","wikidata":"https://www.wikidata.org/wiki/Q1207505","display_name":"Quality (philosophy)","level":2,"score":0.39320001006126404},{"id":"https://openalex.org/C2778029271","wikidata":"https://www.wikidata.org/wiki/Q5421931","display_name":"Extension (predicate logic)","level":2,"score":0.3504999876022339},{"id":"https://openalex.org/C137293760","wikidata":"https://www.wikidata.org/wiki/Q3621696","display_name":"Language model","level":2,"score":0.32749998569488525},{"id":"https://openalex.org/C2778112365","wikidata":"https://www.wikidata.org/wiki/Q3511065","display_name":"Sequence (biology)","level":2,"score":0.3221000134944916},{"id":"https://openalex.org/C142575187","wikidata":"https://www.wikidata.org/wiki/Q3358290","display_name":"Pyramid (geometry)","level":2,"score":0.3172999918460846},{"id":"https://openalex.org/C154945302","wikidata":"https://www.wikidata.org/wiki/Q11660","display_name":"Artificial intelligence","level":1,"score":0.2924000024795532},{"id":"https://openalex.org/C100776233","wikidata":"https://www.wikidata.org/wiki/Q2532492","display_name":"Bridge (graph theory)","level":2,"score":0.2793999910354614},{"id":"https://openalex.org/C125411270","wikidata":"https://www.wikidata.org/wiki/Q18653","display_name":"Encoding (memory)","level":2,"score":0.2687000036239624},{"id":"https://openalex.org/C94375191","wikidata":"https://www.wikidata.org/wiki/Q11205","display_name":"Arithmetic","level":1,"score":0.2581000030040741},{"id":"https://openalex.org/C80444323","wikidata":"https://www.wikidata.org/wiki/Q2878974","display_name":"Theoretical computer science","level":1,"score":0.2572000026702881},{"id":"https://openalex.org/C179799912","wikidata":"https://www.wikidata.org/wiki/Q205084","display_name":"Computational complexity theory","level":2,"score":0.25360000133514404},{"id":"https://openalex.org/C58166","wikidata":"https://www.wikidata.org/wiki/Q224821","display_name":"Fuzzy logic","level":2,"score":0.2533999979496002}],"mesh":[],"locations_count":2,"locations":[{"id":"pmh:oai:arXiv.org:2510.12966","doi":null,"is_oa":true,"landing_page_url":"http://arxiv.org/abs/2510.12966","pdf_url":"https://arxiv.org/pdf/2510.12966","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"},{"id":"doi:10.48550/arxiv.2510.12966","doi":"https://doi.org/10.48550/arxiv.2510.12966","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2510.12966","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"}],"best_oa_location":{"id":"pmh:oai:arXiv.org:2510.12966","doi":null,"is_oa":true,"landing_page_url":"http://arxiv.org/abs/2510.12966","pdf_url":"https://arxiv.org/pdf/2510.12966","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"},"sustainable_development_goals":[],"study_designs":[],"awards":[],"funders":[],"has_content":{"pdf":true,"grobid_xml":true},"content_urls":{"pdf":"https://content.openalex.org/works/W4415274198.pdf","grobid_xml":"https://content.openalex.org/works/W4415274198.grobid-xml"},"referenced_works_count":6,"referenced_works":["https://openalex.org/W2963995027","https://openalex.org/W4319166707","https://openalex.org/W4385749968","https://openalex.org/W4391590701","https://openalex.org/W4401306886","https://openalex.org/W4412944498"],"related_works":["https://openalex.org/W4391556558","https://openalex.org/W7164012452","https://openalex.org/W4392011687","https://openalex.org/W4415274683","https://openalex.org/W7165466198","https://openalex.org/W7166842263","https://openalex.org/W4411120344","https://openalex.org/W4403346564","https://openalex.org/W7167449285","https://openalex.org/W7127321651"],"abstract_inverted_index":{"Speculative":[0,72],"Decoding":[1,73],"(SD)":[2],"accelerates":[3],"inference":[4,210],"in":[5,62,95],"large":[6],"language":[7],"models":[8,48,126],"by":[9,22,36],"using":[10],"a":[11,23,37,165,174,195],"smaller":[12,46,99,124],"draft":[13,40,47,87,125,176],"model":[14,41,84,100,177,187],"to":[15,90,101,137,152,198,208],"propose":[16],"tokens,":[17],"which":[18],"are":[19,33],"then":[20],"verified":[21],"larger":[24],"target":[25,59,89,181,186],"model.":[26],"However,":[27],"the":[28,58,86,92,120],"throughput":[29],"gains":[30],"of":[31,77,122],"SD":[32,78],"fundamentally":[34],"limited":[35],"trade-off":[38],"between":[39,85],"size":[42],"and":[43,66,88,118,178,203],"token":[44],"acceptance:":[45],"generate":[49],"tokens":[50,161],"more":[51],"quickly":[52],"but":[53],"exhibit":[54],"greater":[55],"divergence":[56,140],"from":[57],"model,":[60,182],"resulting":[61],"lower":[63],"acceptance":[64,116,135],"rates":[65,117],"reduced":[67],"speedups.":[68],"We":[69],"introduce":[70],"Pyramid":[71],"(PyramidSD),":[74],"an":[75,81,179],"extension":[76],"that":[79],"inserts":[80],"intermediate":[82],"qualifier":[83],"bridge":[91],"distributional":[93],"gap":[94],"output":[96],"predictions,":[97],"allowing":[98,119],"be":[102,205],"used":[103],"for":[104,189],"drafting.":[105],"This":[106],"hierarchical":[107],"decoding":[108,201],"strategy":[109],"improves":[110],"alignment":[111],"across":[112],"models,":[113],"enabling":[114],"higher":[115],"use":[121],"significantly":[123],"without":[127],"sacrificing":[128],"overall":[129],"performance.":[130],"PyramidSD":[131,149,183,193],"builds":[132],"on":[133,164],"fuzzy":[134],"criteria":[136],"support":[138],"relaxed":[139],"thresholds":[141],"at":[142],"each":[143],"stage,":[144],"improving":[145],"throughput.":[146,191],"In":[147,170],"experiments,":[148],"achieves":[150],"up":[151],"1.91x":[153],"generation":[154],"speed":[155],"over":[156],"standard":[157],"SD,":[158],"reaching":[159],"124":[160],"per":[162],"second":[163],"consumer":[166],"GPU":[167],"(RTX":[168],"4090).":[169],"small-memory":[171],"settings":[172],"with":[173],"1B-parameter":[175],"8B":[180],"minimally":[184],"trades":[185],"quality":[188],"improved":[190],"Overall,":[192],"offers":[194],"practical":[196],"approach":[197],"enhancing":[199],"speculative":[200],"efficiency":[202],"can":[204],"readily":[206],"applied":[207],"existing":[209],"pipelines.":[211]},"counts_by_year":[],"updated_date":"2026-11-10T11:39:43.021295","created_date":"2025-10-17T00:00:00"}