{"id":"https://openalex.org/W7213232249","doi":"https://doi.org/10.48550/arxiv.2609.15313","title":"Reducing the Output-Mode Gap in Speech Language Models via Joint-Output On-Policy Distillation","display_name":"Reducing the Output-Mode Gap in Speech Language Models via Joint-Output On-Policy Distillation","publication_year":2026,"publication_date":"2026-09-14","ids":{"openalex":"https://openalex.org/W7213232249","doi":"https://doi.org/10.48550/arxiv.2609.15313"},"language":"en","primary_location":{"id":"pmh:oai:arXiv.org:2609.15313","doi":null,"is_oa":true,"landing_page_url":"https://arxiv.org/abs/2609.15313","pdf_url":"https://arxiv.org/pdf/2609.15313","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/2609.15313","any_repository_has_fulltext":true},"authorships":[{"author_position":"first","author":{"id":"https://openalex.org/A5006527442","display_name":"Daxin Tan","orcid":null,"observed_orcids":[]},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Tan, Daxin","raw_affiliation_strings":["AI Lab , Leibniz Research Center , Huawei"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"AI Lab , Leibniz Research Center , Huawei","institution_ids":[]}]},{"author_position":"middle","author":{"id":"https://openalex.org/A5128749554","display_name":"Dehua Tao","orcid":null,"observed_orcids":[]},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Tao, Dehua","raw_affiliation_strings":["AI Lab , Leibniz Research Center , Huawei"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"AI Lab , Leibniz Research Center , Huawei","institution_ids":[]}]},{"author_position":"middle","author":{"id":"https://openalex.org/A5126738311","display_name":"Chengxi Deng","orcid":null,"observed_orcids":[]},"institutions":[{"id":"https://openalex.org/I177725633","display_name":"Chinese University of Hong Kong","ror":"https://ror.org/00t33hh48","country_code":"HK","type":"education","lineage":["https://openalex.org/I177725633"]}],"countries":["HK"],"is_corresponding":false,"raw_author_name":"Deng, Chengxi","raw_affiliation_strings":["The Chinese University of Hong Kong"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"The Chinese University of Hong Kong","institution_ids":["https://openalex.org/I177725633"]}]},{"author_position":"middle","author":{"id":"https://openalex.org/A5123089402","display_name":"Hanlin Zhang","orcid":null,"observed_orcids":[]},"institutions":[{"id":"https://openalex.org/I168719708","display_name":"City University of Hong Kong","ror":"https://ror.org/03q8dnn23","country_code":"HK","type":"education","lineage":["https://openalex.org/I168719708"]}],"countries":["HK"],"is_corresponding":false,"raw_author_name":"Zhang, Hanlin","raw_affiliation_strings":["City University of Hong Kong"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"City University of Hong Kong","institution_ids":["https://openalex.org/I168719708"]}]},{"author_position":"last","author":{"id":"https://openalex.org/A5123876131","display_name":"Xiao Chen","orcid":null,"observed_orcids":[]},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Chen, Xiao","raw_affiliation_strings":["AI Lab , Leibniz Research Center , Huawei"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"AI Lab , Leibniz Research Center , Huawei","institution_ids":[]}]}],"institutions":[],"countries_distinct_count":1,"institutions_distinct_count":2,"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/T10201","display_name":"Speech Recognition and Synthesis","score":0.926800012588501,"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/T10201","display_name":"Speech Recognition and Synthesis","score":0.926800012588501,"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/T12031","display_name":"Speech and dialogue systems","score":0.05900000035762787,"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":0.007400000002235174,"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/speech-large-language-models","display_name":"speech large language models","score":0.8579999804496765},{"id":"https://openalex.org/keywords/on-policy-distillation","display_name":"on-policy distillation","score":0.9200000166893005},{"id":"https://openalex.org/keywords/soft-targets","display_name":"soft targets","score":0.5070000290870667},{"id":"https://openalex.org/keywords/gsm8k","display_name":"GSM8K","score":0.8069999814033508}],"concepts":[{"id":"https://openalex.org/C41008148","wikidata":"https://www.wikidata.org/wiki/Q21198","display_name":"Computer science","level":0,"score":0.7756999731063843},{"id":"https://openalex.org/C2779343474","wikidata":"https://www.wikidata.org/wiki/Q3109175","display_name":"Context (archaeology)","level":2,"score":0.614799976348877},{"id":"https://openalex.org/C28490314","wikidata":"https://www.wikidata.org/wiki/Q189436","display_name":"Speech recognition","level":1,"score":0.6079999804496765},{"id":"https://openalex.org/C137293760","wikidata":"https://www.wikidata.org/wiki/Q3621696","display_name":"Language model","level":2,"score":0.5837000012397766},{"id":"https://openalex.org/C57493831","wikidata":"https://www.wikidata.org/wiki/Q3134666","display_name":"Projection (relational algebra)","level":2,"score":0.492900013923645},{"id":"https://openalex.org/C159877910","wikidata":"https://www.wikidata.org/wiki/Q2202883","display_name":"Autoregressive model","level":2,"score":0.4271000027656555},{"id":"https://openalex.org/C154945302","wikidata":"https://www.wikidata.org/wiki/Q11660","display_name":"Artificial intelligence","level":1,"score":0.41200000047683716},{"id":"https://openalex.org/C204321447","wikidata":"https://www.wikidata.org/wiki/Q30642","display_name":"Natural language processing","level":1,"score":0.40059998631477356},{"id":"https://openalex.org/C2993167945","wikidata":"https://www.wikidata.org/wiki/Q196626","display_name":"Linguistic context","level":3,"score":0.3935999870300293},{"id":"https://openalex.org/C61797465","wikidata":"https://www.wikidata.org/wiki/Q1188986","display_name":"Term (time)","level":2,"score":0.38420000672340393},{"id":"https://openalex.org/C2776187449","wikidata":"https://www.wikidata.org/wiki/Q1513879","display_name":"Natural language generation","level":3,"score":0.3779999911785126},{"id":"https://openalex.org/C48677424","wikidata":"https://www.wikidata.org/wiki/Q6888088","display_name":"Mode (computer interface)","level":2,"score":0.3718999922275543},{"id":"https://openalex.org/C2985684807","wikidata":"https://www.wikidata.org/wiki/Q1513879","display_name":"Text generation","level":2,"score":0.36550000309944153},{"id":"https://openalex.org/C155635449","wikidata":"https://www.wikidata.org/wiki/Q4674699","display_name":"Acoustic model","level":3,"score":0.35010001063346863},{"id":"https://openalex.org/C61328038","wikidata":"https://www.wikidata.org/wiki/Q3358061","display_name":"Speech processing","level":2,"score":0.3458999991416931},{"id":"https://openalex.org/C149364088","wikidata":"https://www.wikidata.org/wiki/Q185917","display_name":"Translation (biology)","level":4,"score":0.319599986076355},{"id":"https://openalex.org/C204030448","wikidata":"https://www.wikidata.org/wiki/Q101017","display_name":"Distillation","level":2,"score":0.31520000100135803},{"id":"https://openalex.org/C18555067","wikidata":"https://www.wikidata.org/wiki/Q8375051","display_name":"Joint (building)","level":2,"score":0.3125},{"id":"https://openalex.org/C183322885","wikidata":"https://www.wikidata.org/wiki/Q17007702","display_name":"Context model","level":3,"score":0.3018999993801117},{"id":"https://openalex.org/C81917197","wikidata":"https://www.wikidata.org/wiki/Q628760","display_name":"Selection (genetic algorithm)","level":2,"score":0.2922999858856201},{"id":"https://openalex.org/C2776182073","wikidata":"https://www.wikidata.org/wiki/Q7575395","display_name":"Speech enhancement","level":3,"score":0.28870001435279846},{"id":"https://openalex.org/C195324797","wikidata":"https://www.wikidata.org/wiki/Q33742","display_name":"Natural language","level":2,"score":0.2856999933719635},{"id":"https://openalex.org/C166955791","wikidata":"https://www.wikidata.org/wiki/Q629579","display_name":"Macro","level":2,"score":0.2831999957561493},{"id":"https://openalex.org/C204201278","wikidata":"https://www.wikidata.org/wiki/Q1332614","display_name":"Voice activity detection","level":3,"score":0.2669000029563904}],"mesh":[],"locations_count":2,"locations":[{"id":"pmh:oai:arXiv.org:2609.15313","doi":null,"is_oa":true,"landing_page_url":"https://arxiv.org/abs/2609.15313","pdf_url":"https://arxiv.org/pdf/2609.15313","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.2609.15313","doi":"https://doi.org/10.48550/arxiv.2609.15313","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2609.15313","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":null,"license_id":null,"version":null,"is_accepted":false,"is_published":false,"raw_source_name":null,"raw_type":"Preprint"}],"best_oa_location":{"id":"pmh:oai:arXiv.org:2609.15313","doi":null,"is_oa":true,"landing_page_url":"https://arxiv.org/abs/2609.15313","pdf_url":"https://arxiv.org/pdf/2609.15313","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":[{"display_name":"Peace, justice, and strong institutions","score":0.4286484122276306,"id":"https://metadata.un.org/sdg/16"}],"study_designs":[],"awards":[],"funders":[],"has_content":{"pdf":true,"grobid_xml":true},"content_urls":{"pdf":"https://content.openalex.org/works/W7213232249.pdf","grobid_xml":"https://content.openalex.org/works/W7213232249.grobid-xml"},"referenced_works_count":0,"referenced_works":[],"related_works":["https://openalex.org/W7127173562","https://openalex.org/W7214386216","https://openalex.org/W7139136507","https://openalex.org/W7213234831","https://openalex.org/W7202135553","https://openalex.org/W7141544855","https://openalex.org/W7221834589","https://openalex.org/W7140143788","https://openalex.org/W4389524018","https://openalex.org/W4417248906"],"abstract_inverted_index":{"Autoregressive":[0],"generation":[1,14,25,92,144],"of":[2,35,111],"interleaved":[3,124,143],"text":[4,40,55,99],"and":[5,137,159,174],"acoustic":[6,31],"tokens":[7,32],"is":[8],"a":[9,108,186],"common":[10],"approach":[11],"to":[12,153,162],"spoken-response":[13],"in":[15,57,171,189],"speech":[16,44],"large":[17],"language":[18],"models.":[19],"Although":[20],"this":[21,68],"design":[22],"enables":[23],"streaming":[24],"with":[26,168],"explicit":[27],"textual":[28],"guidance,":[29],"generated":[30,56],"become":[33],"part":[34],"the":[36,53,70,85,101,112,117,121],"context":[37],"for":[38,52,62],"subsequent":[39],"predictions.":[41,133],"Given":[42],"identical":[43],"inputs,":[45],"we":[46,77],"observe":[47],"markedly":[48],"lower":[49],"answer":[50],"accuracy":[51,173,191],"internal":[54],"speech-to-text-and-speech":[58],"(S2TS)":[59],"mode":[60],"than":[61,178],"speech-to-text":[63],"(S2T)":[64],"responses.":[65],"We":[66],"term":[67],"discrepancy":[69],"\\emph{output-mode":[71],"gap}":[72],"(OMG).":[73],"To":[74],"reduce":[75],"OMG,":[76],"propose":[78],"\\emph{Joint-Output":[79],"On-Policy":[80],"Distillation}":[81],"(JO-OPD),":[82],"which":[83],"distills":[84],"model's":[86],"stronger":[87],"S2T":[88,102,172],"policy":[89],"into":[90],"joint":[91],"using":[93],"student-generated":[94],"S2TS":[95],"trajectories.":[96],"At":[97],"each":[98],"position,":[100],"teacher":[103],"provides":[104],"soft":[105],"targets":[106],"from":[107,120,151,160],"text-only":[109],"projection":[110],"student's":[113],"preceding":[114],"outputs,":[115],"while":[116],"student":[118],"predicts":[119],"corresponding":[122],"full":[123],"history.":[125],"A":[126],"preservation":[127],"objective":[128],"further":[129,184],"regularizes":[130],"native":[131],"non-text":[132],"Experiments":[134],"on":[135,157,165,192],"Step-Audio-2-mini":[136],"Baichuan-Audio-Instruct":[138],"reveal":[139],"OMG":[140,150],"across":[141],"two":[142],"architectures.":[145],"On":[146],"Step-Audio-2-mini,":[147],"JO-OPD":[148],"reduces":[149],"42.87":[152],"16.26":[154],"percentage":[155],"points":[156,164],"Spoken-MQA":[158],"29.72":[161],"13.04":[163],"speech-rendered":[166],"GSM8K,":[167],"little":[169],"change":[170],"substantially":[175],"larger":[176],"reductions":[177],"matched":[179],"SFT":[180],"baselines.":[181],"ASR-based":[182],"evaluation":[183],"shows":[185],"7.49-point":[187],"improvement":[188],"spoken-answer":[190],"Spoken-MQA.":[193]},"counts_by_year":[],"updated_date":"2026-10-28T10:33:24.094632","created_date":"2026-09-16T00:00:00"}