{"id":"https://openalex.org/W7215110389","doi":"https://doi.org/10.48550/arxiv.2609.38666","title":"Understanding Off- vs On-Policy Distillation: A Tale of Distinct Training Objectives","display_name":"Understanding Off- vs On-Policy Distillation: A Tale of Distinct Training Objectives","publication_year":2026,"publication_date":"2026-09-29","ids":{"openalex":"https://openalex.org/W7215110389","doi":"https://doi.org/10.48550/arxiv.2609.38666"},"language":"en","primary_location":{"id":"pmh:oai:arXiv.org:2609.38666","doi":null,"is_oa":true,"landing_page_url":"https://arxiv.org/abs/2609.38666","pdf_url":"https://arxiv.org/pdf/2609.38666","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","is_traditional":true,"indexed_in":["arxiv","datacite"],"open_access":{"is_oa":true,"oa_status":"green","oa_url":"https://arxiv.org/pdf/2609.38666","any_repository_has_fulltext":true},"authorships":[{"author_position":"first","author":{"id":"https://openalex.org/A5113779780","display_name":"Qiwei Di","orcid":null,"observed_orcids":[]},"institutions":[{"id":"https://openalex.org/I161318765","display_name":"University of California, Los Angeles","ror":"https://ror.org/046rm7j60","country_code":"US","type":"education","lineage":["https://openalex.org/I161318765"]}],"countries":["US"],"is_corresponding":false,"raw_author_name":"Di, Qiwei","raw_affiliation_strings":["Department of Computer Science , University of California , Los Angeles , CA 90095 , USA;"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"Department of Computer Science , University of California , Los Angeles , CA 90095 , USA;","institution_ids":["https://openalex.org/I161318765"]}]},{"author_position":"middle","author":{"id":"https://openalex.org/A5019561171","display_name":"Xuheng Li","orcid":"https://orcid.org/0009-0003-2006-8619","observed_orcids":["https://orcid.org/0009-0003-2006-8619"]},"institutions":[{"id":"https://openalex.org/I161318765","display_name":"University of California, Los Angeles","ror":"https://ror.org/046rm7j60","country_code":"US","type":"education","lineage":["https://openalex.org/I161318765"]}],"countries":["US"],"is_corresponding":false,"raw_author_name":"Li, Xuheng","raw_affiliation_strings":["Department of Computer Science , University of California , Los Angeles , CA 90095 , USA;"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"Department of Computer Science , University of California , Los Angeles , CA 90095 , USA;","institution_ids":["https://openalex.org/I161318765"]}]},{"author_position":"middle","author":{"id":"https://openalex.org/A5043890914","display_name":"Kaixuan Ji","orcid":"https://orcid.org/0000-0002-6641-0155","observed_orcids":["https://orcid.org/0000-0002-6641-0155"]},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Ji, Kaixuan","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5153701683","display_name":"Chenggong Zhang","orcid":null,"observed_orcids":[]},"institutions":[{"id":"https://openalex.org/I161318765","display_name":"University of California, Los Angeles","ror":"https://ror.org/046rm7j60","country_code":"US","type":"education","lineage":["https://openalex.org/I161318765"]}],"countries":["US"],"is_corresponding":false,"raw_author_name":"Zhang, Chenggong","raw_affiliation_strings":["Department of Computer Science , University of California , Los Angeles , CA 90095 , USA;"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"Department of Computer Science , University of California , Los Angeles , CA 90095 , USA;","institution_ids":["https://openalex.org/I161318765"]}]},{"author_position":"middle","author":{"id":"https://openalex.org/A5063820180","display_name":"Heyang Zhao","orcid":null,"observed_orcids":[]},"institutions":[{"id":"https://openalex.org/I161318765","display_name":"University of California, Los Angeles","ror":"https://ror.org/046rm7j60","country_code":"US","type":"education","lineage":["https://openalex.org/I161318765"]}],"countries":["US"],"is_corresponding":false,"raw_author_name":"Zhao, Heyang","raw_affiliation_strings":["Department of Computer Science , University of California , Los Angeles , CA 90095 , USA;"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"Department of Computer Science , University of California , Los Angeles , CA 90095 , USA;","institution_ids":["https://openalex.org/I161318765"]}]},{"author_position":"last","author":{"id":"https://openalex.org/A5120284571","display_name":"Quanquan Gu","orcid":null,"observed_orcids":[]},"institutions":[{"id":"https://openalex.org/I161318765","display_name":"University of California, Los Angeles","ror":"https://ror.org/046rm7j60","country_code":"US","type":"education","lineage":["https://openalex.org/I161318765"]}],"countries":["US"],"is_corresponding":false,"raw_author_name":"Gu, Quanquan","raw_affiliation_strings":["Department of Computer Science , University of California , Los Angeles , CA 90095 , USA;"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"Department of Computer Science , University of California , Los Angeles , CA 90095 , USA;","institution_ids":["https://openalex.org/I161318765"]}]}],"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/T10462","display_name":"Reinforcement Learning in Robotics","score":0.8999000191688538,"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/T10462","display_name":"Reinforcement Learning in Robotics","score":0.8999000191688538,"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/T11307","display_name":"Domain Adaptation and Few-Shot Learning","score":0.01860000006854534,"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/T12101","display_name":"Advanced Bandit Algorithms Research","score":0.015799999237060547,"subfield":{"id":"https://openalex.org/subfields/1803","display_name":"Management Science and Operations Research"},"field":{"id":"https://openalex.org/fields/18","display_name":"Decision Sciences"},"domain":{"id":"https://openalex.org/domains/2","display_name":"Social Sciences"}}],"keywords":[{"id":"https://openalex.org/keywords/on-policy-distillation","display_name":"on-policy distillation","score":0.9139999747276306},{"id":"https://openalex.org/keywords/multiple-teachers","display_name":"multiple teachers","score":0.7139999866485596},{"id":"https://openalex.org/keywords/reverse-kl-divergence","display_name":"reverse KL divergence","score":0.9959999918937683},{"id":"https://openalex.org/keywords/regret-bounds","display_name":"regret bounds","score":0.722000002861023},{"id":"https://openalex.org/keywords/forgetting","display_name":"forgetting","score":0.34700000286102295},{"id":"https://openalex.org/keywords/tabular-reinforcement-learning","display_name":"tabular reinforcement learning","score":0.5040000081062317},{"id":"https://openalex.org/keywords/function-approximation","display_name":"function approximation","score":0.8059999942779541}],"concepts":[{"id":"https://openalex.org/C50817715","wikidata":"https://www.wikidata.org/wiki/Q79895177","display_name":"Regret","level":2,"score":0.7348999977111816},{"id":"https://openalex.org/C7149132","wikidata":"https://www.wikidata.org/wiki/Q1377840","display_name":"Forgetting","level":2,"score":0.7031000256538391},{"id":"https://openalex.org/C207390915","wikidata":"https://www.wikidata.org/wiki/Q1230525","display_name":"Divergence (linguistics)","level":2,"score":0.6579999923706055},{"id":"https://openalex.org/C171752962","wikidata":"https://www.wikidata.org/wiki/Q255166","display_name":"Kullback\u2013Leibler divergence","level":2,"score":0.5260999798774719},{"id":"https://openalex.org/C774472","wikidata":"https://www.wikidata.org/wiki/Q6760393","display_name":"Margin (machine learning)","level":2,"score":0.4925999939441681},{"id":"https://openalex.org/C41008148","wikidata":"https://www.wikidata.org/wiki/Q21198","display_name":"Computer science","level":0,"score":0.48910000920295715},{"id":"https://openalex.org/C88626702","wikidata":"https://www.wikidata.org/wiki/Q1128903","display_name":"Continuation","level":2,"score":0.4860000014305115},{"id":"https://openalex.org/C39927690","wikidata":"https://www.wikidata.org/wiki/Q11197","display_name":"Logarithm","level":2,"score":0.4778999984264374},{"id":"https://openalex.org/C33923547","wikidata":"https://www.wikidata.org/wiki/Q395","display_name":"Mathematics","level":0,"score":0.4472000002861023},{"id":"https://openalex.org/C14036430","wikidata":"https://www.wikidata.org/wiki/Q3736076","display_name":"Function (biology)","level":2,"score":0.4052000045776367},{"id":"https://openalex.org/C119857082","wikidata":"https://www.wikidata.org/wiki/Q2539","display_name":"Machine learning","level":1,"score":0.3926999866962433},{"id":"https://openalex.org/C154945302","wikidata":"https://www.wikidata.org/wiki/Q11660","display_name":"Artificial intelligence","level":1,"score":0.38580000400543213},{"id":"https://openalex.org/C2777211547","wikidata":"https://www.wikidata.org/wiki/Q17141490","display_name":"Training (meteorology)","level":2,"score":0.34200000762939453},{"id":"https://openalex.org/C81917197","wikidata":"https://www.wikidata.org/wiki/Q628760","display_name":"Selection (genetic algorithm)","level":2,"score":0.32899999618530273},{"id":"https://openalex.org/C204030448","wikidata":"https://www.wikidata.org/wiki/Q101017","display_name":"Distillation","level":2,"score":0.3005000054836273},{"id":"https://openalex.org/C80191262","wikidata":"https://www.wikidata.org/wiki/Q5477668","display_name":"Fragility","level":2,"score":0.2856000065803528},{"id":"https://openalex.org/C2781067378","wikidata":"https://www.wikidata.org/wiki/Q17027399","display_name":"Interpretability","level":2,"score":0.28369998931884766},{"id":"https://openalex.org/C141603448","wikidata":"https://www.wikidata.org/wiki/Q134830","display_name":"Prefix","level":2,"score":0.27459999918937683},{"id":"https://openalex.org/C201236551","wikidata":"https://www.wikidata.org/wiki/Q7310809","display_name":"Relative value","level":2,"score":0.27070000767707825},{"id":"https://openalex.org/C2781249084","wikidata":"https://www.wikidata.org/wiki/Q908656","display_name":"Preference","level":2,"score":0.25920000672340393},{"id":"https://openalex.org/C189950617","wikidata":"https://www.wikidata.org/wiki/Q937228","display_name":"Property (philosophy)","level":2,"score":0.25769999623298645},{"id":"https://openalex.org/C2780801425","wikidata":"https://www.wikidata.org/wiki/Q5164392","display_name":"Construct (python library)","level":2,"score":0.2533000111579895},{"id":"https://openalex.org/C165064840","wikidata":"https://www.wikidata.org/wiki/Q1321061","display_name":"Matching (statistics)","level":2,"score":0.2513999938964844}],"mesh":[],"locations_count":2,"locations":[{"id":"pmh:oai:arXiv.org:2609.38666","doi":null,"is_oa":true,"landing_page_url":"https://arxiv.org/abs/2609.38666","pdf_url":"https://arxiv.org/pdf/2609.38666","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.2609.38666","doi":"https://doi.org/10.48550/arxiv.2609.38666","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2609.38666","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:2609.38666","doi":null,"is_oa":true,"landing_page_url":"https://arxiv.org/abs/2609.38666","pdf_url":"https://arxiv.org/pdf/2609.38666","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":[{"id":"https://metadata.un.org/sdg/4","score":0.6164000034332275,"display_name":"Quality education"}],"study_designs":[],"awards":[],"funders":[],"has_content":{"pdf":true,"grobid_xml":true},"content_urls":{"pdf":"https://content.openalex.org/works/W7215110389.pdf","grobid_xml":"https://content.openalex.org/works/W7215110389.grobid-xml"},"referenced_works_count":40,"referenced_works":["https://openalex.org/W4391631876","https://openalex.org/W4407424870","https://openalex.org/W4414939355","https://openalex.org/W4415274511","https://openalex.org/W4415276156","https://openalex.org/W4415967437","https://openalex.org/W6891914185","https://openalex.org/W7080558933","https://openalex.org/W7118867773","https://openalex.org/W7124432582","https://openalex.org/W7125861034","https://openalex.org/W7125915072","https://openalex.org/W7126034091","https://openalex.org/W7128814765","https://openalex.org/W7130402983","https://openalex.org/W7133318564","https://openalex.org/W7134813685","https://openalex.org/W7141388625","https://openalex.org/W7143342743","https://openalex.org/W7148454362","https://openalex.org/W7153282135","https://openalex.org/W7154206010","https://openalex.org/W7154505894","https://openalex.org/W7155410664","https://openalex.org/W7160313503","https://openalex.org/W7160438937","https://openalex.org/W7160854431","https://openalex.org/W7160893033","https://openalex.org/W7162525570","https://openalex.org/W7163131677","https://openalex.org/W7163590749","https://openalex.org/W7163707237","https://openalex.org/W7164026309","https://openalex.org/W7165056263","https://openalex.org/W7166701728","https://openalex.org/W7166885264","https://openalex.org/W7167630691","https://openalex.org/W7168964786","https://openalex.org/W7170033572","https://openalex.org/W7172299738"],"related_works":["https://openalex.org/W7222327894","https://openalex.org/W7165634569","https://openalex.org/W7221072085","https://openalex.org/W7141388625","https://openalex.org/W7204908525","https://openalex.org/W7165678529","https://openalex.org/W7204437245","https://openalex.org/W7164026309","https://openalex.org/W7221829010","https://openalex.org/W7160893647"],"abstract_inverted_index":{"On-policy":[0],"distillation":[1,33],"(OPD)":[2],"learns":[3],"from":[4,34,44],"teacher":[5],"feedback":[6],"on":[7,148],"student-generated":[8],"responses":[9],"and":[10,25,74,86,107],"has":[11],"shown":[12],"promise":[13],"in":[14,82],"reducing":[15],"forgetting":[16],"relative":[17],"to":[18,90,112,131,138],"supervised":[19],"fine-tuning":[20],"(SFT).":[21],"However,":[22],"its":[23,41],"benefits":[24,106],"fragility":[26,108],"remain":[27],"incompletely":[28],"understood.":[29],"We":[30,65],"study":[31],"sequential":[32],"multiple":[35],"teachers,":[36],"where":[37],"the":[38,45,83,88,105],"student":[39],"minimizes":[40],"average":[42],"divergence":[43,50],"teachers.":[46],"Forward":[47],"Kullback--Leibler":[48],"(KL)":[49],"yields":[51,59],"a":[52,60,120,146],"weighted":[53,62],"arithmetic":[54],"mixture,":[55],"while":[56],"reverse":[57,115],"KL":[58,116],"normalized":[61],"geometric":[63],"aggregate.":[64],"develop":[66],"algorithms":[67],"that":[68,101,133,151],"learn":[69],"these":[70,95],"targets":[71],"under":[72,124],"off-policy":[73],"on-policy":[75],"feedback,":[76,126],"respectively,":[77],"establishing":[78],"logarithmic":[79],"regret":[80],"bounds":[81],"tabular":[84],"setting":[85],"extending":[87],"analysis":[89],"function":[91],"approximation.":[92],"By":[93],"analyzing":[94],"aggregation":[96],"targets,":[97],"we":[98],"identify":[99],"mechanisms":[100],"help":[102],"explain":[103],"both":[104],"of":[109],"OPD.":[110],"Relative":[111],"forward":[113],"KL,":[114],"can":[117,152],"better":[118],"retain":[119],"confident":[121],"expert's":[122],"preferences":[123],"uninformative":[125],"but":[127],"is":[128],"more":[129],"sensitive":[130],"teachers":[132],"assign":[134],"very":[135],"low":[136],"probabilities":[137],"correct":[139],"responses.":[140],"Its":[141],"token-level":[142],"conditionals":[143],"also":[144],"reveal":[145],"dependence":[147],"continuation":[149],"distributions":[150],"favor":[153],"incorrect":[154],"prefixes":[155],"over":[156],"long":[157],"horizons.":[158]},"counts_by_year":[],"updated_date":"2026-11-07T11:25:07.637180","created_date":"2026-10-02T00:00:00"}