{"id":"https://openalex.org/W7155051642","doi":"https://doi.org/10.48550/arxiv.2604.17535","title":"OPSDL: On-Policy Self-Distillation for Long-Context Language Models","display_name":"OPSDL: On-Policy Self-Distillation for Long-Context Language Models","publication_year":2026,"publication_date":"2026-04-19","ids":{"openalex":"https://openalex.org/W7155051642","doi":"https://doi.org/10.48550/arxiv.2604.17535"},"language":null,"primary_location":{"id":"doi:10.48550/arxiv.2604.17535","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2604.17535","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.2604.17535","any_repository_has_fulltext":true},"authorships":[{"author_position":"first","author":{"id":"https://openalex.org/A5134157576","display_name":"Xinsen Zhang","orcid":null},"institutions":[],"countries":[],"is_corresponding":true,"raw_author_name":"Zhang, Xinsen","raw_affiliation_strings":[],"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5134121373","display_name":"Zhenkai Ding","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Ding, Zhenkai","raw_affiliation_strings":[],"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5081439632","display_name":"Tianjun Pan","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Pan, Tianjun","raw_affiliation_strings":[],"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5134150129","display_name":"Run Yang","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Yang, Run","raw_affiliation_strings":[],"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5134164428","display_name":"Chun Kang","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Kang, Chun","raw_affiliation_strings":[],"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5129737859","display_name":"Xue Xiong","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Xiong, Xue","raw_affiliation_strings":[],"affiliations":[]},{"author_position":"last","author":{"id":"https://openalex.org/A5134160714","display_name":"Jingnan Gu","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Gu, Jingnan","raw_affiliation_strings":[],"affiliations":[]}],"institutions":[],"countries_distinct_count":0,"institutions_distinct_count":7,"corresponding_author_ids":["https://openalex.org/A5134157576"],"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":0.2741999924182892,"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":0.2741999924182892,"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":0.15649999678134918,"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/T11714","display_name":"Multimodal Machine Learning Applications","score":0.0746999979019165,"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/language-model","display_name":"Language model","score":0.682200014591217},{"id":"https://openalex.org/keywords/context","display_name":"Context (archaeology)","score":0.6765000224113464},{"id":"https://openalex.org/keywords/divergence","display_name":"Divergence (linguistics)","score":0.5958999991416931},{"id":"https://openalex.org/keywords/scalability","display_name":"Scalability","score":0.5958999991416931},{"id":"https://openalex.org/keywords/language-understanding","display_name":"Language understanding","score":0.3991999924182892},{"id":"https://openalex.org/keywords/sample","display_name":"Sample (material)","score":0.3831999897956848}],"concepts":[{"id":"https://openalex.org/C41008148","wikidata":"https://www.wikidata.org/wiki/Q21198","display_name":"Computer science","level":0,"score":0.779699981212616},{"id":"https://openalex.org/C137293760","wikidata":"https://www.wikidata.org/wiki/Q3621696","display_name":"Language model","level":2,"score":0.682200014591217},{"id":"https://openalex.org/C2779343474","wikidata":"https://www.wikidata.org/wiki/Q3109175","display_name":"Context (archaeology)","level":2,"score":0.6765000224113464},{"id":"https://openalex.org/C48044578","wikidata":"https://www.wikidata.org/wiki/Q727490","display_name":"Scalability","level":2,"score":0.5958999991416931},{"id":"https://openalex.org/C207390915","wikidata":"https://www.wikidata.org/wiki/Q1230525","display_name":"Divergence (linguistics)","level":2,"score":0.5958999991416931},{"id":"https://openalex.org/C154945302","wikidata":"https://www.wikidata.org/wiki/Q11660","display_name":"Artificial intelligence","level":1,"score":0.53329998254776},{"id":"https://openalex.org/C2983448237","wikidata":"https://www.wikidata.org/wiki/Q1078276","display_name":"Language understanding","level":2,"score":0.3991999924182892},{"id":"https://openalex.org/C119857082","wikidata":"https://www.wikidata.org/wiki/Q2539","display_name":"Machine learning","level":1,"score":0.39579999446868896},{"id":"https://openalex.org/C198531522","wikidata":"https://www.wikidata.org/wiki/Q485146","display_name":"Sample (material)","level":2,"score":0.3831999897956848},{"id":"https://openalex.org/C204321447","wikidata":"https://www.wikidata.org/wiki/Q30642","display_name":"Natural language processing","level":1,"score":0.37709999084472656},{"id":"https://openalex.org/C204323151","wikidata":"https://www.wikidata.org/wiki/Q905424","display_name":"Range (aeronautics)","level":2,"score":0.366100013256073},{"id":"https://openalex.org/C67186912","wikidata":"https://www.wikidata.org/wiki/Q367664","display_name":"Data modeling","level":2,"score":0.335999995470047},{"id":"https://openalex.org/C2779843651","wikidata":"https://www.wikidata.org/wiki/Q7390335","display_name":"SIGNAL (programming language)","level":2,"score":0.33469998836517334},{"id":"https://openalex.org/C183322885","wikidata":"https://www.wikidata.org/wiki/Q17007702","display_name":"Context model","level":3,"score":0.3249000012874603},{"id":"https://openalex.org/C112972136","wikidata":"https://www.wikidata.org/wiki/Q7595718","display_name":"Stability (learning theory)","level":2,"score":0.25940001010894775}],"mesh":[],"locations_count":1,"locations":[{"id":"doi:10.48550/arxiv.2604.17535","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2604.17535","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.2604.17535","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2604.17535","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":[{"id":"https://metadata.un.org/sdg/4","display_name":"Quality Education","score":0.7198008894920349}],"awards":[],"funders":[],"has_content":{"grobid_xml":false,"pdf":false},"content_urls":null,"referenced_works_count":0,"referenced_works":[],"related_works":[],"abstract_inverted_index":{"Extending":[0],"the":[1,53,70,82,107,111,123,196],"effective":[2],"context":[3,168],"length":[4],"of":[5,56,134,153,198],"large":[6],"language":[7],"models":[8,154],"(LLMs)":[9],"remains":[10],"a":[11,78,90,151,201],"central":[12],"challenge":[13],"for":[14,51,206],"real-world":[15],"applications.":[16],"While":[17],"recent":[18,60],"post-training":[19,172],"methods":[20,62],"have":[21],"made":[22],"progress":[23],"in":[24,97],"long-context":[25,98,148,207],"scaling,":[26],"they":[27],"either":[28],"rely":[29,68],"on":[30,69,106,147],"high-quality":[31],"supervision":[32,115],"data":[33],"or":[34],"sparse":[35],"sequence-level":[36],"rewards,":[37],"leading":[38],"to":[39,75,92,157],"unstable":[40],"and":[41,67,137,163,177,203],"inefficient":[42],"optimization.":[43],"We":[44,144],"propose":[45],"OPSDL,":[46],"an":[47],"On-Policy":[48],"Self-Distillation":[49],"method":[50],"enhancing":[52],"Long-context":[54],"capabilities":[55],"LLMs.":[57],"Unlike":[58],"other":[59],"self-distillation":[61],"that":[63],"inject":[64],"privileged":[65],"information":[66],"model's":[71,83],"in-context":[72],"learning":[73],"ability":[74],"act":[76],"as":[77,89,175,200],"teacher,":[79],"OPSDL":[80,146,199],"leverages":[81],"own":[84,95],"inherently":[85],"strong":[86],"short-context":[87,191],"capability":[88],"self-teacher":[91,112],"supervise":[93],"its":[94],"generation":[96],"scenarios.":[99],"The":[100],"model":[101],"first":[102],"generates":[103],"responses":[104],"conditioned":[105],"full":[108],"long-context,":[109],"then":[110],"provides":[113],"per-token":[114],"signals":[116],"via":[117],"point-wise":[118],"reverse":[119],"KL":[120],"divergence":[121],"under":[122],"relevant":[124,135],"extracted":[125],"short-context.":[126],"This":[127],"dense":[128],"token-level":[129],"signal":[130],"encourages":[131],"faithful":[132],"use":[133],"evidence":[136],"mitigates":[138],"hallucinations":[139],"induced":[140],"by":[141],"irrelevant":[142],"context.":[143],"evaluate":[145],"benchmarks":[149],"across":[150,166],"range":[152],"from":[155],"7B":[156],"32B":[158],"parameters.":[159],"Results":[160],"show":[161],"consistent":[162],"substantial":[164],"improvements":[165],"varying":[167],"lengths,":[169],"outperforming":[170],"standard":[171],"approaches":[173],"such":[174],"SFT":[176],"DPO":[178],"with":[179],"higher":[180],"sample":[181],"efficiency.":[182],"Notably,":[183],"these":[184],"gains":[185],"are":[186],"achieved":[187],"without":[188],"degrading":[189],"general":[190],"performance.":[192],"These":[193],"findings":[194],"highlight":[195],"effectiveness":[197],"scalable":[202],"stable":[204],"approach":[205],"learning.":[208]},"counts_by_year":[],"updated_date":"2026-04-22T06:07:44.442478","created_date":"2026-04-22T00:00:00"}