{"id":"https://openalex.org/W7161301568","doi":"https://doi.org/10.48550/arxiv.2605.15113","title":"Learning from Language Feedback via Variational Policy Distillation","display_name":"Learning from Language Feedback via Variational Policy Distillation","publication_year":2026,"publication_date":"2026-05-14","ids":{"openalex":"https://openalex.org/W7161301568","doi":"https://doi.org/10.48550/arxiv.2605.15113"},"language":null,"primary_location":{"id":"doi:10.48550/arxiv.2605.15113","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2605.15113","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":"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"},"type":"preprint","indexed_in":["datacite"],"open_access":{"is_oa":true,"oa_status":"green","oa_url":"https://doi.org/10.48550/arxiv.2605.15113","any_repository_has_fulltext":true},"authorships":[{"author_position":"first","author":{"id":"https://openalex.org/A5136266220","display_name":"Yang Li","orcid":"https://orcid.org/0000-0001-5672-4110"},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Li, Yang","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5081408444","display_name":"Erik Nijkamp","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Nijkamp, Erik","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5048500970","display_name":"Semih Yavuz","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Yavuz, Semih","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"last","author":{"id":"https://openalex.org/A5005443526","display_name":"Shafiq Joty","orcid":"https://orcid.org/0000-0002-9222-2641"},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Joty, Shafiq","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]}],"institutions":[],"countries_distinct_count":0,"institutions_distinct_count":0,"corresponding_author_ids":[],"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":0.2215999960899353,"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":0.2215999960899353,"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/T11948","display_name":"Machine Learning in Materials Science","score":0.15600000321865082,"subfield":{"id":"https://openalex.org/subfields/2505","display_name":"Materials Chemistry"},"field":{"id":"https://openalex.org/fields/25","display_name":"Materials Science"},"domain":{"id":"https://openalex.org/domains/3","display_name":"Physical Sciences"}},{"id":"https://openalex.org/T12072","display_name":"Machine Learning and Algorithms","score":0.09870000183582306,"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/reinforcement-learning","display_name":"Reinforcement learning","score":0.5771999955177307},{"id":"https://openalex.org/keywords/trajectory","display_name":"Trajectory","score":0.5514000058174133},{"id":"https://openalex.org/keywords/distillation","display_name":"Distillation","score":0.5285000205039978},{"id":"https://openalex.org/keywords/outcome","display_name":"Outcome (game theory)","score":0.48919999599456787},{"id":"https://openalex.org/keywords/language-model","display_name":"Language model","score":0.4553999900817871},{"id":"https://openalex.org/keywords/scheme","display_name":"Scheme (mathematics)","score":0.373199999332428}],"concepts":[{"id":"https://openalex.org/C41008148","wikidata":"https://www.wikidata.org/wiki/Q21198","display_name":"Computer science","level":0,"score":0.6739000082015991},{"id":"https://openalex.org/C97541855","wikidata":"https://www.wikidata.org/wiki/Q830687","display_name":"Reinforcement learning","level":2,"score":0.5771999955177307},{"id":"https://openalex.org/C13662910","wikidata":"https://www.wikidata.org/wiki/Q193139","display_name":"Trajectory","level":2,"score":0.5514000058174133},{"id":"https://openalex.org/C204030448","wikidata":"https://www.wikidata.org/wiki/Q101017","display_name":"Distillation","level":2,"score":0.5285000205039978},{"id":"https://openalex.org/C148220186","wikidata":"https://www.wikidata.org/wiki/Q7111912","display_name":"Outcome (game theory)","level":2,"score":0.48919999599456787},{"id":"https://openalex.org/C137293760","wikidata":"https://www.wikidata.org/wiki/Q3621696","display_name":"Language model","level":2,"score":0.4553999900817871},{"id":"https://openalex.org/C154945302","wikidata":"https://www.wikidata.org/wiki/Q11660","display_name":"Artificial intelligence","level":1,"score":0.42340001463890076},{"id":"https://openalex.org/C77618280","wikidata":"https://www.wikidata.org/wiki/Q1155772","display_name":"Scheme (mathematics)","level":2,"score":0.373199999332428},{"id":"https://openalex.org/C48145219","wikidata":"https://www.wikidata.org/wiki/Q1335365","display_name":"Security token","level":2,"score":0.3492000102996826},{"id":"https://openalex.org/C165064840","wikidata":"https://www.wikidata.org/wiki/Q1321061","display_name":"Matching (statistics)","level":2,"score":0.3463999927043915},{"id":"https://openalex.org/C85847156","wikidata":"https://www.wikidata.org/wiki/Q59015987","display_name":"Verifiable secret sharing","level":3,"score":0.34630000591278076},{"id":"https://openalex.org/C77967617","wikidata":"https://www.wikidata.org/wiki/Q4677561","display_name":"Active learning (machine learning)","level":2,"score":0.3393999934196472},{"id":"https://openalex.org/C2776760102","wikidata":"https://www.wikidata.org/wiki/Q5139990","display_name":"Code (set theory)","level":3,"score":0.30169999599456787},{"id":"https://openalex.org/C119857082","wikidata":"https://www.wikidata.org/wiki/Q2539","display_name":"Machine learning","level":1,"score":0.2971999943256378}],"mesh":[],"locations_count":1,"locations":[{"id":"doi:10.48550/arxiv.2605.15113","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2605.15113","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":"cc-by","license_id":"https://openalex.org/licenses/cc-by","version":null,"is_accepted":false,"is_published":null,"raw_source_name":null,"raw_type":"Preprint"}],"best_oa_location":{"id":"doi:10.48550/arxiv.2605.15113","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2605.15113","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":"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"},"sustainable_development_goals":[{"score":0.761086106300354,"id":"https://metadata.un.org/sdg/4","display_name":"Quality Education"}],"awards":[],"funders":[],"has_content":{"grobid_xml":false,"pdf":false},"content_urls":null,"referenced_works_count":0,"referenced_works":[],"related_works":[],"abstract_inverted_index":{"Reinforcement":[0],"learning":[1,77],"from":[2,7,78,142],"verifiable":[3],"rewards":[4],"(RLVR)":[5],"suffers":[6],"sparse":[8],"outcome":[9],"signals,":[10],"creating":[11],"severe":[12],"exploration":[13],"bottlenecks":[14],"on":[15,40,99,127,159,181],"complex":[16],"reasoning":[17,161,184],"tasks.":[18],"Recent":[19],"on-policy":[20,130],"self-distillation":[21,174,195],"methods":[22],"attempt":[23],"to":[24,31,45,138,197],"address":[25],"this":[26,123],"by":[27,177],"utilizing":[28],"language":[29,79],"feedback":[30,80,109,158],"generate":[32],"dense,":[33],"token-level":[34],"supervision.":[35],"However,":[36],"these":[37],"approaches":[38],"rely":[39],"a":[41,73,82,111],"fixed,":[42],"passive":[43,150],"teacher":[44,95],"interpret":[46],"the":[47,50,54,92,94,118,120,135,147,190],"feedback.":[48],"As":[49],"student":[51,121],"policy":[52],"improves,":[53],"teacher's":[55,136],"zero-shot":[56],"assessment":[57],"capabilities":[58],"plateau,":[59],"ultimately":[60],"halting":[61],"further":[62],"learning.":[63],"To":[64],"overcome":[65],"this,":[66],"we":[67,188],"propose":[68],"Variational":[69,83],"Policy":[70],"Distillation":[71],"(VPD),":[72],"framework":[74,180],"that":[75],"formalizes":[76],"as":[81],"Expectation-Maximization":[84],"(EM)":[85],"problem.":[86],"VPD":[87,145,166],"co-evolves":[88],"both":[89,169],"policies:":[90],"in":[91],"E-step,":[93],"is":[96],"actively":[97],"refined":[98],"trajectory":[100],"outcomes":[101],"via":[102],"an":[103],"adaptive":[104],"trust-region":[105],"update,":[106],"translating":[107],"textual":[108,143],"into":[110],"dynamically":[112],"improved":[113],"target":[114],"token":[115],"distribution.":[116],"In":[117],"M-step,":[119],"internalizes":[122],"dense":[124],"distributional":[125],"guidance":[126],"its":[128],"own":[129],"rollouts.":[131],"By":[132],"continuously":[133],"improving":[134],"ability":[137],"extract":[139],"actionable":[140],"signals":[141],"critique,":[144],"overcomes":[146],"limitations":[148],"of":[149,156,193],"distillation.":[151],"Evaluated":[152],"across":[153],"diverse":[154],"sources":[155],"diagnostic":[157],"scientific":[160],"and":[162,172,185],"code":[163],"generation":[164],"tasks,":[165],"consistently":[167],"outperforms":[168],"standard":[170],"RLVR":[171],"existing":[173],"baselines.":[175],"Finally,":[176],"stress-testing":[178],"our":[179],"rigid":[182],"mathematical":[183],"cold-start":[186],"regimes,":[187],"illuminate":[189],"fundamental":[191],"bounds":[192],"feedback-driven":[194],"compared":[196],"pure":[198],"environment-driven":[199],"RL.":[200]},"counts_by_year":[],"updated_date":"2026-07-01T06:00:48.157686","created_date":"2026-05-16T00:00:00"}