{"id":"https://openalex.org/W7166842386","doi":"https://doi.org/10.18653/v1/2026.findings-acl.879","title":"Understanding and Preventing Entropy Collapse in RLVR with On-Policy Entropy Flow Optimization","display_name":"Understanding and Preventing Entropy Collapse in RLVR with On-Policy Entropy Flow Optimization","publication_year":2026,"publication_date":"2026-01-01","ids":{"openalex":"https://openalex.org/W7166842386","doi":"https://doi.org/10.18653/v1/2026.findings-acl.879"},"language":null,"primary_location":{"id":"doi:10.18653/v1/2026.findings-acl.879","is_oa":true,"landing_page_url":"https://doi.org/10.18653/v1/2026.findings-acl.879","pdf_url":"https://aclanthology.org/2026.findings-acl.879.pdf","source":null,"license":"cc-by","license_id":"https://openalex.org/licenses/cc-by","version":"publishedVersion","is_accepted":true,"is_published":true,"raw_source_name":"Findings of the Association for Computational Linguistics: ACL 2026","raw_type":"proceedings-article"},"type":"conference-paper","indexed_in":["crossref"],"open_access":{"is_oa":true,"oa_status":"gold","oa_url":"https://aclanthology.org/2026.findings-acl.879.pdf","any_repository_has_fulltext":null},"authorships":[{"author_position":"first","author":{"id":"https://openalex.org/A5139819852","display_name":"Huimin Xu","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Huimin Xu","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5139802967","display_name":"Shuai Zhao","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Shuai Zhao","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5139715013","display_name":"Xiaobao Wu","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Xiaobao Wu","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"last","author":{"id":"https://openalex.org/A5139773614","display_name":"Anh Tuan Luu","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Anh Tuan Luu","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":0.0,"has_fulltext":true,"cited_by_count":0,"citation_normalized_percentile":{"value":0.83232259,"is_in_top_1_percent":false,"is_in_top_10_percent":false},"cited_by_percentile_year":null,"biblio":{"volume":null,"issue":null,"first_page":"17759","last_page":"17771"},"is_retracted":false,"is_paratext":false,"is_xpac":false,"primary_topic":{"id":"https://openalex.org/T10462","display_name":"Reinforcement Learning in Robotics","score":0.23759999871253967,"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":0.23759999871253967,"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/T11206","display_name":"Model Reduction and Neural Networks","score":0.05909999832510948,"subfield":{"id":"https://openalex.org/subfields/3109","display_name":"Statistical and Nonlinear Physics"},"field":{"id":"https://openalex.org/fields/31","display_name":"Physics and Astronomy"},"domain":{"id":"https://openalex.org/domains/3","display_name":"Physical Sciences"}},{"id":"https://openalex.org/T11689","display_name":"Adversarial Robustness in Machine Learning","score":0.05290000140666962,"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/entropy","display_name":"Entropy (arrow of time)","score":0.5343999862670898},{"id":"https://openalex.org/keywords/flow","display_name":"Flow (mathematics)","score":0.2939000129699707},{"id":"https://openalex.org/keywords/heat-flow","display_name":"Heat flow","score":0.25290000438690186},{"id":"https://openalex.org/keywords/entropy-production","display_name":"Entropy production","score":0.218299999833107}],"concepts":[{"id":"https://openalex.org/C106301342","wikidata":"https://www.wikidata.org/wiki/Q4117933","display_name":"Entropy (arrow of time)","level":2,"score":0.5343999862670898},{"id":"https://openalex.org/C33923547","wikidata":"https://www.wikidata.org/wiki/Q395","display_name":"Mathematics","level":0,"score":0.4350999891757965},{"id":"https://openalex.org/C121864883","wikidata":"https://www.wikidata.org/wiki/Q677916","display_name":"Statistical physics","level":1,"score":0.3779999911785126},{"id":"https://openalex.org/C28826006","wikidata":"https://www.wikidata.org/wiki/Q33521","display_name":"Applied mathematics","level":1,"score":0.3400000035762787},{"id":"https://openalex.org/C41008148","wikidata":"https://www.wikidata.org/wiki/Q21198","display_name":"Computer science","level":0,"score":0.3384000062942505},{"id":"https://openalex.org/C126255220","wikidata":"https://www.wikidata.org/wiki/Q141495","display_name":"Mathematical optimization","level":1,"score":0.3070000112056732},{"id":"https://openalex.org/C38349280","wikidata":"https://www.wikidata.org/wiki/Q1434290","display_name":"Flow (mathematics)","level":2,"score":0.2939000129699707},{"id":"https://openalex.org/C121332964","wikidata":"https://www.wikidata.org/wiki/Q413","display_name":"Physics","level":0,"score":0.2872999906539917},{"id":"https://openalex.org/C2985596519","wikidata":"https://www.wikidata.org/wiki/Q179635","display_name":"Heat flow","level":3,"score":0.25290000438690186},{"id":"https://openalex.org/C207282930","wikidata":"https://www.wikidata.org/wiki/Q5380811","display_name":"Entropy production","level":2,"score":0.218299999833107}],"mesh":[],"locations_count":1,"locations":[{"id":"doi:10.18653/v1/2026.findings-acl.879","is_oa":true,"landing_page_url":"https://doi.org/10.18653/v1/2026.findings-acl.879","pdf_url":"https://aclanthology.org/2026.findings-acl.879.pdf","source":null,"license":"cc-by","license_id":"https://openalex.org/licenses/cc-by","version":"publishedVersion","is_accepted":true,"is_published":true,"raw_source_name":"Findings of the Association for Computational Linguistics: ACL 2026","raw_type":"proceedings-article"}],"best_oa_location":{"id":"doi:10.18653/v1/2026.findings-acl.879","is_oa":true,"landing_page_url":"https://doi.org/10.18653/v1/2026.findings-acl.879","pdf_url":"https://aclanthology.org/2026.findings-acl.879.pdf","source":null,"license":"cc-by","license_id":"https://openalex.org/licenses/cc-by","version":"publishedVersion","is_accepted":true,"is_published":true,"raw_source_name":"Findings of the Association for Computational Linguistics: ACL 2026","raw_type":"proceedings-article"},"sustainable_development_goals":[{"id":"https://metadata.un.org/sdg/13","score":0.5122987031936646,"display_name":"Climate action"}],"awards":[],"funders":[{"id":"https://openalex.org/F4320335534","display_name":"Air Traffic Management Research Institute","ror":null}],"has_content":{"pdf":true,"grobid_xml":true},"content_urls":{"pdf":"https://content.openalex.org/works/W7166842386.pdf","grobid_xml":"https://content.openalex.org/works/W7166842386.grobid-xml"},"referenced_works_count":0,"referenced_works":[],"related_works":[],"abstract_inverted_index":{"Reinforcement":[0],"learning":[1],"with":[2],"verifiable":[3],"rewards":[4],"(RLVR)":[5],"has":[6],"become":[7],"an":[8,118],"effective":[9],"paradigm":[10],"for":[11],"improving":[12],"the":[13,102],"reasoning":[14,144],"ability":[15],"of":[16,93,104],"large":[17],"language":[18],"models.However,":[19],"widely":[20],"used":[21],"RLVR":[22,98],"algorithms,":[23],"such":[24],"as":[25],"GRPO,":[26],"often":[27],"suffer":[28],"from":[29,66],"entropy":[30,41,49,64,69,86,94,106,120,135],"collapse,":[31],"leading":[32],"to":[33,131,134],"premature":[34],"determinism":[35],"and":[36,43,100,127,152],"unstable":[37],"optimization.Existing":[38],"remedies,":[39],"including":[40],"regularization":[42],"ratio-based":[44],"clipping":[45],"heuristics,":[46],"either":[47],"control":[48],"in":[50,82,96],"a":[51,67,83,90],"coarse-grained":[52],"manner":[53],"or":[54],"rely":[55],"on":[56,141],"approximate":[57],"on-policy":[58],"training.In":[59],"this":[60,109],"paper,":[61],"we":[62,111],"revisit":[63],"collapse":[65,95],"token-level":[68],"flow":[70,121],"perspective.Our":[71],"analysis":[72],"reveals":[73],"that":[74,124,147],"entropy-decreasing":[75,128],"tokens":[76],"consistently":[77],"outweigh":[78],"entropy-increasing":[79,126],"ones,":[80],"resulting":[81],"severely":[84],"imbalanced":[85],"flow.This":[87],"perspective":[88],"provides":[89],"unified":[91],"explanation":[92],"existing":[97],"algorithms":[99],"highlights":[101],"importance":[103],"balancing":[105,122],"dynamics.Motivated":[107],"by":[108],"analysis,":[110],"propose":[112],"On-Policy":[113],"Entropy":[114],"Flow":[115],"Optimization":[116],"(OPEFO),":[117],"adaptive":[119],"mechanism":[123],"rescales":[125],"updates":[129],"according":[130],"their":[132],"contributions":[133],"change,":[136],"while":[137],"remaining":[138],"strict":[139],"on-policy.Experiments":[140],"six":[142],"mathematical":[143],"benchmarks":[145],"demonstrate":[146],"OPEFO":[148],"improves":[149],"training":[150],"stability":[151],"final":[153],"performance":[154],"1":[155],".":[156]},"counts_by_year":[],"updated_date":"2026-07-29T14:22:42.915294","created_date":"2026-07-02T00:00:00"}