{"id":"https://openalex.org/W7131395617","doi":"https://doi.org/10.48550/arxiv.2602.19362","title":"LLMs Can Learn to Reason Via Off-Policy RL","display_name":"LLMs Can Learn to Reason Via Off-Policy RL","publication_year":2026,"publication_date":"2026-02-22","ids":{"openalex":"https://openalex.org/W7131395617","doi":"https://doi.org/10.48550/arxiv.2602.19362"},"language":null,"primary_location":{"id":"doi:10.48550/arxiv.2602.19362","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2602.19362","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":"article"},"type":"preprint","indexed_in":["datacite"],"open_access":{"is_oa":true,"oa_status":"green","oa_url":"https://doi.org/10.48550/arxiv.2602.19362","any_repository_has_fulltext":true},"authorships":[{"author_position":"first","author":{"id":"https://openalex.org/A5126669337","display_name":"Daniel Ritter","orcid":null},"institutions":[],"countries":[],"is_corresponding":true,"raw_author_name":"Ritter, Daniel","raw_affiliation_strings":[],"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5095045747","display_name":"Owen Oertell","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Oertell, Owen","raw_affiliation_strings":[],"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5126680364","display_name":"Bradley Guo","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Guo, Bradley","raw_affiliation_strings":[],"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5126839356","display_name":"Jonathan Chang","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Chang, Jonathan","raw_affiliation_strings":[],"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5022968100","display_name":"Kiant\u00e9 Brantley","orcid":"https://orcid.org/0000-0002-8395-594X"},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Brantley, Kiant\u00e9","raw_affiliation_strings":[],"affiliations":[]},{"author_position":"last","author":{"id":"https://openalex.org/A5126797234","display_name":"Wen Sun","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Sun, Wen","raw_affiliation_strings":[],"affiliations":[]}],"institutions":[],"countries_distinct_count":0,"institutions_distinct_count":6,"corresponding_author_ids":["https://openalex.org/A5126669337"],"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":0.2775000035762787,"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":0.2775000035762787,"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/T12535","display_name":"Machine Learning and Data Classification","score":0.125,"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/T13702","display_name":"Machine Learning in Healthcare","score":0.07810000330209732,"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/inference","display_name":"Inference","score":0.8264999985694885},{"id":"https://openalex.org/keywords/reinforcement-learning","display_name":"Reinforcement learning","score":0.612500011920929},{"id":"https://openalex.org/keywords/thompson-sampling","display_name":"Thompson sampling","score":0.46560001373291016},{"id":"https://openalex.org/keywords/sampling","display_name":"Sampling (signal processing)","score":0.41370001435279846},{"id":"https://openalex.org/keywords/policy-learning","display_name":"Policy learning","score":0.3971000015735626},{"id":"https://openalex.org/keywords/training-set","display_name":"Training set","score":0.33899998664855957}],"concepts":[{"id":"https://openalex.org/C2776214188","wikidata":"https://www.wikidata.org/wiki/Q408386","display_name":"Inference","level":2,"score":0.8264999985694885},{"id":"https://openalex.org/C41008148","wikidata":"https://www.wikidata.org/wiki/Q21198","display_name":"Computer science","level":0,"score":0.642300009727478},{"id":"https://openalex.org/C97541855","wikidata":"https://www.wikidata.org/wiki/Q830687","display_name":"Reinforcement learning","level":2,"score":0.612500011920929},{"id":"https://openalex.org/C119857082","wikidata":"https://www.wikidata.org/wiki/Q2539","display_name":"Machine learning","level":1,"score":0.5333999991416931},{"id":"https://openalex.org/C154945302","wikidata":"https://www.wikidata.org/wiki/Q11660","display_name":"Artificial intelligence","level":1,"score":0.5139999985694885},{"id":"https://openalex.org/C73602740","wikidata":"https://www.wikidata.org/wiki/Q7795822","display_name":"Thompson sampling","level":3,"score":0.46560001373291016},{"id":"https://openalex.org/C140779682","wikidata":"https://www.wikidata.org/wiki/Q210868","display_name":"Sampling (signal processing)","level":3,"score":0.41370001435279846},{"id":"https://openalex.org/C2779436431","wikidata":"https://www.wikidata.org/wiki/Q30672407","display_name":"Policy learning","level":2,"score":0.3971000015735626},{"id":"https://openalex.org/C51632099","wikidata":"https://www.wikidata.org/wiki/Q3985153","display_name":"Training set","level":2,"score":0.33899998664855957},{"id":"https://openalex.org/C179518139","wikidata":"https://www.wikidata.org/wiki/Q5140297","display_name":"Coding (social sciences)","level":2,"score":0.3303999900817871},{"id":"https://openalex.org/C91306197","wikidata":"https://www.wikidata.org/wiki/Q45767","display_name":"Competition (biology)","level":2,"score":0.3296000063419342},{"id":"https://openalex.org/C2777211547","wikidata":"https://www.wikidata.org/wiki/Q17141490","display_name":"Training (meteorology)","level":2,"score":0.3172000050544739},{"id":"https://openalex.org/C149782125","wikidata":"https://www.wikidata.org/wiki/Q160039","display_name":"Econometrics","level":1,"score":0.3140000104904175},{"id":"https://openalex.org/C52740198","wikidata":"https://www.wikidata.org/wiki/Q1539564","display_name":"Importance sampling","level":3,"score":0.3034999966621399},{"id":"https://openalex.org/C158600405","wikidata":"https://www.wikidata.org/wiki/Q5054566","display_name":"Causal inference","level":2,"score":0.28630000352859497},{"id":"https://openalex.org/C75778745","wikidata":"https://www.wikidata.org/wiki/Q342626","display_name":"Lag","level":2,"score":0.2563000023365021}],"mesh":[],"locations_count":1,"locations":[{"id":"doi:10.48550/arxiv.2602.19362","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2602.19362","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":"article"}],"best_oa_location":{"id":"doi:10.48550/arxiv.2602.19362","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2602.19362","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":"article"},"sustainable_development_goals":[],"awards":[],"funders":[],"has_content":{"pdf":false,"grobid_xml":false},"content_urls":null,"referenced_works_count":0,"referenced_works":[],"related_works":[],"abstract_inverted_index":{"Reinforcement":[0],"learning":[1],"(RL)":[2],"approaches":[3],"for":[4,160],"Large":[5],"Language":[6],"Models":[7],"(LLMs)":[8],"frequently":[9],"use":[10],"on-policy":[11],"algorithms,":[12],"such":[13],"as":[14],"PPO":[15],"or":[16,62],"GRPO.":[17],"However,":[18],"policy":[19,104],"lag":[20],"from":[21],"distributed":[22],"training":[23,29,68,175],"architectures":[24],"and":[25,30,69,84,119,176],"differences":[26],"between":[27,173],"the":[28,37,67,75,122,155,174],"inference":[31,70,76,177],"policies":[32,71],"break":[33],"this":[34,51,79],"assumption,":[35],"making":[36,50],"data":[38,53],"off-policy":[39,52,88,181],"by":[40,63,72],"design.":[41],"To":[42],"rectify":[43],"this,":[44],"prior":[45,183],"work":[46],"has":[47],"focused":[48],"on":[49,115,131],"appear":[54],"more":[55,64,168,180],"on-policy,":[56],"either":[57],"via":[58,147],"importance":[59,113],"sampling":[60,114],"(IS),":[61],"closely":[65],"aligning":[66],"explicitly":[73],"modifying":[74],"engine.":[77],"In":[78],"work,":[80],"we":[81],"embrace":[82],"off-policyness":[83],"propose":[85],"a":[86,125],"novel":[87],"RL":[89],"algorithm":[90],"that":[91,108,144],"does":[92],"not":[93],"require":[94],"these":[95],"modifications:":[96],"Optimal":[97],"Advantage-based":[98],"Policy":[99],"Optimization":[100],"with":[101,112,165],"Lagged":[102],"Inference":[103],"(OAPL).":[105],"We":[106,140],"show":[107],"OAPL":[109,148,158],"outperforms":[110],"GRPO":[111],"competition":[116],"math":[117],"benchmarks,":[118],"can":[120],"match":[121],"performance":[123],"of":[124,167],"publicly":[126],"available":[127],"coding":[128],"model,":[129],"DeepCoder,":[130],"LiveCodeBench,":[132],"while":[133],"using":[134],"3x":[135],"fewer":[136],"generations":[137],"during":[138],"training.":[139],"further":[141],"empirically":[142],"demonstrate":[143],"models":[145],"trained":[146],"have":[149],"improved":[150],"test":[151],"time":[152],"scaling":[153],"under":[154],"Pass@k":[156],"metric.":[157],"allows":[159],"efficient,":[161],"effective":[162],"post-training":[163],"even":[164],"lags":[166],"than":[169,182],"400":[170],"gradient":[171],"steps":[172],"policies,":[178],"100x":[179],"approaches.":[184]},"counts_by_year":[],"updated_date":"2026-02-26T06:34:08.959763","created_date":"2026-02-26T00:00:00"}