{"id":"https://openalex.org/W4415084114","doi":"https://doi.org/10.48550/arxiv.2503.01067","title":"All Roads Lead to Likelihood: The Value of Reinforcement Learning in Fine-Tuning","display_name":"All Roads Lead to Likelihood: The Value of Reinforcement Learning in Fine-Tuning","publication_year":2025,"publication_date":"2025-03-03","ids":{"openalex":"https://openalex.org/W4415084114","doi":"https://doi.org/10.48550/arxiv.2503.01067"},"language":"en","primary_location":{"id":"pmh:oai:arXiv.org:2503.01067","doi":null,"is_oa":true,"landing_page_url":"http://arxiv.org/abs/2503.01067","pdf_url":"https://arxiv.org/pdf/2503.01067","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/2503.01067","any_repository_has_fulltext":true},"authorships":[{"author_position":"first","author":{"id":"https://openalex.org/A5111918435","display_name":"Gokul Swamy","orcid":null,"observed_orcids":[]},"institutions":[{"id":"https://openalex.org/I74973139","display_name":"Carnegie Mellon University","ror":"https://ror.org/05x2bcf33","country_code":"US","type":"education","lineage":["https://openalex.org/I74973139"]}],"countries":["US"],"is_corresponding":false,"raw_author_name":"Swamy, Gokul","raw_affiliation_strings":["Carnegie Mellon University"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"Carnegie Mellon University","institution_ids":["https://openalex.org/I74973139"]}]},{"author_position":"middle","author":{"id":"https://openalex.org/A5057995939","display_name":"Sanjiban Choudhury","orcid":"https://orcid.org/0000-0003-2762-8888","observed_orcids":["https://orcid.org/0000-0003-2762-8888"]},"institutions":[{"id":"https://openalex.org/I205783295","display_name":"Cornell University","ror":"https://ror.org/05bnh6r87","country_code":"US","type":"education","lineage":["https://openalex.org/I205783295"]}],"countries":["US"],"is_corresponding":false,"raw_author_name":"Choudhury, Sanjiban","raw_affiliation_strings":["Cornell University"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"Cornell University","institution_ids":["https://openalex.org/I205783295"]}]},{"author_position":"middle","author":{"id":"https://openalex.org/A5032266950","display_name":"Wen Yao Sun","orcid":"https://orcid.org/0000-0001-7086-4910","observed_orcids":["https://orcid.org/0000-0001-7086-4910"]},"institutions":[{"id":"https://openalex.org/I205783295","display_name":"Cornell University","ror":"https://ror.org/05bnh6r87","country_code":"US","type":"education","lineage":["https://openalex.org/I205783295"]}],"countries":["US"],"is_corresponding":false,"raw_author_name":"Sun, Wen","raw_affiliation_strings":["Cornell University"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"Cornell University","institution_ids":["https://openalex.org/I205783295"]}]},{"author_position":"middle","author":{"id":"https://openalex.org/A5001070941","display_name":"Zhiwei Steven Wu","orcid":"https://orcid.org/0000-0002-8125-8227","observed_orcids":["https://orcid.org/0000-0002-8125-8227"]},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Wu, Zhiwei Steven","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"last","author":{"id":"https://openalex.org/A5112411386","display_name":"James Andrew Bagnell","orcid":null,"observed_orcids":[]},"institutions":[{"id":"https://openalex.org/I74973139","display_name":"Carnegie Mellon University","ror":"https://ror.org/05x2bcf33","country_code":"US","type":"education","lineage":["https://openalex.org/I74973139"]}],"countries":["US"],"is_corresponding":false,"raw_author_name":"Bagnell, J. Andrew","raw_affiliation_strings":["Carnegie Mellon University"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"Carnegie Mellon University","institution_ids":["https://openalex.org/I74973139"]}]}],"institutions":[],"countries_distinct_count":1,"institutions_distinct_count":2,"corresponding_author_ids":[],"corresponding_institution_ids":[],"apc_list":null,"apc_paid":null,"fwci":null,"has_fulltext":true,"cited_by_count":57,"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":0.6464999914169312,"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":0.6464999914169312,"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/T10028","display_name":"Topic Modeling","score":0.14030000567436218,"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":0.062199998646974564,"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/foundation-model-fine-tuning","display_name":"foundation model fine-tuning","score":0.8489999771118164},{"id":"https://openalex.org/keywords/reward-modeling","display_name":"reward modeling","score":0.6869999766349792},{"id":"https://openalex.org/keywords/two-stage-training","display_name":"two-stage training","score":0.6359999775886536},{"id":"https://openalex.org/keywords/maximum-likelihood-estimation","display_name":"maximum likelihood estimation","score":0.7149999737739563},{"id":"https://openalex.org/keywords/information-theoretic-limits","display_name":"information-theoretic limits","score":0.6700000166893005},{"id":"https://openalex.org/keywords/preference-data","display_name":"preference data","score":0.6010000109672546},{"id":"https://openalex.org/keywords/online-feedback","display_name":"online feedback","score":0.3930000066757202}],"concepts":[{"id":"https://openalex.org/C97541855","wikidata":"https://www.wikidata.org/wiki/Q830687","display_name":"Reinforcement learning","level":2,"score":0.7116000056266785},{"id":"https://openalex.org/C2776207758","wikidata":"https://www.wikidata.org/wiki/Q5303302","display_name":"Downstream (manufacturing)","level":2,"score":0.5680000185966492},{"id":"https://openalex.org/C2776291640","wikidata":"https://www.wikidata.org/wiki/Q2912517","display_name":"Value (mathematics)","level":2,"score":0.5666000247001648},{"id":"https://openalex.org/C41008148","wikidata":"https://www.wikidata.org/wiki/Q21198","display_name":"Computer science","level":0,"score":0.5649999976158142},{"id":"https://openalex.org/C190839683","wikidata":"https://www.wikidata.org/wiki/Q2448197","display_name":"Train","level":2,"score":0.5630999803543091},{"id":"https://openalex.org/C2780586882","wikidata":"https://www.wikidata.org/wiki/Q7520643","display_name":"Simple (philosophy)","level":2,"score":0.5209000110626221},{"id":"https://openalex.org/C2778572836","wikidata":"https://www.wikidata.org/wiki/Q380933","display_name":"Space (punctuation)","level":2,"score":0.5099999904632568},{"id":"https://openalex.org/C92424840","wikidata":"https://www.wikidata.org/wiki/Q7912772","display_name":"Value of information","level":2,"score":0.48969998955726624},{"id":"https://openalex.org/C2781249084","wikidata":"https://www.wikidata.org/wiki/Q908656","display_name":"Preference","level":2,"score":0.45989999175071716},{"id":"https://openalex.org/C154945302","wikidata":"https://www.wikidata.org/wiki/Q11660","display_name":"Artificial intelligence","level":1,"score":0.44110000133514404},{"id":"https://openalex.org/C67203356","wikidata":"https://www.wikidata.org/wiki/Q1321905","display_name":"Reinforcement","level":2,"score":0.40869998931884766},{"id":"https://openalex.org/C2777093003","wikidata":"https://www.wikidata.org/wiki/Q6508345","display_name":"Lead (geology)","level":2,"score":0.3912000060081482},{"id":"https://openalex.org/C119857082","wikidata":"https://www.wikidata.org/wiki/Q2539","display_name":"Machine learning","level":1,"score":0.3513000011444092},{"id":"https://openalex.org/C2780102126","wikidata":"https://www.wikidata.org/wiki/Q10928179","display_name":"Online and offline","level":2,"score":0.3418000042438507},{"id":"https://openalex.org/C2780966255","wikidata":"https://www.wikidata.org/wiki/Q5474306","display_name":"Foundation (evidence)","level":2,"score":0.3301999866962433},{"id":"https://openalex.org/C113336015","wikidata":"https://www.wikidata.org/wiki/Q574010","display_name":"Complete information","level":2,"score":0.32030001282691956},{"id":"https://openalex.org/C75553542","wikidata":"https://www.wikidata.org/wiki/Q178161","display_name":"A priori and a posteriori","level":2,"score":0.3091999888420105},{"id":"https://openalex.org/C177264268","wikidata":"https://www.wikidata.org/wiki/Q1514741","display_name":"Set (abstract data type)","level":2,"score":0.30630001425743103},{"id":"https://openalex.org/C120936955","wikidata":"https://www.wikidata.org/wiki/Q2155640","display_name":"Empirical research","level":2,"score":0.30559998750686646},{"id":"https://openalex.org/C2777851325","wikidata":"https://www.wikidata.org/wiki/Q7094102","display_name":"Online model","level":2,"score":0.28619998693466187},{"id":"https://openalex.org/C87868495","wikidata":"https://www.wikidata.org/wiki/Q750843","display_name":"Information processing","level":2,"score":0.27160000801086426},{"id":"https://openalex.org/C149782125","wikidata":"https://www.wikidata.org/wiki/Q160039","display_name":"Econometrics","level":1,"score":0.26980000734329224},{"id":"https://openalex.org/C51632099","wikidata":"https://www.wikidata.org/wiki/Q3985153","display_name":"Training set","level":2,"score":0.2632000148296356},{"id":"https://openalex.org/C13662910","wikidata":"https://www.wikidata.org/wiki/Q193139","display_name":"Trajectory","level":2,"score":0.2567000091075897}],"mesh":[],"locations_count":2,"locations":[{"id":"pmh:oai:arXiv.org:2503.01067","doi":null,"is_oa":true,"landing_page_url":"http://arxiv.org/abs/2503.01067","pdf_url":"https://arxiv.org/pdf/2503.01067","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.2503.01067","doi":"https://doi.org/10.48550/arxiv.2503.01067","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2503.01067","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":null,"license_id":null,"version":null,"is_accepted":false,"is_published":false,"raw_source_name":null,"raw_type":"Preprint"}],"best_oa_location":{"id":"pmh:oai:arXiv.org:2503.01067","doi":null,"is_oa":true,"landing_page_url":"http://arxiv.org/abs/2503.01067","pdf_url":"https://arxiv.org/pdf/2503.01067","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":[],"study_designs":[],"awards":[{"id":"https://openalex.org/G3167606647","display_name":"RI: Small: Towards Provably Efficient Representation Learning in Reinforcement Learning via Rich Function Approximation","funder_award_id":"2154711","funder_id":"https://openalex.org/F4320306076","funder_display_name":"National Science Foundation"},{"id":"https://openalex.org/G5453458660","display_name":"Collaborative Research: RI: Medium: Superhuman Imitation Learning from Heterogeneous Demonstrations","funder_award_id":"2312956","funder_id":"https://openalex.org/F4320306076","funder_display_name":"National Science Foundation"},{"id":"https://openalex.org/G7439794172","display_name":"Collaborative Research: Inverse Task Planning from Few-Shot Vision Language Demonstrations","funder_award_id":"2327973","funder_id":"https://openalex.org/F4320306076","funder_display_name":"National Science Foundation"},{"id":"https://openalex.org/G8153128647","display_name":"CAREER: Towards Real-world Reinforcement Learning","funder_award_id":"2339395","funder_id":"https://openalex.org/F4320306076","funder_display_name":"National Science Foundation"}],"funders":[{"id":"https://openalex.org/F4320306076","display_name":"National Science Foundation","ror":"https://ror.org/021nxhr62"},{"id":"https://openalex.org/F4320332180","display_name":"Defense Advanced Research Projects Agency","ror":"https://ror.org/02caytj08"},{"id":"https://openalex.org/F4320337345","display_name":"Office of Naval Research","ror":"https://ror.org/00rk2pe57"}],"has_content":{"pdf":true,"grobid_xml":true},"content_urls":{"pdf":"https://content.openalex.org/works/W4415084114.pdf","grobid_xml":"https://content.openalex.org/works/W4415084114.grobid-xml"},"referenced_works_count":45,"referenced_works":["https://openalex.org/W2155831576","https://openalex.org/W2736601468","https://openalex.org/W2897097528","https://openalex.org/W2908510526","https://openalex.org/W4223908421","https://openalex.org/W4327810158","https://openalex.org/W4377121527","https://openalex.org/W4380137125","https://openalex.org/W4386559757","https://openalex.org/W4387801352","https://openalex.org/W4387947637","https://openalex.org/W4389821477","https://openalex.org/W4390573302","https://openalex.org/W4390723302","https://openalex.org/W4391673254","https://openalex.org/W4392679398","https://openalex.org/W4392828102","https://openalex.org/W4393063868","https://openalex.org/W4394906585","https://openalex.org/W4394972204","https://openalex.org/W4395444023","https://openalex.org/W4396945613","https://openalex.org/W4399061478","https://openalex.org/W4399198758","https://openalex.org/W4399198767","https://openalex.org/W4401306886","https://openalex.org/W4403346252","https://openalex.org/W4403853601","https://openalex.org/W4403964131","https://openalex.org/W4404792842","https://openalex.org/W4405035625","https://openalex.org/W4405035967","https://openalex.org/W4405766390","https://openalex.org/W4406071657","https://openalex.org/W4406231668","https://openalex.org/W4406961529","https://openalex.org/W4407632903","https://openalex.org/W4407759093","https://openalex.org/W4414281281","https://openalex.org/W4415329957","https://openalex.org/W4415795501","https://openalex.org/W4415797532","https://openalex.org/W4415800425","https://openalex.org/W6891892315","https://openalex.org/W7133208539"],"related_works":["https://openalex.org/W7080122776","https://openalex.org/W4415936065","https://openalex.org/W4402712402","https://openalex.org/W4407759720","https://openalex.org/W7171801068","https://openalex.org/W6910652885","https://openalex.org/W4377927976","https://openalex.org/W7165685735","https://openalex.org/W4393335877","https://openalex.org/W4415795502"],"abstract_inverted_index":{"From":[0],"a":[1,20,30,50,85,134,181],"first-principles":[2],"perspective,":[3,76],"it":[4,42,138],"may":[5],"seem":[6],"odd":[7],"that":[8,130,163],"the":[9,60,106,119,124,128,144,150,155,185],"strongest":[10],"results":[11],"in":[12,110],"foundation":[13],"model":[14,32,87],"fine-tuning":[15],"(FT)":[16],"are":[17,164],"achieved":[18],"via":[19,66,82,94],"relatively":[21,140,145,168],"complex,":[22],"two-stage":[23,173],"training":[24],"procedure.":[25],"Specifically,":[26],"one":[27],"first":[28],"trains":[29],"reward":[31,86],"(RM)":[33],"on":[34,63,105,131],"some":[35],"dataset":[36,65],"(e.g.,":[37],"human":[38],"preferences)":[39],"before":[40],"using":[41],"to":[43,142,178],"provide":[44],"online":[45,174],"feedback":[46],"as":[47],"part":[48],"of":[49,108,184,188],"downstream":[51,156],"reinforcement":[52],"learning":[53],"(RL)":[54],"procedure,":[55],"rather":[56],"than":[57,193],"directly":[58],"optimizing":[59],"policy":[61],"parameters":[62],"said":[64],"offline":[67,194],"maximum":[68],"likelihood":[69],"estimation.":[70],"In":[71],"fact,":[72],"from":[73,149],"an":[74],"information-theoretic":[75],"we":[77,101,122],"can":[78],"only":[79,159,176],"lose":[80],"information":[81,93],"passing":[83],"through":[84,112],"and":[88,115],"cannot":[89],"create":[90],"any":[91],"new":[92],"on-policy":[95],"sampling.":[96],"To":[97],"explain":[98],"this":[99],"discrepancy,":[100],"scrutinize":[102],"several":[103],"hypotheses":[104,120],"value":[107],"RL":[109,157],"FT":[111,175],"both":[113],"theoretical":[114],"empirical":[116],"lenses.":[117],"Of":[118],"considered,":[121],"find":[123],"most":[125],"support":[126],"for":[127,166],"explanation":[129],"problems":[132],"with":[133],"generation-verification":[135],"gap,":[136],"(1)":[137],"is":[139],"easy":[141],"learn":[143],"simple":[146,169],"RM":[147],"(verifier)":[148],"preference":[151],"data.":[152],"Then,":[153],"(2)":[154],"procedure":[158],"returns":[160],"policies":[161],"(generators)":[162],"optimal":[165],"such":[167],"verifiers.":[170],"Thus,":[171],"end-to-end,":[172],"has":[177],"search":[179],"over":[180],"reduced":[182],"subset":[183],"full":[186],"space":[187],"policies,":[189],"requiring":[190],"less":[191],"data":[192],"FT.":[195]},"counts_by_year":[{"year":2026,"cited_by_count":23},{"year":2025,"cited_by_count":34}],"updated_date":"2026-11-09T13:22:03.289127","created_date":"2025-10-12T00:00:00"}