{"id":"https://openalex.org/W7155416398","doi":"https://doi.org/10.48550/arxiv.2604.19857","title":"Rethinking Reinforcement Fine-Tuning in LVLM: Convergence, Reward Decomposition, and Generalization","display_name":"Rethinking Reinforcement Fine-Tuning in LVLM: Convergence, Reward Decomposition, and Generalization","publication_year":2026,"publication_date":"2026-04-21","ids":{"openalex":"https://openalex.org/W7155416398","doi":"https://doi.org/10.48550/arxiv.2604.19857"},"language":null,"primary_location":{"id":"doi:10.48550/arxiv.2604.19857","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2604.19857","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,"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"},"type":"preprint","indexed_in":["datacite"],"open_access":{"is_oa":true,"oa_status":"green","oa_url":"https://doi.org/10.48550/arxiv.2604.19857","any_repository_has_fulltext":true},"authorships":[{"author_position":"first","author":{"id":"https://openalex.org/A5021383025","display_name":"Carter Adams","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Adams, Carter","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5029891035","display_name":"Rafael Oliveira","orcid":"https://orcid.org/0000-0002-3586-5026"},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Oliveira, Rafael","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5134415241","display_name":"Gabriel Almeida","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Almeida, Gabriel","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"last","author":{"id":"https://openalex.org/A5134388915","display_name":"Sofia Torres","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Torres, Sofia","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/T10462","display_name":"Reinforcement Learning in Robotics","score":0.41429999470710754,"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":0.41429999470710754,"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":0.30959999561309814,"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"}},{"id":"https://openalex.org/T10653","display_name":"Robot Manipulation and Learning","score":0.031700000166893005,"subfield":{"id":"https://openalex.org/subfields/2207","display_name":"Control and Systems Engineering"},"field":{"id":"https://openalex.org/fields/22","display_name":"Engineering"},"domain":{"id":"https://openalex.org/domains/3","display_name":"Physical Sciences"}}],"keywords":[{"id":"https://openalex.org/keywords/generalization","display_name":"Generalization","score":0.7605999708175659},{"id":"https://openalex.org/keywords/verifiable-secret-sharing","display_name":"Verifiable secret sharing","score":0.715399980545044},{"id":"https://openalex.org/keywords/set","display_name":"Set (abstract data type)","score":0.550599992275238},{"id":"https://openalex.org/keywords/markov-decision-process","display_name":"Markov decision process","score":0.5163000226020813},{"id":"https://openalex.org/keywords/convergence","display_name":"Convergence (economics)","score":0.5013999938964844},{"id":"https://openalex.org/keywords/decomposition","display_name":"Decomposition","score":0.46549999713897705},{"id":"https://openalex.org/keywords/reinforcement-learning","display_name":"Reinforcement learning","score":0.46219998598098755},{"id":"https://openalex.org/keywords/representation","display_name":"Representation (politics)","score":0.4584999978542328}],"concepts":[{"id":"https://openalex.org/C177148314","wikidata":"https://www.wikidata.org/wiki/Q170084","display_name":"Generalization","level":2,"score":0.7605999708175659},{"id":"https://openalex.org/C85847156","wikidata":"https://www.wikidata.org/wiki/Q59015987","display_name":"Verifiable secret sharing","level":3,"score":0.715399980545044},{"id":"https://openalex.org/C177264268","wikidata":"https://www.wikidata.org/wiki/Q1514741","display_name":"Set (abstract data type)","level":2,"score":0.550599992275238},{"id":"https://openalex.org/C41008148","wikidata":"https://www.wikidata.org/wiki/Q21198","display_name":"Computer science","level":0,"score":0.5401999950408936},{"id":"https://openalex.org/C106189395","wikidata":"https://www.wikidata.org/wiki/Q176789","display_name":"Markov decision process","level":3,"score":0.5163000226020813},{"id":"https://openalex.org/C2777303404","wikidata":"https://www.wikidata.org/wiki/Q759757","display_name":"Convergence (economics)","level":2,"score":0.5013999938964844},{"id":"https://openalex.org/C124681953","wikidata":"https://www.wikidata.org/wiki/Q339062","display_name":"Decomposition","level":2,"score":0.46549999713897705},{"id":"https://openalex.org/C97541855","wikidata":"https://www.wikidata.org/wiki/Q830687","display_name":"Reinforcement learning","level":2,"score":0.46219998598098755},{"id":"https://openalex.org/C2776359362","wikidata":"https://www.wikidata.org/wiki/Q2145286","display_name":"Representation (politics)","level":3,"score":0.4584999978542328},{"id":"https://openalex.org/C98763669","wikidata":"https://www.wikidata.org/wiki/Q176645","display_name":"Markov chain","level":2,"score":0.42719998955726624},{"id":"https://openalex.org/C159886148","wikidata":"https://www.wikidata.org/wiki/Q176645","display_name":"Markov process","level":2,"score":0.42649999260902405},{"id":"https://openalex.org/C12713177","wikidata":"https://www.wikidata.org/wiki/Q1900281","display_name":"Perspective (graphical)","level":2,"score":0.41040000319480896},{"id":"https://openalex.org/C28719098","wikidata":"https://www.wikidata.org/wiki/Q44946","display_name":"Point (geometry)","level":2,"score":0.39640000462532043},{"id":"https://openalex.org/C67203356","wikidata":"https://www.wikidata.org/wiki/Q1321905","display_name":"Reinforcement","level":2,"score":0.3799999952316284},{"id":"https://openalex.org/C112972136","wikidata":"https://www.wikidata.org/wiki/Q7595718","display_name":"Stability (learning theory)","level":2,"score":0.36559998989105225},{"id":"https://openalex.org/C154945302","wikidata":"https://www.wikidata.org/wiki/Q11660","display_name":"Artificial intelligence","level":1,"score":0.362199991941452},{"id":"https://openalex.org/C2781311116","wikidata":"https://www.wikidata.org/wiki/Q83306","display_name":"Group (periodic table)","level":2,"score":0.3488999903202057},{"id":"https://openalex.org/C17098449","wikidata":"https://www.wikidata.org/wiki/Q176814","display_name":"Partially observable Markov decision process","level":4,"score":0.31209999322891235},{"id":"https://openalex.org/C80444323","wikidata":"https://www.wikidata.org/wiki/Q2878974","display_name":"Theoretical computer science","level":1,"score":0.2906000018119812},{"id":"https://openalex.org/C119857082","wikidata":"https://www.wikidata.org/wiki/Q2539","display_name":"Machine learning","level":1,"score":0.2897999882698059},{"id":"https://openalex.org/C57869625","wikidata":"https://www.wikidata.org/wiki/Q1783502","display_name":"Rate of convergence","level":3,"score":0.28690001368522644},{"id":"https://openalex.org/C33923547","wikidata":"https://www.wikidata.org/wiki/Q395","display_name":"Mathematics","level":0,"score":0.2838999927043915},{"id":"https://openalex.org/C2776502983","wikidata":"https://www.wikidata.org/wiki/Q690182","display_name":"Contrast (vision)","level":2,"score":0.2782999873161316},{"id":"https://openalex.org/C126255220","wikidata":"https://www.wikidata.org/wiki/Q141495","display_name":"Mathematical optimization","level":1,"score":0.26649999618530273},{"id":"https://openalex.org/C26517878","wikidata":"https://www.wikidata.org/wiki/Q228039","display_name":"Key (lock)","level":2,"score":0.25029999017715454}],"mesh":[],"locations_count":1,"locations":[{"id":"doi:10.48550/arxiv.2604.19857","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2604.19857","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,"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":"Preprint"}],"best_oa_location":{"id":"doi:10.48550/arxiv.2604.19857","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2604.19857","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,"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"},"sustainable_development_goals":[{"score":0.7029914259910583,"id":"https://metadata.un.org/sdg/16","display_name":"Peace, Justice and strong institutions"}],"awards":[],"funders":[],"has_content":{"pdf":false,"grobid_xml":false},"content_urls":null,"referenced_works_count":0,"referenced_works":[],"related_works":[],"abstract_inverted_index":{"Reinforcement":[0,36],"fine-tuning":[1],"with":[2,18,114,144],"verifiable":[3,62,133],"rewards":[4,63,134],"(RLVR)":[5],"has":[6],"emerged":[7],"as":[8,22],"a":[9,84,106,137,160,177,190],"powerful":[10],"paradigm":[11,44],"for":[12,194],"equipping":[13],"large":[14],"vision-language":[15],"models":[16,110],"(LVLMs)":[17],"agentic":[19,112],"capabilities":[20],"such":[21],"tool":[23,68,116],"use":[24],"and":[25,79,153,173],"multi-step":[26],"reasoning.":[27],"Despite":[28],"striking":[29],"empirical":[30],"successes,":[31],"most":[32],"notably":[33],"Visual":[34],"Agentic":[35],"Fine-Tuning":[37],"(Visual-ARFT),":[38],"the":[39,58,71,100,148,166,199],"theoretical":[40],"underpinnings":[41],"of":[42,61,73,87,150,180],"this":[43,119],"remain":[45],"poorly":[46],"understood.":[47],"In":[48],"particular,":[49],"two":[50],"critical":[51],"questions":[52],"lack":[53],"rigorous":[54],"answers:":[55],"(i)~how":[56],"does":[57,81],"composite":[59,132],"structure":[60],"(format":[64],"compliance,":[65],"answer":[66],"accuracy,":[67],"executability)":[69],"affect":[70],"convergence":[72],"Group":[74],"Relative":[75],"Policy":[76],"Optimization":[77],"(GRPO),":[78],"(ii)~why":[80],"training":[82],"on":[83,147],"small":[85],"set":[86],"tool-augmented":[88,195],"tasks":[89],"transfer":[90,202],"to":[91,136],"out-of-distribution":[92,201],"domains?":[93],"We":[94],"address":[95],"these":[96],"gaps":[97],"by":[98],"introducing":[99],"\\emph{Tool-Augmented":[101],"Markov":[102],"Decision":[103],"Process}":[104],"(TA-MDP),":[105],"formal":[107],"framework":[108],"that":[109,129,164,197],"multimodal":[111],"decision-making":[113],"bounded-depth":[115],"calls.":[117],"Within":[118],"framework,":[120],"we":[121,127,158,188],"establish":[122,189],"three":[123],"main":[124],"results.":[125],"First,":[126],"prove":[128],"GRPO":[130],"under":[131],"converges":[135],"first-order":[138],"stationary":[139],"point":[140],"at":[141],"rate":[142],"$O(1/\\sqrt{T})$":[143],"explicit":[145],"dependence":[146],"number":[149],"reward":[151,182],"components":[152],"group":[154],"size":[155],"(\\textbf{Theorem~1}).":[156],"Second,":[157],"derive":[159],"\\emph{Reward":[161],"Decomposition":[162],"Theorem}":[163],"bounds":[165],"sub-optimality":[167],"gap":[168],"between":[169],"decomposed":[170],"per-component":[171],"optimization":[172],"joint":[174],"optimization,":[175],"providing":[176],"precise":[178],"characterization":[179],"when":[181],"decomposition":[183],"is":[184],"beneficial":[185],"(\\textbf{Theorem~2}).":[186],"Third,":[187],"PAC-Bayes":[191],"generalization":[192],"bound":[193],"policies":[196],"explains":[198],"strong":[200],"observed":[203],"in":[204],"Visual-ARFT":[205],"(\\textbf{Theorem~3}).":[206]},"counts_by_year":[],"updated_date":"2026-07-28T07:46:37.118299","created_date":"2026-04-24T00:00:00"}