{"id":"https://openalex.org/W7163514553","doi":"https://doi.org/10.48550/arxiv.2606.04272","title":"RL Excursions during Pre-Training: Re-examining Policy Optimization for LLM training","display_name":"RL Excursions during Pre-Training: Re-examining Policy Optimization for LLM training","publication_year":2026,"publication_date":"2026-06-02","ids":{"openalex":"https://openalex.org/W7163514553","doi":"https://doi.org/10.48550/arxiv.2606.04272"},"language":null,"primary_location":{"id":"doi:10.48550/arxiv.2606.04272","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2606.04272","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.2606.04272","any_repository_has_fulltext":true},"authorships":[{"author_position":"first","author":{"id":"https://openalex.org/A5039871170","display_name":"Rachit Bansal","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Bansal, Rachit","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5137809498","display_name":"Clara Mohri","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Mohri, Clara","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5137809560","display_name":"Tian Qin","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Qin, Tian","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5137816834","display_name":"David Alvarez-Melis","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Alvarez-Melis, David","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"last","author":{"id":"https://openalex.org/A5133908067","display_name":"Sham Kakade","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Kakade, Sham","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":0.8562999963760376,"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":0.8562999963760376,"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":0.014499999582767487,"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/T11689","display_name":"Adversarial Robustness in Machine Learning","score":0.013700000010430813,"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":0.7023000121116638},{"id":"https://openalex.org/keywords/merge","display_name":"Merge (version control)","score":0.6345000267028809},{"id":"https://openalex.org/keywords/sharpening","display_name":"Sharpening","score":0.5645999908447266},{"id":"https://openalex.org/keywords/pipeline","display_name":"Pipeline (software)","score":0.5462999939918518},{"id":"https://openalex.org/keywords/lever","display_name":"Lever","score":0.5110999941825867},{"id":"https://openalex.org/keywords/training","display_name":"Training (meteorology)","score":0.4740000069141388},{"id":"https://openalex.org/keywords/training-set","display_name":"Training set","score":0.4332999885082245}],"concepts":[{"id":"https://openalex.org/C97541855","wikidata":"https://www.wikidata.org/wiki/Q830687","display_name":"Reinforcement learning","level":2,"score":0.7023000121116638},{"id":"https://openalex.org/C41008148","wikidata":"https://www.wikidata.org/wiki/Q21198","display_name":"Computer science","level":0,"score":0.6661999821662903},{"id":"https://openalex.org/C197129107","wikidata":"https://www.wikidata.org/wiki/Q1921621","display_name":"Merge (version control)","level":2,"score":0.6345000267028809},{"id":"https://openalex.org/C2781137444","wikidata":"https://www.wikidata.org/wiki/Q237105","display_name":"Sharpening","level":2,"score":0.5645999908447266},{"id":"https://openalex.org/C43521106","wikidata":"https://www.wikidata.org/wiki/Q2165493","display_name":"Pipeline (software)","level":2,"score":0.5462999939918518},{"id":"https://openalex.org/C107524782","wikidata":"https://www.wikidata.org/wiki/Q40164","display_name":"Lever","level":2,"score":0.5110999941825867},{"id":"https://openalex.org/C2777211547","wikidata":"https://www.wikidata.org/wiki/Q17141490","display_name":"Training (meteorology)","level":2,"score":0.4740000069141388},{"id":"https://openalex.org/C51632099","wikidata":"https://www.wikidata.org/wiki/Q3985153","display_name":"Training set","level":2,"score":0.4332999885082245},{"id":"https://openalex.org/C154945302","wikidata":"https://www.wikidata.org/wiki/Q11660","display_name":"Artificial intelligence","level":1,"score":0.4323999881744385},{"id":"https://openalex.org/C2776748549","wikidata":"https://www.wikidata.org/wiki/Q201610","display_name":"Status quo","level":2,"score":0.38040000200271606},{"id":"https://openalex.org/C119857082","wikidata":"https://www.wikidata.org/wiki/Q2539","display_name":"Machine learning","level":1,"score":0.3662000000476837},{"id":"https://openalex.org/C2781235140","wikidata":"https://www.wikidata.org/wiki/Q275131","display_name":"Scratch","level":2,"score":0.3537999987602234},{"id":"https://openalex.org/C206588197","wikidata":"https://www.wikidata.org/wiki/Q846574","display_name":"Reuse","level":2,"score":0.29170000553131104},{"id":"https://openalex.org/C153083717","wikidata":"https://www.wikidata.org/wiki/Q6535263","display_name":"Leverage (statistics)","level":2,"score":0.28139999508857727},{"id":"https://openalex.org/C42058472","wikidata":"https://www.wikidata.org/wiki/Q810214","display_name":"Base (topology)","level":2,"score":0.26080000400543213},{"id":"https://openalex.org/C4554734","wikidata":"https://www.wikidata.org/wiki/Q593744","display_name":"Knowledge base","level":2,"score":0.25040000677108765}],"mesh":[],"locations_count":1,"locations":[{"id":"doi:10.48550/arxiv.2606.04272","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2606.04272","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.2606.04272","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2606.04272","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":[],"awards":[],"funders":[],"has_content":{"grobid_xml":false,"pdf":false},"content_urls":null,"referenced_works_count":0,"referenced_works":[],"related_works":[],"abstract_inverted_index":{"The":[0,110],"standard":[1],"LLM":[2,24,155],"training":[3,22,141,156],"pipeline":[4,55],"applies":[5],"reinforcement":[6],"learning":[7],"(RL)":[8],"only":[9,105],"after":[10],"pre-training":[11,39,68],"and":[12,27,31,49,130],"supervised":[13],"fine-tuning":[14],"(SFT).":[15],"We":[16,41],"question":[17],"this":[18],"status":[19],"quo":[20],"by":[21,34,119,133],"a":[23,72],"from":[25,159],"scratch":[26],"applying":[28,87],"RL,":[29,120],"SFT,":[30],"SFT":[32,131],"followed":[33],"RL":[35,44,76,88,107,129],"directly":[36,89],"to":[37,90],"intermediate":[38],"checkpoints.":[40],"find":[42,65],"that":[43,66,154],"is":[45,71],"effective":[46],"very":[47],"early,":[48],"often":[50],"matches":[51],"the":[52,94,97,114],"full":[53],"SFT$\\to$RL":[54],"early":[56],"as":[57],"well.":[58],"Through":[59],"experiments":[60],"on":[61],"harder":[62],"problems,":[63],"we":[64,127],"targeted":[67],"data":[69],"composition":[70],"strong":[73],"lever":[74],"for":[75],"effectiveness,":[77],"even":[78],"more":[79],"so":[80],"than":[81],"model":[82,115],"scale.":[83],"Beyond":[84],"reasoning":[85],"accuracy,":[86],"base":[91],"checkpoints":[92],"expands":[93],"model's":[95],"distribution;":[96],"sharpening":[98],"effect":[99],"reported":[100],"in":[101],"recent":[102],"work":[103],"arises":[104],"when":[106],"follows":[108],"SFT.":[109,125],"general":[111,148],"capabilities":[112],"of":[113,163],"remain":[116],"essentially":[117],"unchanged":[118],"while":[121,146],"they":[122],"degrade":[123],"following":[124],"Finally,":[126],"merge":[128],"objectives":[132],"parallel":[134],"averaging,":[135],"which":[136],"outperforms":[137],"across":[138,144],"all":[139],"other":[140],"methods":[142],"discussed,":[143],"metrics,":[145],"preserving":[147],"capabilities.":[149],"Together,":[150],"these":[151],"results":[152],"suggest":[153],"might":[157],"benefit":[158],"an":[160],"expanded":[161],"use":[162],"RL.":[164]},"counts_by_year":[],"updated_date":"2026-07-01T06:00:48.157686","created_date":"2026-06-05T00:00:00"}