{"id":"https://openalex.org/W7162493339","doi":"https://doi.org/10.48550/arxiv.2605.26606","title":"Spend Your Rollouts Where It Counts: Rollout Allocation for Group-Based RL Post-Training","display_name":"Spend Your Rollouts Where It Counts: Rollout Allocation for Group-Based RL Post-Training","publication_year":2026,"publication_date":"2026-05-26","ids":{"openalex":"https://openalex.org/W7162493339","doi":"https://doi.org/10.48550/arxiv.2605.26606"},"language":null,"primary_location":{"id":"doi:10.48550/arxiv.2605.26606","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2605.26606","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":"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.2605.26606","any_repository_has_fulltext":true},"authorships":[{"author_position":"first","author":{"id":"https://openalex.org/A5103046203","display_name":"Woojeong Kim","orcid":"https://orcid.org/0009-0001-3005-8014"},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Kim, Woojeong","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5137136038","display_name":"Ziyi Yang","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Yang, Ziyi","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5113817523","display_name":"Jing Nathan Yan","orcid":"https://orcid.org/0009-0004-0157-5065"},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Yan, Jing Nathan","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"last","author":{"id":"https://openalex.org/A5137102556","display_name":"Jialu Liu","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Liu, Jialu","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/T10028","display_name":"Topic Modeling","score":0.23600000143051147,"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":0.23600000143051147,"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/T10181","display_name":"Natural Language Processing Techniques","score":0.1306000053882599,"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/T11636","display_name":"Artificial Intelligence in Healthcare and Education","score":0.06650000065565109,"subfield":{"id":"https://openalex.org/subfields/2718","display_name":"Health Informatics"},"field":{"id":"https://openalex.org/fields/27","display_name":"Medicine"},"domain":{"id":"https://openalex.org/domains/4","display_name":"Health Sciences"}}],"keywords":[{"id":"https://openalex.org/keywords/reinforcement-learning","display_name":"Reinforcement learning","score":0.5906000137329102},{"id":"https://openalex.org/keywords/fraction","display_name":"Fraction (chemistry)","score":0.5566999912261963},{"id":"https://openalex.org/keywords/baseline","display_name":"Baseline (sea)","score":0.5195000171661377},{"id":"https://openalex.org/keywords/budget-constraint","display_name":"Budget constraint","score":0.42100000381469727},{"id":"https://openalex.org/keywords/importance-sampling","display_name":"Importance sampling","score":0.4138999879360199},{"id":"https://openalex.org/keywords/sampling","display_name":"Sampling (signal processing)","score":0.38179999589920044},{"id":"https://openalex.org/keywords/function","display_name":"Function (biology)","score":0.32269999384880066}],"concepts":[{"id":"https://openalex.org/C41008148","wikidata":"https://www.wikidata.org/wiki/Q21198","display_name":"Computer science","level":0,"score":0.7594000101089478},{"id":"https://openalex.org/C97541855","wikidata":"https://www.wikidata.org/wiki/Q830687","display_name":"Reinforcement learning","level":2,"score":0.5906000137329102},{"id":"https://openalex.org/C149629883","wikidata":"https://www.wikidata.org/wiki/Q660926","display_name":"Fraction (chemistry)","level":2,"score":0.5566999912261963},{"id":"https://openalex.org/C12725497","wikidata":"https://www.wikidata.org/wiki/Q810247","display_name":"Baseline (sea)","level":2,"score":0.5195000171661377},{"id":"https://openalex.org/C8505890","wikidata":"https://www.wikidata.org/wiki/Q605095","display_name":"Budget constraint","level":2,"score":0.42100000381469727},{"id":"https://openalex.org/C52740198","wikidata":"https://www.wikidata.org/wiki/Q1539564","display_name":"Importance sampling","level":3,"score":0.4138999879360199},{"id":"https://openalex.org/C140779682","wikidata":"https://www.wikidata.org/wiki/Q210868","display_name":"Sampling (signal processing)","level":3,"score":0.38179999589920044},{"id":"https://openalex.org/C14036430","wikidata":"https://www.wikidata.org/wiki/Q3736076","display_name":"Function (biology)","level":2,"score":0.32269999384880066},{"id":"https://openalex.org/C73602740","wikidata":"https://www.wikidata.org/wiki/Q7795822","display_name":"Thompson sampling","level":3,"score":0.3165000081062317},{"id":"https://openalex.org/C77618280","wikidata":"https://www.wikidata.org/wiki/Q1155772","display_name":"Scheme (mathematics)","level":2,"score":0.2890999913215637},{"id":"https://openalex.org/C126255220","wikidata":"https://www.wikidata.org/wiki/Q141495","display_name":"Mathematical optimization","level":1,"score":0.28110000491142273},{"id":"https://openalex.org/C206729178","wikidata":"https://www.wikidata.org/wiki/Q2271896","display_name":"Scheduling (production processes)","level":2,"score":0.2667999863624573},{"id":"https://openalex.org/C112758219","wikidata":"https://www.wikidata.org/wiki/Q16038819","display_name":"Duration (music)","level":2,"score":0.2653000056743622},{"id":"https://openalex.org/C2986087404","wikidata":"https://www.wikidata.org/wiki/Q15946010","display_name":"Online learning","level":2,"score":0.26499998569488525},{"id":"https://openalex.org/C17923572","wikidata":"https://www.wikidata.org/wiki/Q7250160","display_name":"Propensity score matching","level":2,"score":0.2646999955177307},{"id":"https://openalex.org/C137836250","wikidata":"https://www.wikidata.org/wiki/Q984063","display_name":"Optimization problem","level":2,"score":0.2572000026702881}],"mesh":[],"locations_count":1,"locations":[{"id":"doi:10.48550/arxiv.2605.26606","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2605.26606","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":"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.2605.26606","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2605.26606","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":"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":[{"id":"https://metadata.un.org/sdg/10","score":0.5596667528152466,"display_name":"Reduced inequalities"}],"awards":[],"funders":[],"has_content":{"grobid_xml":false,"pdf":false},"content_urls":null,"referenced_works_count":0,"referenced_works":[],"related_works":[],"abstract_inverted_index":{"Reinforcement":[0],"learning":[1,53],"(RL)":[2],"is":[3,89],"the":[4,14,21,70,120,123],"dominant":[5],"paradigm":[6],"for":[7,100],"post-training":[8],"large":[9],"language":[10],"models.":[11],"However,":[12],"in":[13,63,172],"online,":[15],"on-policy":[16],"setting,":[17],"rollout":[18,97],"generation":[19],"dominates":[20],"computational":[22],"cost":[23],"of":[24,65,119],"training.":[25],"Group-based":[26],"policy":[27,71],"optimization":[28],"methods":[29],"compute":[30],"advantages":[31],"from":[32,108,144],"multiple":[33,137],"rollouts":[34,50,125],"per":[35],"prompt,":[36],"yet":[37],"they":[38],"indiscriminately":[39],"allocate":[40],"budget":[41],"to":[42,128,146,162],"prompts":[43,130,133],"with":[44,153],"collapsed":[45],"reward":[46,67],"distributions,":[47],"wasting":[48],"expensive":[49],"on":[51],"negligible":[52],"signals.":[54],"We":[55,92],"demonstrate":[56],"that":[57],"group-based":[58,101],"updates":[59],"are":[60,126,134],"most":[61],"effective":[62],"regimes":[64],"high":[66],"variance.":[68],"Since":[69],"evolves":[72],"throughout":[73],"training,":[74],"prompt":[75,88,106],"informativeness":[76,115],"must":[77],"be":[78],"estimated":[79],"online":[80],"rather":[81],"than":[82,165,170],"precomputed,":[83],"but":[84],"exhaustively":[85],"evaluating":[86],"every":[87],"computationally":[90],"prohibitive.":[91],"introduce":[93],"Pilot-Commit,":[94],"a":[95,110,117],"budget-aware":[96],"allocation":[98],"framework":[99],"RL":[102],"post-training.":[103],"Pilot-Commit":[104,149],"decouples":[105],"evaluation":[107],"exploitation:":[109],"pilot":[111],"stage":[112],"estimates":[113],"per-prompt":[114],"using":[116],"fraction":[118],"budget,":[121],"and":[122,141,167],"remaining":[124],"allocated":[127],"high-leverage":[129],"while":[131],"low-signal":[132],"skipped.":[135],"Across":[136],"math":[138],"reasoning":[139],"benchmarks":[140],"model":[142],"scales":[143],"1.5B":[145],"14B":[147],"parameters,":[148],"matches":[150],"baseline":[151],"accuracy":[152,160],"significantly":[154],"lower":[155],"sampling":[156],"costs,":[157],"reaching":[158],"target":[159],"up":[161],"$1.9\\times$":[163],"faster":[164,169],"GRPO":[166],"$4.0\\times$":[168],"DAPO":[171],"cumulative":[173],"rollouts.":[174]},"counts_by_year":[],"updated_date":"2026-07-28T07:46:37.118299","created_date":"2026-05-28T00:00:00"}