{"id":"https://openalex.org/W7148854760","doi":"https://doi.org/10.48550/arxiv.2604.01597","title":"Learning from the Right Rollouts: Data Attribution for PPO-based LLM Post-Training","display_name":"Learning from the Right Rollouts: Data Attribution for PPO-based LLM Post-Training","publication_year":2026,"publication_date":"2026-04-02","ids":{"openalex":"https://openalex.org/W7148854760","doi":"https://doi.org/10.48550/arxiv.2604.01597"},"language":null,"primary_location":{"id":"doi:10.48550/arxiv.2604.01597","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2604.01597","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.2604.01597","any_repository_has_fulltext":true},"authorships":[{"author_position":"first","author":{"id":"https://openalex.org/A5100938790","display_name":"Dong Shu","orcid":"https://orcid.org/0009-0009-9785-3934"},"institutions":[],"countries":[],"is_corresponding":true,"raw_author_name":"Shu, Dong","raw_affiliation_strings":[],"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5132900554","display_name":"Denghui Zhang","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Zhang, Denghui","raw_affiliation_strings":[],"affiliations":[]},{"author_position":"last","author":{"id":"https://openalex.org/A5132895071","display_name":"Jessica Hullman","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Hullman, Jessica","raw_affiliation_strings":[],"affiliations":[]}],"institutions":[],"countries_distinct_count":0,"institutions_distinct_count":3,"corresponding_author_ids":["https://openalex.org/A5100938790"],"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":0.5831000208854675,"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":0.5831000208854675,"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/T12101","display_name":"Advanced Bandit Algorithms Research","score":0.0681999996304512,"subfield":{"id":"https://openalex.org/subfields/1803","display_name":"Management Science and Operations Research"},"field":{"id":"https://openalex.org/fields/18","display_name":"Decision Sciences"},"domain":{"id":"https://openalex.org/domains/2","display_name":"Social Sciences"}},{"id":"https://openalex.org/T11307","display_name":"Domain Adaptation and Few-Shot Learning","score":0.033799998462200165,"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/process","display_name":"Process (computing)","score":0.6855999827384949},{"id":"https://openalex.org/keywords/attribution","display_name":"Attribution","score":0.6103000044822693},{"id":"https://openalex.org/keywords/noisy-data","display_name":"Noisy data","score":0.5088000297546387},{"id":"https://openalex.org/keywords/training-set","display_name":"Training set","score":0.48179998993873596},{"id":"https://openalex.org/keywords/training","display_name":"Training (meteorology)","score":0.42829999327659607},{"id":"https://openalex.org/keywords/data-collection","display_name":"Data collection","score":0.3125999867916107}],"concepts":[{"id":"https://openalex.org/C98045186","wikidata":"https://www.wikidata.org/wiki/Q205663","display_name":"Process (computing)","level":2,"score":0.6855999827384949},{"id":"https://openalex.org/C143299363","wikidata":"https://www.wikidata.org/wiki/Q900584","display_name":"Attribution","level":2,"score":0.6103000044822693},{"id":"https://openalex.org/C154945302","wikidata":"https://www.wikidata.org/wiki/Q11660","display_name":"Artificial intelligence","level":1,"score":0.6044999957084656},{"id":"https://openalex.org/C41008148","wikidata":"https://www.wikidata.org/wiki/Q21198","display_name":"Computer science","level":0,"score":0.5799000263214111},{"id":"https://openalex.org/C2781170535","wikidata":"https://www.wikidata.org/wiki/Q30587856","display_name":"Noisy data","level":2,"score":0.5088000297546387},{"id":"https://openalex.org/C51632099","wikidata":"https://www.wikidata.org/wiki/Q3985153","display_name":"Training set","level":2,"score":0.48179998993873596},{"id":"https://openalex.org/C119857082","wikidata":"https://www.wikidata.org/wiki/Q2539","display_name":"Machine learning","level":1,"score":0.4499000012874603},{"id":"https://openalex.org/C2777211547","wikidata":"https://www.wikidata.org/wiki/Q17141490","display_name":"Training (meteorology)","level":2,"score":0.42829999327659607},{"id":"https://openalex.org/C15744967","wikidata":"https://www.wikidata.org/wiki/Q9418","display_name":"Psychology","level":0,"score":0.35429999232292175},{"id":"https://openalex.org/C133462117","wikidata":"https://www.wikidata.org/wiki/Q4929239","display_name":"Data collection","level":2,"score":0.3125999867916107},{"id":"https://openalex.org/C5465570","wikidata":"https://www.wikidata.org/wiki/Q5326898","display_name":"Early stopping","level":3,"score":0.30709999799728394},{"id":"https://openalex.org/C112972136","wikidata":"https://www.wikidata.org/wiki/Q7595718","display_name":"Stability (learning theory)","level":2,"score":0.302700012922287},{"id":"https://openalex.org/C67186912","wikidata":"https://www.wikidata.org/wiki/Q367664","display_name":"Data modeling","level":2,"score":0.28439998626708984},{"id":"https://openalex.org/C2776145971","wikidata":"https://www.wikidata.org/wiki/Q30673951","display_name":"Labeled data","level":2,"score":0.27300000190734863},{"id":"https://openalex.org/C137836250","wikidata":"https://www.wikidata.org/wiki/Q984063","display_name":"Optimization problem","level":2,"score":0.259799987077713},{"id":"https://openalex.org/C180747234","wikidata":"https://www.wikidata.org/wiki/Q23373","display_name":"Cognitive psychology","level":1,"score":0.25519999861717224}],"mesh":[],"locations_count":1,"locations":[{"id":"doi:10.48550/arxiv.2604.01597","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2604.01597","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.2604.01597","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2604.01597","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":{"Traditional":[0],"RL":[1,63],"algorithms":[2],"like":[3],"Proximal":[4],"Policy":[5],"Optimization":[6],"(PPO)":[7],"typically":[8],"train":[9],"on":[10],"the":[11,17,62],"entire":[12],"rollout":[13],"buffer,":[14],"operating":[15],"under":[16],"assumption":[18],"that":[19,57,83,93,103],"all":[20],"generated":[21],"episodes":[22,30,82],"provide":[23],"a":[24,54,75,87],"beneficial":[25],"optimization":[26],"signal.":[27],"However,":[28],"these":[29],"frequently":[31],"contain":[32],"noisy":[33],"or":[34],"unfaithful":[35,120],"reasoning,":[36],"which":[37],"can":[38],"degrade":[39],"model":[40],"performance":[41],"and":[42,80,98],"slow":[43],"down":[44],"training.":[45],"In":[46],"this":[47],"paper,":[48],"we":[49],"propose":[50],"\\textbf{Influence-Guided":[51],"PPO":[52,99],"(I-PPO)},":[53],"novel":[55],"framework":[56],"integrates":[58],"data":[59],"attribution":[60],"into":[61],"post-training":[64],"loop.":[65],"By":[66],"calculating":[67],"an":[68,109],"influence":[69],"score":[70],"for":[71],"each":[72],"episode":[73],"using":[74],"gradient-based":[76],"approximation,":[77],"I-PPO":[78,94],"identifies":[79],"eliminates":[81],"are":[84],"anti-aligned":[85],"with":[86],"validation":[88],"gradient.":[89],"Our":[90],"experiments":[91],"demonstrate":[92],"consistently":[95],"outperforms":[96],"SFT":[97],"baselines.":[100],"We":[101],"show":[102],"our":[104],"filtering":[105],"process":[106],"acts":[107],"as":[108],"intrinsic":[110],"early":[111],"stopping":[112],"mechanism,":[113],"accelerating":[114],"training":[115],"efficiency":[116],"while":[117],"effectively":[118],"reducing":[119],"CoT":[121],"reasoning.":[122]},"counts_by_year":[],"updated_date":"2026-04-04T06:15:33.020886","created_date":"2026-04-04T00:00:00"}