{"id":"https://openalex.org/W4416035874","doi":"https://doi.org/10.18653/v1/2025.findings-emnlp.57","title":"Unearthing Gems from Stones: Policy Optimization with Negative Sample Augmentation for LLM Reasoning","display_name":"Unearthing Gems from Stones: Policy Optimization with Negative Sample Augmentation for LLM Reasoning","publication_year":2025,"publication_date":"2025-01-01","ids":{"openalex":"https://openalex.org/W4416035874","doi":"https://doi.org/10.18653/v1/2025.findings-emnlp.57"},"language":null,"primary_location":{"id":"doi:10.18653/v1/2025.findings-emnlp.57","is_oa":true,"landing_page_url":"https://doi.org/10.18653/v1/2025.findings-emnlp.57","pdf_url":"https://aclanthology.org/2025.findings-emnlp.57.pdf","source":null,"license":"cc-by","license_id":"https://openalex.org/licenses/cc-by","version":"publishedVersion","is_accepted":true,"is_published":true,"raw_source_name":"Findings of the Association for Computational Linguistics: EMNLP 2025","raw_type":"proceedings-article"},"type":"article","indexed_in":["crossref"],"open_access":{"is_oa":true,"oa_status":"gold","oa_url":"https://aclanthology.org/2025.findings-emnlp.57.pdf","any_repository_has_fulltext":null},"authorships":[{"author_position":"first","author":{"id":"https://openalex.org/A5100299281","display_name":"Zhaohui Yang","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Zhaohui Yang","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5072799796","display_name":"Yuxiao Ye","orcid":"https://orcid.org/0000-0003-4891-7240"},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Yuxiao Ye","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5018604247","display_name":"Shilei Jiang","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Shilei Jiang","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5108126327","display_name":"Shihong Deng","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Shihong Deng","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5024273159","display_name":"Hu Chen","orcid":"https://orcid.org/0009-0003-5600-5935"},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Chen Hu","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5036100020","display_name":"Linjing Li","orcid":"https://orcid.org/0000-0002-8737-099X"},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Linjing Li","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"last","author":{"id":"https://openalex.org/A5060116992","display_name":"Daxin Jiang","orcid":"https://orcid.org/0000-0002-6657-5806"},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Daxin Jiang","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]}],"institutions":[],"countries_distinct_count":0,"institutions_distinct_count":7,"corresponding_author_ids":[],"corresponding_institution_ids":[],"apc_list":null,"apc_paid":null,"fwci":0.0,"has_fulltext":true,"cited_by_count":0,"citation_normalized_percentile":{"value":0.15845208,"is_in_top_1_percent":false,"is_in_top_10_percent":false},"cited_by_percentile_year":null,"biblio":{"volume":null,"issue":null,"first_page":"1061","last_page":"1075"},"is_retracted":false,"is_paratext":false,"is_xpac":false,"primary_topic":{"id":"https://openalex.org/T11010","display_name":"Logic, Reasoning, and Knowledge","score":0.2524999976158142,"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/T11010","display_name":"Logic, Reasoning, and Knowledge","score":0.2524999976158142,"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/T12026","display_name":"Explainable Artificial Intelligence (XAI)","score":0.09300000220537186,"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/T12072","display_name":"Machine Learning and Algorithms","score":0.04740000143647194,"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/sample","display_name":"Sample (material)","score":0.690500020980835},{"id":"https://openalex.org/keywords/case-based-reasoning","display_name":"Case-based reasoning","score":0.26440000534057617}],"concepts":[{"id":"https://openalex.org/C198531522","wikidata":"https://www.wikidata.org/wiki/Q485146","display_name":"Sample (material)","level":2,"score":0.690500020980835},{"id":"https://openalex.org/C41008148","wikidata":"https://www.wikidata.org/wiki/Q21198","display_name":"Computer science","level":0,"score":0.5385000109672546},{"id":"https://openalex.org/C154945302","wikidata":"https://www.wikidata.org/wiki/Q11660","display_name":"Artificial intelligence","level":1,"score":0.41609999537467957},{"id":"https://openalex.org/C42475967","wikidata":"https://www.wikidata.org/wiki/Q194292","display_name":"Operations research","level":1,"score":0.3122999966144562},{"id":"https://openalex.org/C149782125","wikidata":"https://www.wikidata.org/wiki/Q160039","display_name":"Econometrics","level":1,"score":0.30570000410079956},{"id":"https://openalex.org/C539667460","wikidata":"https://www.wikidata.org/wiki/Q2414942","display_name":"Management science","level":1,"score":0.29330000281333923},{"id":"https://openalex.org/C162324750","wikidata":"https://www.wikidata.org/wiki/Q8134","display_name":"Economics","level":0,"score":0.28459998965263367},{"id":"https://openalex.org/C20162079","wikidata":"https://www.wikidata.org/wiki/Q1151406","display_name":"Case-based reasoning","level":2,"score":0.26440000534057617},{"id":"https://openalex.org/C127413603","wikidata":"https://www.wikidata.org/wiki/Q11023","display_name":"Engineering","level":0,"score":0.2556999921798706},{"id":"https://openalex.org/C144237770","wikidata":"https://www.wikidata.org/wiki/Q747534","display_name":"Mathematical economics","level":1,"score":0.24719999730587006}],"mesh":[],"locations_count":1,"locations":[{"id":"doi:10.18653/v1/2025.findings-emnlp.57","is_oa":true,"landing_page_url":"https://doi.org/10.18653/v1/2025.findings-emnlp.57","pdf_url":"https://aclanthology.org/2025.findings-emnlp.57.pdf","source":null,"license":"cc-by","license_id":"https://openalex.org/licenses/cc-by","version":"publishedVersion","is_accepted":true,"is_published":true,"raw_source_name":"Findings of the Association for Computational Linguistics: EMNLP 2025","raw_type":"proceedings-article"}],"best_oa_location":{"id":"doi:10.18653/v1/2025.findings-emnlp.57","is_oa":true,"landing_page_url":"https://doi.org/10.18653/v1/2025.findings-emnlp.57","pdf_url":"https://aclanthology.org/2025.findings-emnlp.57.pdf","source":null,"license":"cc-by","license_id":"https://openalex.org/licenses/cc-by","version":"publishedVersion","is_accepted":true,"is_published":true,"raw_source_name":"Findings of the Association for Computational Linguistics: EMNLP 2025","raw_type":"proceedings-article"},"sustainable_development_goals":[],"awards":[],"funders":[{"id":"https://openalex.org/F4320321133","display_name":"Chinese Academy of Sciences","ror":"https://ror.org/034t30j35"}],"has_content":{"pdf":true,"grobid_xml":true},"content_urls":{"pdf":"https://content.openalex.org/works/W4416035874.pdf","grobid_xml":"https://content.openalex.org/works/W4416035874.grobid-xml"},"referenced_works_count":0,"referenced_works":[],"related_works":[],"abstract_inverted_index":{"Recent":[0],"advances":[1],"in":[2,23],"reasoning":[3,136],"language":[4],"models":[5],"have":[6],"witnessed":[7],"a":[8,89],"paradigm":[9],"shift":[10],"from":[11],"short":[12],"to":[13,69,118,154],"long":[14,24],"CoT":[15,25],"pattern.Given":[16],"the":[17,28,139],"substantial":[18],"computational":[19],"cost":[20],"of":[21,30,76],"rollouts":[22],"models,":[26],"maximizing":[27],"utility":[29],"fixed":[31],"training":[32,141],"datasets":[33],"becomes":[34],"crucial.Our":[35],"analysis":[36],"reveals":[37],"that":[38,94,128],"negative":[39,57,124],"responses":[40],"contain":[41],"valuable":[42],"components":[43],"such":[44],"as":[45],"self-reflection":[46],"and":[47,108,111,147,150],"errorcorrection":[48],"steps,":[49],"yet":[50],"primary":[51],"existing":[52],"methods":[53],"either":[54],"completely":[55],"discard":[56],"samples":[58],"(RFT)":[59],"or":[60],"apply":[61],"equal":[62],"penalization":[63],"across":[64],"all":[65],"tokens":[66],"(RL),":[67],"failing":[68],"leverage":[70],"these":[71],"potential":[72],"learning":[73],"signals.In":[74],"light":[75],"this,":[77],"we":[78],"propose":[79],"Behavior":[80],"Constrained":[81],"Policy":[82],"Gradient":[83],"with":[84,115],"Negative":[85],"Sample":[86],"Augmentation":[87],"(BCPG-NSA),":[88],"fine-grained":[90],"offline":[91],"RL":[92],"framework":[93],"encompasses":[95],"three":[96],"stages:":[97],"1)":[98],"sample":[99,145],"segmentation,":[100],"2)":[101],"consensus-based":[102],"step":[103],"correctness":[104],"assessment":[105],"combining":[106],"LLM":[107],"PRM":[109],"judgers,":[110],"3)":[112],"policy":[113],"optimization":[114],"NSA":[116],"designed":[117],"effectively":[119],"mine":[120],"positive":[121],"steps":[122],"within":[123],"samples.Experimental":[125],"results":[126],"show":[127],"BCPG-NSA":[129],"outperforms":[130],"baselines":[131],"on":[132],"several":[133],"challenging":[134],"math/coding":[135],"benchmarks":[137],"using":[138],"same":[140],"dataset,":[142],"achieving":[143],"improved":[144],"efficiency":[146],"demonstrating":[148],"robustness":[149],"scalability":[151],"when":[152],"extended":[153],"multiple":[155],"iterations.":[156]},"counts_by_year":[],"updated_date":"2026-06-11T09:08:48.828518","created_date":"2025-11-08T00:00:00"}