{"id":"https://openalex.org/W4412945699","doi":"https://doi.org/10.18653/v1/2025.acl-long.305","title":"Disentangling Biased Knowledge from Reasoning in Large Language Models via Machine Unlearning","display_name":"Disentangling Biased Knowledge from Reasoning in Large Language Models via Machine Unlearning","publication_year":2025,"publication_date":"2025-01-01","ids":{"openalex":"https://openalex.org/W4412945699","doi":"https://doi.org/10.18653/v1/2025.acl-long.305"},"language":"en","primary_location":{"id":"doi:10.18653/v1/2025.acl-long.305","doi":"https://doi.org/10.18653/v1/2025.acl-long.305","is_oa":true,"landing_page_url":"https://doi.org/10.18653/v1/2025.acl-long.305","pdf_url":"https://aclanthology.org/2025.acl-long.305.pdf","source":{"id":"https://openalex.org/S4393915740","display_name":"Proceedings of the conference - Association for Computational Linguistics. Meeting","issn_l":"0736-587X","issn":["0736-587X"],"is_oa":false,"is_in_doaj":false,"is_core":false,"listed_in":["norway-1"],"host_organization":null,"host_organization_name":null,"host_organization_lineage":[],"host_organization_lineage_names":[],"type":"conference"},"license":null,"license_id":null,"version":"publishedVersion","is_accepted":true,"is_published":true,"raw_source_name":"Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)","raw_type":"proceedings-article"},"volume":null,"type":"conference-paper","indexed_in":["crossref"],"open_access":{"is_oa":true,"oa_status":"bronze","oa_url":"https://aclanthology.org/2025.acl-long.305.pdf","any_repository_has_fulltext":false},"authorships":[{"author_position":"first","author":{"id":"https://openalex.org/A5102018211","display_name":"Zheyuan Liu","orcid":"https://orcid.org/0009-0007-5458-6712","observed_orcids":["https://orcid.org/0009-0007-5458-6712"]},"institutions":[{"id":"https://openalex.org/I107639228","display_name":"University of Notre Dame","ror":"https://ror.org/00mkhxb43","country_code":"US","type":"education","lineage":["https://openalex.org/I107639228"]}],"countries":["US"],"is_corresponding":false,"raw_author_name":"Zheyuan Liu","raw_affiliation_strings":["University of Notre Dame , 2 Amazon"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"University of Notre Dame , 2 Amazon","institution_ids":["https://openalex.org/I107639228"]}]},{"author_position":"middle","author":{"id":"https://openalex.org/A5032620055","display_name":"Suraj Maharjan","orcid":"https://orcid.org/0009-0007-4963-9170","observed_orcids":["https://orcid.org/0009-0007-4963-9170","https://orcid.org/0009-0007-9391-5654"]},"institutions":[{"id":"https://openalex.org/I107639228","display_name":"University of Notre Dame","ror":"https://ror.org/00mkhxb43","country_code":"US","type":"education","lineage":["https://openalex.org/I107639228"]}],"countries":["US"],"is_corresponding":false,"raw_author_name":"Suraj Maharjan","raw_affiliation_strings":["University of Notre Dame , 2 Amazon"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"University of Notre Dame , 2 Amazon","institution_ids":["https://openalex.org/I107639228"]}]},{"author_position":"middle","author":{"id":"https://openalex.org/A5082020546","display_name":"Fanyou Wu","orcid":"https://orcid.org/0000-0002-4894-5738","observed_orcids":["https://orcid.org/0000-0002-4894-5738"]},"institutions":[{"id":"https://openalex.org/I107639228","display_name":"University of Notre Dame","ror":"https://ror.org/00mkhxb43","country_code":"US","type":"education","lineage":["https://openalex.org/I107639228"]}],"countries":["US"],"is_corresponding":false,"raw_author_name":"Fanyou Wu","raw_affiliation_strings":["University of Notre Dame , 2 Amazon"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"University of Notre Dame , 2 Amazon","institution_ids":["https://openalex.org/I107639228"]}]},{"author_position":"middle","author":{"id":"https://openalex.org/A5026929017","display_name":"Rahil Parikh","orcid":"https://orcid.org/0000-0002-2710-3742","observed_orcids":["https://orcid.org/0000-0002-2710-3742"]},"institutions":[{"id":"https://openalex.org/I107639228","display_name":"University of Notre Dame","ror":"https://ror.org/00mkhxb43","country_code":"US","type":"education","lineage":["https://openalex.org/I107639228"]}],"countries":["US"],"is_corresponding":false,"raw_author_name":"Rahil Parikh","raw_affiliation_strings":["University of Notre Dame , 2 Amazon"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"University of Notre Dame , 2 Amazon","institution_ids":["https://openalex.org/I107639228"]}]},{"author_position":"middle","author":{"id":"https://openalex.org/A5018651448","display_name":"Belhassen Bayar","orcid":"https://orcid.org/0000-0003-1382-2153","observed_orcids":["https://orcid.org/0000-0003-1382-2153"]},"institutions":[{"id":"https://openalex.org/I107639228","display_name":"University of Notre Dame","ror":"https://ror.org/00mkhxb43","country_code":"US","type":"education","lineage":["https://openalex.org/I107639228"]}],"countries":["US"],"is_corresponding":false,"raw_author_name":"Belhassen Bayar","raw_affiliation_strings":["University of Notre Dame , 2 Amazon"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"University of Notre Dame , 2 Amazon","institution_ids":["https://openalex.org/I107639228"]}]},{"author_position":"middle","author":{"id":"https://openalex.org/A5038858575","display_name":"Srinivasan H. Sengamedu","orcid":"https://orcid.org/0000-0003-1847-8398","observed_orcids":["https://orcid.org/0000-0003-1847-8398"]},"institutions":[{"id":"https://openalex.org/I107639228","display_name":"University of Notre Dame","ror":"https://ror.org/00mkhxb43","country_code":"US","type":"education","lineage":["https://openalex.org/I107639228"]}],"countries":["US"],"is_corresponding":false,"raw_author_name":"Srinivasan H. Sengamedu","raw_affiliation_strings":["University of Notre Dame , 2 Amazon"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"University of Notre Dame , 2 Amazon","institution_ids":["https://openalex.org/I107639228"]}]},{"author_position":"last","author":{"id":"https://openalex.org/A5074821819","display_name":"Meng Jiang","orcid":"https://orcid.org/0000-0002-3009-519X","observed_orcids":["https://orcid.org/0000-0002-3009-519X"]},"institutions":[{"id":"https://openalex.org/I107639228","display_name":"University of Notre Dame","ror":"https://ror.org/00mkhxb43","country_code":"US","type":"education","lineage":["https://openalex.org/I107639228"]}],"countries":["US"],"is_corresponding":false,"raw_author_name":"Meng Jiang","raw_affiliation_strings":["University of Notre Dame , 2 Amazon"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"University of Notre Dame , 2 Amazon","institution_ids":["https://openalex.org/I107639228"]}]}],"institutions":[],"countries_distinct_count":1,"institutions_distinct_count":1,"corresponding_author_ids":[],"corresponding_institution_ids":["https://openalex.org/I107639228"],"apc_list":null,"apc_paid":null,"fwci":2.0546,"has_fulltext":true,"cited_by_count":2,"citation_normalized_percentile":{"value":0.88664914,"is_in_top_1_percent":false,"is_in_top_10_percent":false},"cited_by_percentile_year":{"min":96,"max":97},"biblio":{"volume":null,"issue":null,"first_page":"6105","last_page":"6123"},"is_retracted":false,"is_paratext":false,"is_xpac":false,"primary_topic":{"id":"https://openalex.org/T10883","display_name":"Ethics and Social Impacts of AI","score":0.8539999723434448,"subfield":{"id":"https://openalex.org/subfields/3311","display_name":"Safety Research"},"field":{"id":"https://openalex.org/fields/33","display_name":"Social Sciences"},"domain":{"id":"https://openalex.org/domains/2","display_name":"Social Sciences"}},"topics":[{"id":"https://openalex.org/T10883","display_name":"Ethics and Social Impacts of AI","score":0.8539999723434448,"subfield":{"id":"https://openalex.org/subfields/3311","display_name":"Safety Research"},"field":{"id":"https://openalex.org/fields/33","display_name":"Social Sciences"},"domain":{"id":"https://openalex.org/domains/2","display_name":"Social Sciences"}},{"id":"https://openalex.org/T10764","display_name":"Privacy-Preserving Technologies in Data","score":0.07689999788999557,"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/T10028","display_name":"Topic Modeling","score":0.010099999606609344,"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/machine-unlearning","display_name":"machine unlearning","score":0.8730000257492065},{"id":"https://openalex.org/keywords/large-language-models","display_name":"large language models","score":0.9490000009536743},{"id":"https://openalex.org/keywords/language-model-bias","display_name":"bias in language models","score":0.48399999737739563},{"id":"https://openalex.org/keywords/bias-removal","display_name":"bias removal","score":0.4129999876022339},{"id":"https://openalex.org/keywords/reasoning","display_name":"reasoning","score":0.30399999022483826}],"concepts":[{"id":"https://openalex.org/C41008148","wikidata":"https://www.wikidata.org/wiki/Q21198","display_name":"Computer science","level":0,"score":0.6844103336334229},{"id":"https://openalex.org/C154945302","wikidata":"https://www.wikidata.org/wiki/Q11660","display_name":"Artificial intelligence","level":1,"score":0.4261760711669922},{"id":"https://openalex.org/C204321447","wikidata":"https://www.wikidata.org/wiki/Q30642","display_name":"Natural language processing","level":1,"score":0.42452311515808105},{"id":"https://openalex.org/C188147891","wikidata":"https://www.wikidata.org/wiki/Q147638","display_name":"Cognitive science","level":1,"score":0.4202449917793274},{"id":"https://openalex.org/C15744967","wikidata":"https://www.wikidata.org/wiki/Q9418","display_name":"Psychology","level":0,"score":0.1946728229522705}],"mesh":[],"locations_count":1,"locations":[{"id":"doi:10.18653/v1/2025.acl-long.305","doi":"https://doi.org/10.18653/v1/2025.acl-long.305","is_oa":true,"landing_page_url":"https://doi.org/10.18653/v1/2025.acl-long.305","pdf_url":"https://aclanthology.org/2025.acl-long.305.pdf","source":{"id":"https://openalex.org/S4393915740","display_name":"Proceedings of the conference - Association for Computational Linguistics. Meeting","issn_l":"0736-587X","issn":["0736-587X"],"is_oa":false,"is_in_doaj":false,"is_core":false,"listed_in":["norway-1"],"host_organization":null,"host_organization_name":null,"host_organization_lineage":[],"host_organization_lineage_names":[],"type":"conference"},"license":null,"license_id":null,"version":"publishedVersion","is_accepted":true,"is_published":true,"raw_source_name":"Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)","raw_type":"proceedings-article"}],"best_oa_location":{"id":"doi:10.18653/v1/2025.acl-long.305","doi":"https://doi.org/10.18653/v1/2025.acl-long.305","is_oa":true,"landing_page_url":"https://doi.org/10.18653/v1/2025.acl-long.305","pdf_url":"https://aclanthology.org/2025.acl-long.305.pdf","source":{"id":"https://openalex.org/S4393915740","display_name":"Proceedings of the conference - Association for Computational Linguistics. Meeting","issn_l":"0736-587X","issn":["0736-587X"],"is_oa":false,"is_in_doaj":false,"is_core":false,"listed_in":["norway-1"],"host_organization":null,"host_organization_name":null,"host_organization_lineage":[],"host_organization_lineage_names":[],"type":"conference"},"license":null,"license_id":null,"version":"publishedVersion","is_accepted":true,"is_published":true,"raw_source_name":"Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)","raw_type":"proceedings-article"},"sustainable_development_goals":[],"study_designs":[],"awards":[],"funders":[],"has_content":{"grobid_xml":true,"pdf":true},"content_urls":{"pdf":"https://content.openalex.org/works/W4412945699.pdf","grobid_xml":"https://content.openalex.org/works/W4412945699.grobid-xml"},"referenced_works_count":0,"referenced_works":[],"related_works":["https://openalex.org/W4412887980","https://openalex.org/W7196965540","https://openalex.org/W7196936216","https://openalex.org/W7213495915","https://openalex.org/W4402684036","https://openalex.org/W7166814103","https://openalex.org/W7196994526","https://openalex.org/W4411119535","https://openalex.org/W4415796441","https://openalex.org/W4411120115"],"abstract_inverted_index":{"The":[0],"rapid":[1],"development":[2],"of":[3,54],"Large":[4],"Language":[5],"Models":[6],"(LLMs)":[7],"has":[8],"led":[9],"to":[10,40,50,67,166],"their":[11],"widespread":[12],"adoption":[13],"across":[14,148],"various":[15],"domains,":[16],"leveraging":[17],"vast":[18],"pre-training":[19],"knowledge":[20,44,57,118],"and":[21,81,138,159],"impressive":[22],"generalization":[23],"capabilities.However,":[24],"these":[25,84],"models":[26],"often":[27],"inherit":[28],"biased":[29,43,117,128],"knowledge,":[30],"resulting":[31],"in":[32,35,95,124],"unfair":[33],"decisions":[34],"sensitive":[36],"applications.It":[37],"is":[38],"challenging":[39],"remove":[41],"this":[42,61,102],"without":[45],"compromising":[46],"reasoning":[47,121,161],"abilities":[48],"due":[49],"the":[51,55,69],"entangled":[52],"nature":[53],"learned":[56],"within":[58],"LLMs.To":[59],"solve":[60],"problem,":[62],"existing":[63,167],"approaches":[64],"have":[65,86],"attempted":[66],"mitigate":[68],"bias":[70],"using":[71,130],"techniques":[72],"such":[73],"as":[74],"finetuning":[75],"with":[76,135],"unbiased":[77,136],"datasets,":[78],"model":[79],"merging,":[80],"gradient":[82],"ascent.While":[83],"methods":[85],"experimentally":[87],"proven":[88],"effective,":[89],"they":[90],"can":[91],"still":[92],"be":[93],"sub-optimum":[94],"fully":[96],"disentangling":[97],"biases":[98],"from":[99],"reasoning.To":[100],"address":[101],"gap,":[103],"we":[104],"propose":[105],"Selective":[106],"Disentanglement":[107],"Unlearning":[108],"(SDU),":[109],"a":[110,131],"novel":[111],"unlearning":[112],"framework":[113],"that":[114,152],"selectively":[115],"removes":[116],"while":[119],"preserving":[120],"capabilities.SDU":[122],"operates":[123],"three":[125],"stages:":[126],"identifying":[127],"parameters":[129],"shadow":[132],"LLM,":[133],"fine-tuning":[134],"data,":[137],"performing":[139],"selective":[140],"parameter":[141],"updates":[142],"based":[143],"on":[144],"weight":[145],"saliency.Experimental":[146],"results":[147],"multiple":[149],"LLMs":[150],"show":[151],"SDU":[153],"improves":[154],"fairness":[155],"accuracy":[156],"by":[157,163],"14.7%":[158],"enhances":[160],"performance":[162],"62.6%":[164],"compared":[165],"baselines.2":[168]},"counts_by_year":[{"year":2026,"cited_by_count":2}],"updated_date":"2026-10-29T09:56:32.521200","created_date":"2025-10-10T00:00:00"}