{"id":"https://openalex.org/W4415933347","doi":"https://doi.org/10.18653/v1/2026.findings-acl.1893","title":"Debiasing LLMs by Masking Unfairness-Driving Attention Heads","display_name":"Debiasing LLMs by Masking Unfairness-Driving Attention Heads","publication_year":2026,"publication_date":"2026-01-01","ids":{"openalex":"https://openalex.org/W4415933347","doi":"https://doi.org/10.18653/v1/2026.findings-acl.1893"},"language":"en","primary_location":{"id":"doi:10.18653/v1/2026.findings-acl.1893","is_oa":true,"landing_page_url":"https://doi.org/10.18653/v1/2026.findings-acl.1893","pdf_url":"https://aclanthology.org/2026.findings-acl.1893.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: ACL 2026","raw_type":"proceedings-article"},"type":"conference-paper","indexed_in":["arxiv","crossref","datacite"],"open_access":{"is_oa":true,"oa_status":"gold","oa_url":"https://aclanthology.org/2026.findings-acl.1893.pdf","any_repository_has_fulltext":true},"authorships":[{"author_position":"first","author":{"id":null,"display_name":"Tingxu Han","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Tingxu Han","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5107951446","display_name":"Wei Song","orcid":"https://orcid.org/0009-0002-1879-2256"},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Wei Song","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":null,"display_name":"Ziqi Ding","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Ziqi Ding","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5040935769","display_name":"Ziming Li","orcid":"https://orcid.org/0000-0003-0964-6724"},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Ziming Li","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":null,"display_name":"Chunrong Fang","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Chunrong Fang","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5101867789","display_name":"Yuekang Li","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Yuekang Li","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5101979289","display_name":"Dongfang Liu","orcid":"https://orcid.org/0000-0001-6995-4775"},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Dongfang Liu","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5100422933","display_name":"Zhenyu Chen","orcid":"https://orcid.org/0000-0002-4989-7109"},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Zhenyu Chen","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"last","author":{"id":"https://openalex.org/A5044589781","display_name":"Zhenting Wang","orcid":"https://orcid.org/0000-0001-6527-5089"},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Zhenting Wang","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":0.0,"has_fulltext":true,"cited_by_count":0,"citation_normalized_percentile":{"value":0.00202293,"is_in_top_1_percent":false,"is_in_top_10_percent":false},"cited_by_percentile_year":null,"biblio":{"volume":null,"issue":null,"first_page":"37980","last_page":"37992"},"is_retracted":false,"is_paratext":false,"is_xpac":false,"primary_topic":{"id":"https://openalex.org/T11636","display_name":"Artificial Intelligence in Healthcare and Education","score":0.21879999339580536,"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"}},"topics":[{"id":"https://openalex.org/T11636","display_name":"Artificial Intelligence in Healthcare and Education","score":0.21879999339580536,"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"}},{"id":"https://openalex.org/T10883","display_name":"Ethics and Social Impacts of AI","score":0.1665000021457672,"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/T13702","display_name":"Machine Learning in Healthcare","score":0.1152999997138977,"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/debiasing","display_name":"Debiasing","score":0.9962000250816345},{"id":"https://openalex.org/keywords/masking","display_name":"Masking (illustration)","score":0.7480999827384949},{"id":"https://openalex.org/keywords/differential","display_name":"Differential (mechanical device)","score":0.4438000023365021},{"id":"https://openalex.org/keywords/gender-bias","display_name":"Gender bias","score":0.43059998750686646},{"id":"https://openalex.org/keywords/cognitive-bias","display_name":"Cognitive bias","score":0.421999990940094},{"id":"https://openalex.org/keywords/work","display_name":"Work (physics)","score":0.3871999979019165},{"id":"https://openalex.org/keywords/cluster","display_name":"Cluster (spacecraft)","score":0.36570000648498535}],"concepts":[{"id":"https://openalex.org/C2779458634","wikidata":"https://www.wikidata.org/wiki/Q24963715","display_name":"Debiasing","level":2,"score":0.9962000250816345},{"id":"https://openalex.org/C2777402240","wikidata":"https://www.wikidata.org/wiki/Q6783436","display_name":"Masking (illustration)","level":2,"score":0.7480999827384949},{"id":"https://openalex.org/C180747234","wikidata":"https://www.wikidata.org/wiki/Q23373","display_name":"Cognitive psychology","level":1,"score":0.5745000243186951},{"id":"https://openalex.org/C15744967","wikidata":"https://www.wikidata.org/wiki/Q9418","display_name":"Psychology","level":0,"score":0.5527999997138977},{"id":"https://openalex.org/C41008148","wikidata":"https://www.wikidata.org/wiki/Q21198","display_name":"Computer science","level":0,"score":0.4943999946117401},{"id":"https://openalex.org/C93226319","wikidata":"https://www.wikidata.org/wiki/Q193137","display_name":"Differential (mechanical device)","level":2,"score":0.4438000023365021},{"id":"https://openalex.org/C2983427547","wikidata":"https://www.wikidata.org/wiki/Q93200","display_name":"Gender bias","level":2,"score":0.43059998750686646},{"id":"https://openalex.org/C189216375","wikidata":"https://www.wikidata.org/wiki/Q1127759","display_name":"Cognitive bias","level":3,"score":0.421999990940094},{"id":"https://openalex.org/C18762648","wikidata":"https://www.wikidata.org/wiki/Q42213","display_name":"Work (physics)","level":2,"score":0.3871999979019165},{"id":"https://openalex.org/C77805123","wikidata":"https://www.wikidata.org/wiki/Q161272","display_name":"Social psychology","level":1,"score":0.3702000081539154},{"id":"https://openalex.org/C164866538","wikidata":"https://www.wikidata.org/wiki/Q367351","display_name":"Cluster (spacecraft)","level":2,"score":0.36570000648498535},{"id":"https://openalex.org/C181447626","wikidata":"https://www.wikidata.org/wiki/Q1224359","display_name":"Differential item functioning","level":4,"score":0.32420000433921814},{"id":"https://openalex.org/C127729010","wikidata":"https://www.wikidata.org/wiki/Q60165","display_name":"Dynamic inconsistency","level":2,"score":0.31709998846054077},{"id":"https://openalex.org/C3019921246","wikidata":"https://www.wikidata.org/wiki/Q104853310","display_name":"Differential effects","level":2,"score":0.31060001254081726},{"id":"https://openalex.org/C2776817677","wikidata":"https://www.wikidata.org/wiki/Q4839818","display_name":"Backward masking","level":3,"score":0.2996000051498413},{"id":"https://openalex.org/C79585631","wikidata":"https://www.wikidata.org/wiki/Q431498","display_name":"Confirmation bias","level":2,"score":0.298799991607666},{"id":"https://openalex.org/C154945302","wikidata":"https://www.wikidata.org/wiki/Q11660","display_name":"Artificial intelligence","level":1,"score":0.27649998664855957},{"id":"https://openalex.org/C75917345","wikidata":"https://www.wikidata.org/wiki/Q2725298","display_name":"Sampling bias","level":3,"score":0.2605000138282776},{"id":"https://openalex.org/C40423286","wikidata":"https://www.wikidata.org/wiki/Q284172","display_name":"Selection bias","level":2,"score":0.2502000033855438}],"mesh":[],"locations_count":3,"locations":[{"id":"doi:10.18653/v1/2026.findings-acl.1893","is_oa":true,"landing_page_url":"https://doi.org/10.18653/v1/2026.findings-acl.1893","pdf_url":"https://aclanthology.org/2026.findings-acl.1893.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: ACL 2026","raw_type":"proceedings-article"},{"id":"pmh:oai:arXiv.org:2510.10142","is_oa":true,"landing_page_url":"http://arxiv.org/abs/2510.10142","pdf_url":"https://arxiv.org/pdf/2510.10142","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":null,"license_id":null,"version":"submittedVersion","is_accepted":false,"is_published":false,"raw_source_name":null,"raw_type":null},{"id":"doi:10.48550/arxiv.2510.10142","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2510.10142","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":null,"license_id":null,"version":null,"is_accepted":false,"is_published":null,"raw_source_name":null,"raw_type":"Preprint"}],"best_oa_location":{"id":"doi:10.18653/v1/2026.findings-acl.1893","is_oa":true,"landing_page_url":"https://doi.org/10.18653/v1/2026.findings-acl.1893","pdf_url":"https://aclanthology.org/2026.findings-acl.1893.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: ACL 2026","raw_type":"proceedings-article"},"sustainable_development_goals":[],"awards":[],"funders":[],"has_content":{"grobid_xml":true,"pdf":true},"content_urls":{"pdf":"https://content.openalex.org/works/W4415933347.pdf","grobid_xml":"https://content.openalex.org/works/W4415933347.grobid-xml"},"referenced_works_count":0,"referenced_works":[],"related_works":[],"abstract_inverted_index":{"Large":[0],"language":[1],"models":[2],"(LLMs)":[3],"increasingly":[4],"mediate":[5],"decisions":[6],"in":[7,87],"domains":[8],"where":[9],"unfair":[10],"treatment":[11],"of":[12,45,77,113],"demographic":[13],"groups":[14],"is":[15,177],"unacceptable.Existing":[16],"work":[17],"probes":[18],"when":[19],"biased":[20],"outputs":[21],"appear,":[22],"but":[23,120],"gives":[24],"little":[25],"insight":[26],"into":[27],"the":[28,74,78,127],"mechanisms":[29],"that":[30,99,116,143],"generate":[31],"them,":[32],"leaving":[33],"existing":[34],"mitigations":[35],"largely":[36,122],"fragile.In":[37],"this":[38],"paper,":[39],"we":[40,93,140],"conduct":[41],"a":[42,51,95,110],"systematic":[43],"investigation":[44],"LLM":[46,79],"unfairness":[47,83,162],"and":[48,80,90,134,153,155,165,169],"propose":[49,141],"DIFFHEADS,":[50],"lightweight":[52],"debiasing":[53],"framework":[54],"for":[55],"LLMs.We":[56],"first":[57,128],"compare":[58],"Direct-Answer":[59],"(DA)":[60],"prompting":[61,65,132],"to":[62,105],"Chain-of-Thought":[63],"(CoT)":[64],"across":[66],"eight":[67],"representative":[68],"open-and":[69],"closedsource":[70],"LLMs.DA":[71],"will":[72],"trigger":[73],"nature-bias":[75],"component":[76],"reduce":[81],"measured":[82],"by":[84,163],"391.9%":[85],"-534.5%":[86],"both":[88],"oneturn":[89],"two-turn":[91],"dialogues.Next,":[92],"define":[94],"token-to-head":[96],"contribution":[97],"score":[98],"traces":[100],"each":[101],"token's":[102],"influence":[103],"back":[104],"individual":[106],"attention":[107],"heads.This":[108],"reveals":[109],"small":[111],"cluster":[112],"bias":[114,135,145],"heads":[115,146],"activate":[117],"under":[118,167],"DA":[119,152,168],"stay":[121],"dormant":[123],"with":[124],"CoT,":[125,154,170],"providing":[126],"causal":[129],"link":[130],"between":[131,151],"strategy":[133],"emergence.Finally,":[136],"building":[137],"on":[138],"this,":[139],"DIFFHEADS":[142],"identifies":[144],"through":[147],"differential":[148],"activation":[149],"analysis":[150],"selectively":[156],"masks":[157],"only":[158],"those":[159],"heads.DIFFHEADS":[160],"reduces":[161],"49.4%,":[164],"40.3%":[166],"respectively,":[171],"without":[172],"harming":[173],"model":[174],"utility.Our":[175],"code":[176],"available":[178],"at":[179],"https://github.":[180],"com/GeniusHTX/DIFFHEADS.":[181]},"counts_by_year":[],"updated_date":"2026-08-02T14:50:37.381335","created_date":"2025-10-15T00:00:00"}