{"id":"https://openalex.org/W7140128870","doi":"https://doi.org/10.18653/v1/2026.findings-eacl.41","title":"Shifting Perspectives: Steering Vectors for Robust Bias Mitigation in LLMs","display_name":"Shifting Perspectives: Steering Vectors for Robust Bias Mitigation in LLMs","publication_year":2026,"publication_date":"2026-01-01","ids":{"openalex":"https://openalex.org/W7140128870","doi":"https://doi.org/10.18653/v1/2026.findings-eacl.41"},"language":"en","primary_location":{"id":"doi:10.18653/v1/2026.findings-eacl.41","doi":"https://doi.org/10.18653/v1/2026.findings-eacl.41","is_oa":true,"landing_page_url":"https://doi.org/10.18653/v1/2026.findings-eacl.41","pdf_url":"https://aclanthology.org/2026.findings-eacl.41.pdf","source":{"id":"https://openalex.org/S7407087051","display_name":"Findings of the Association for Computational Linguistics: EACL","issn_l":null,"issn":null,"is_oa":false,"is_in_doaj":false,"is_core":false,"listed_in":[],"host_organization":null,"host_organization_name":null,"host_organization_lineage":[],"host_organization_lineage_names":[],"type":"conference"},"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: EACL 2026","raw_type":"proceedings-article"},"volume":null,"type":"conference-paper","indexed_in":["crossref"],"open_access":{"is_oa":true,"oa_status":"hybrid","oa_url":"https://aclanthology.org/2026.findings-eacl.41.pdf","any_repository_has_fulltext":true},"authorships":[{"author_position":"first","author":{"id":"https://openalex.org/A5103305942","display_name":"Zara Siddique","orcid":null,"observed_orcids":[]},"institutions":[{"id":"https://openalex.org/I79510175","display_name":"Cardiff University","ror":"https://ror.org/03kk7td41","country_code":"GB","type":"education","lineage":["https://openalex.org/I79510175"]}],"countries":["GB"],"is_corresponding":false,"raw_author_name":"Zara Siddique","raw_affiliation_strings":["School of Computer Science and Informatics , Cardiff University , United Kingdom"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"School of Computer Science and Informatics , Cardiff University , United Kingdom","institution_ids":["https://openalex.org/I79510175"]}]},{"author_position":"middle","author":{"id":"https://openalex.org/A5077247933","display_name":"Irtaza Khalid","orcid":"https://orcid.org/0000-0002-1532-7829","observed_orcids":["https://orcid.org/0000-0002-1532-7829"]},"institutions":[{"id":"https://openalex.org/I79510175","display_name":"Cardiff University","ror":"https://ror.org/03kk7td41","country_code":"GB","type":"education","lineage":["https://openalex.org/I79510175"]}],"countries":["GB"],"is_corresponding":false,"raw_author_name":"Irtaza Khalid","raw_affiliation_strings":["School of Computer Science and Informatics , Cardiff University , United Kingdom"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"School of Computer Science and Informatics , Cardiff University , United Kingdom","institution_ids":["https://openalex.org/I79510175"]}]},{"author_position":"middle","author":{"id":"https://openalex.org/A5008814236","display_name":"Liam D. Turner","orcid":"https://orcid.org/0000-0003-4877-5289","observed_orcids":["https://orcid.org/0000-0003-4877-5289"]},"institutions":[{"id":"https://openalex.org/I79510175","display_name":"Cardiff University","ror":"https://ror.org/03kk7td41","country_code":"GB","type":"education","lineage":["https://openalex.org/I79510175"]}],"countries":["GB"],"is_corresponding":false,"raw_author_name":"Liam Turner","raw_affiliation_strings":["School of Computer Science and Informatics , Cardiff University , United Kingdom"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"School of Computer Science and Informatics , Cardiff University , United Kingdom","institution_ids":["https://openalex.org/I79510175"]}]},{"author_position":"last","author":{"id":"https://openalex.org/A5002661723","display_name":"Luis Espinosa-Anke","orcid":"https://orcid.org/0000-0001-6830-9176","observed_orcids":["https://orcid.org/0000-0001-6830-9176"]},"institutions":[{"id":"https://openalex.org/I79510175","display_name":"Cardiff University","ror":"https://ror.org/03kk7td41","country_code":"GB","type":"education","lineage":["https://openalex.org/I79510175"]}],"countries":["GB"],"is_corresponding":false,"raw_author_name":"Luis Espinosa-Anke","raw_affiliation_strings":["AMPLYFI , United Kingdom","School of Computer Science and Informatics , Cardiff University , United Kingdom"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"AMPLYFI , United Kingdom","institution_ids":[]},{"raw_affiliation_string":"School of Computer Science and Informatics , Cardiff University , United Kingdom","institution_ids":["https://openalex.org/I79510175"]}]}],"institutions":[],"countries_distinct_count":1,"institutions_distinct_count":1,"corresponding_author_ids":[],"corresponding_institution_ids":["https://openalex.org/I79510175"],"apc_list":null,"apc_paid":null,"fwci":9.4205,"has_fulltext":true,"cited_by_count":1,"citation_normalized_percentile":{"value":0.95272932,"is_in_top_1_percent":false,"is_in_top_10_percent":true},"cited_by_percentile_year":{"min":93,"max":96},"biblio":{"volume":null,"issue":null,"first_page":"809","last_page":"820"},"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":0.9595000147819519,"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":0.9595000147819519,"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/T12262","display_name":"Hate Speech and Cyberbullying Detection","score":0.007300000172108412,"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":0.006899999920278788,"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/bias-mitigation","display_name":"bias mitigation","score":0.8100000023841858},{"id":"https://openalex.org/keywords/large-language-models","display_name":"large language models","score":0.8640000224113464},{"id":"https://openalex.org/keywords/steering-vectors","display_name":"steering vectors","score":0.9990000128746033},{"id":"https://openalex.org/keywords/social-bias","display_name":"social bias","score":0.7540000081062317},{"id":"https://openalex.org/keywords/prompting","display_name":"prompting","score":0.4180000126361847},{"id":"https://openalex.org/keywords/fine-tuning","display_name":"fine-tuning","score":0.8939999938011169},{"id":"https://openalex.org/keywords/ai-safety","display_name":"AI safety","score":0.7009999752044678},{"id":"https://openalex.org/keywords/mmlu","display_name":"MMLU","score":0.5410000085830688}],"concepts":[{"id":"https://openalex.org/C41008148","wikidata":"https://www.wikidata.org/wiki/Q21198","display_name":"Computer science","level":0,"score":0.5116000175476074},{"id":"https://openalex.org/C112930515","wikidata":"https://www.wikidata.org/wiki/Q4389547","display_name":"Risk analysis (engineering)","level":1,"score":0.3783999979496002},{"id":"https://openalex.org/C47446073","wikidata":"https://www.wikidata.org/wiki/Q5165890","display_name":"Control theory (sociology)","level":3,"score":0.31529998779296875},{"id":"https://openalex.org/C18762648","wikidata":"https://www.wikidata.org/wiki/Q42213","display_name":"Work (physics)","level":2,"score":0.30559998750686646},{"id":"https://openalex.org/C63479239","wikidata":"https://www.wikidata.org/wiki/Q7353546","display_name":"Robustness (evolution)","level":3,"score":0.3000999987125397},{"id":"https://openalex.org/C162324750","wikidata":"https://www.wikidata.org/wiki/Q8134","display_name":"Economics","level":0,"score":0.29760000109672546},{"id":"https://openalex.org/C99498987","wikidata":"https://www.wikidata.org/wiki/Q2210247","display_name":"Noise (video)","level":3,"score":0.29739999771118164},{"id":"https://openalex.org/C127413603","wikidata":"https://www.wikidata.org/wiki/Q11023","display_name":"Engineering","level":0,"score":0.2955000102519989},{"id":"https://openalex.org/C2775924081","wikidata":"https://www.wikidata.org/wiki/Q55608371","display_name":"Control (management)","level":2,"score":0.2745000123977661},{"id":"https://openalex.org/C26517878","wikidata":"https://www.wikidata.org/wiki/Q228039","display_name":"Key (lock)","level":2,"score":0.26930001378059387},{"id":"https://openalex.org/C165064840","wikidata":"https://www.wikidata.org/wiki/Q1321061","display_name":"Matching (statistics)","level":2,"score":0.26420000195503235},{"id":"https://openalex.org/C32896092","wikidata":"https://www.wikidata.org/wiki/Q189447","display_name":"Risk management","level":2,"score":0.25200000405311584}],"mesh":[],"locations_count":2,"locations":[{"id":"doi:10.18653/v1/2026.findings-eacl.41","doi":"https://doi.org/10.18653/v1/2026.findings-eacl.41","is_oa":true,"landing_page_url":"https://doi.org/10.18653/v1/2026.findings-eacl.41","pdf_url":"https://aclanthology.org/2026.findings-eacl.41.pdf","source":{"id":"https://openalex.org/S7407087051","display_name":"Findings of the Association for Computational Linguistics: EACL","issn_l":null,"issn":null,"is_oa":false,"is_in_doaj":false,"is_core":false,"listed_in":[],"host_organization":null,"host_organization_name":null,"host_organization_lineage":[],"host_organization_lineage_names":[],"type":"conference"},"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: EACL 2026","raw_type":"proceedings-article"},{"id":"pmh:oai:https://orca.cardiff.ac.uk:186135","doi":null,"is_oa":true,"landing_page_url":"https://orca.cardiff.ac.uk/id/eprint/186135/1/2026.findings-eacl.41.pdf","pdf_url":"https://orca.cardiff.ac.uk/id/eprint/186135/1/2026.findings-eacl.41.pdf","source":{"id":"https://openalex.org/S4306401195","display_name":"ORCA Online Research @Cardiff (Cardiff University)","issn_l":null,"issn":null,"is_oa":false,"is_in_doaj":false,"is_core":false,"listed_in":[],"host_organization":"https://openalex.org/I79510175","host_organization_name":"Cardiff University","host_organization_lineage":["https://openalex.org/I79510175"],"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":"PeerReviewed"}],"best_oa_location":{"id":"doi:10.18653/v1/2026.findings-eacl.41","doi":"https://doi.org/10.18653/v1/2026.findings-eacl.41","is_oa":true,"landing_page_url":"https://doi.org/10.18653/v1/2026.findings-eacl.41","pdf_url":"https://aclanthology.org/2026.findings-eacl.41.pdf","source":{"id":"https://openalex.org/S7407087051","display_name":"Findings of the Association for Computational Linguistics: EACL","issn_l":null,"issn":null,"is_oa":false,"is_in_doaj":false,"is_core":false,"listed_in":[],"host_organization":null,"host_organization_name":null,"host_organization_lineage":[],"host_organization_lineage_names":[],"type":"conference"},"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: EACL 2026","raw_type":"proceedings-article"},"sustainable_development_goals":[{"id":"https://metadata.un.org/sdg/10","display_name":"Reduced inequalities","score":0.4074000120162964}],"study_designs":[],"awards":[],"funders":[{"id":"https://openalex.org/F4320314731","display_name":"UK Research and Innovation","ror":"https://ror.org/001aqnf71"}],"has_content":{"pdf":true,"grobid_xml":true},"content_urls":{"pdf":"https://content.openalex.org/works/W7140128870.pdf","grobid_xml":"https://content.openalex.org/works/W7140128870.grobid-xml"},"referenced_works_count":0,"referenced_works":[],"related_works":["https://openalex.org/W4405470208","https://openalex.org/W7127987255","https://openalex.org/W7160745628","https://openalex.org/W7126183608","https://openalex.org/W4416035237","https://openalex.org/W7128690443","https://openalex.org/W4416867973","https://openalex.org/W4405181259","https://openalex.org/W4416553088","https://openalex.org/W4414804972"],"abstract_inverted_index":{"We":[0],"present":[1],"a":[2,31,43,144],"novel":[3],"approach":[4],"to":[5,17,30,56],"bias":[6,34,59,122,136,152],"mitigation":[7,60,123],"in":[8,21,96,103,153],"large":[9],"language":[10],"models":[11],"(LLMs)":[12],"by":[13],"applying":[14],"steering":[15,26,73,110,133],"vectors":[16,74,111,134],"modify":[18],"model":[19],"activations":[20],"forward":[22],"passes.We":[23],"compute":[24],"8":[25],"vectors,":[27],"each":[28],"corresponding":[29],"different":[32],"social":[33],"axis,":[35],"such":[36],"as":[37],"age,":[38],"gender,":[39],"or":[40],"race,":[41],"on":[42,66,80,83,87,116],"training":[44],"subset":[45],"of":[46,54,78,106,119,132],"the":[47,52,67,113,120,128],"BBQ":[48,68],"dataset":[49],"and":[50,85,89,94,99,138,146],"compare":[51],"effectiveness":[53],"these":[55],"3":[57],"additional":[58],"methods":[61,124],"across":[62],"4":[63],"datasets.When":[64],"optimized":[65],"dataset,":[69],"our":[70],"individually":[71],"tuned":[72],"achieve":[75],"average":[76],"improvements":[77,91,100],"12.8%":[79],"BBQ,":[81],"8.3%":[82],"CLEAR-Bias,":[84],"1%":[86],"StereoSet,":[88],"show":[90],"over":[92,101],"prompting":[93],"Self-Debias":[95],"all":[97],"cases,":[98],"fine-tuning":[102],"12":[104],"out":[105],"17":[107],"evaluations.In":[108],"addition,":[109],"showed":[112],"lowest":[114],"impact":[115],"MMLU":[117],"scores":[118],"four":[121],"tested.The":[125],"work":[126],"presents":[127],"first":[129],"systematic":[130],"investigation":[131],"for":[135,150,158],"mitigation,":[137],"we":[139],"demonstrate":[140],"that":[141],"they":[142],"are":[143],"powerful":[145],"computationally":[147],"efficient":[148],"strategy":[149],"reducing":[151],"LLMs,":[154],"with":[155],"broader":[156],"implications":[157],"enhancing":[159],"AI":[160],"safety.":[161],"1":[162]},"counts_by_year":[{"year":2026,"cited_by_count":1}],"updated_date":"2026-10-12T09:40:04.860316","created_date":"2026-02-01T00:00:00"}