{"id":"https://openalex.org/W4417357097","doi":"https://doi.org/10.48550/arxiv.2506.10922","title":"Robustly Improving LLM Fairness in Realistic Settings via Interpretability","display_name":"Robustly Improving LLM Fairness in Realistic Settings via Interpretability","publication_year":2025,"publication_date":"2025-06-12","ids":{"openalex":"https://openalex.org/W4417357097","doi":"https://doi.org/10.48550/arxiv.2506.10922"},"language":"en","primary_location":{"id":"pmh:oai:arXiv.org:2506.10922","is_oa":true,"landing_page_url":"http://arxiv.org/abs/2506.10922","pdf_url":"https://arxiv.org/pdf/2506.10922","source":{"id":"https://openalex.org/S4393918464","display_name":"ArXiv.org","issn_l":"2331-8422","issn":["2331-8422"],"is_oa":true,"is_in_doaj":false,"is_core":false,"host_organization":null,"host_organization_name":null,"host_organization_lineage":[],"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":"text"},"type":"preprint","indexed_in":["arxiv","datacite"],"open_access":{"is_oa":true,"oa_status":"green","oa_url":"https://arxiv.org/pdf/2506.10922","any_repository_has_fulltext":true},"authorships":[{"author_position":"first","author":{"id":"https://openalex.org/A5109695738","display_name":"Adam Karvonen","orcid":null},"institutions":[],"countries":[],"is_corresponding":true,"raw_author_name":"Karvonen, Adam","raw_affiliation_strings":[],"affiliations":[]},{"author_position":"last","author":{"id":"https://openalex.org/A5037987827","display_name":"Samuel D. Marks","orcid":"https://orcid.org/0000-0001-6677-3274"},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Marks, Samuel","raw_affiliation_strings":[],"affiliations":[]}],"institutions":[],"countries_distinct_count":0,"institutions_distinct_count":2,"corresponding_author_ids":["https://openalex.org/A5109695738"],"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/T12970","display_name":"Names, Identity, and Discrimination Research","score":0.21649999916553497,"subfield":{"id":"https://openalex.org/subfields/3312","display_name":"Sociology and Political Science"},"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/T12970","display_name":"Names, Identity, and Discrimination Research","score":0.21649999916553497,"subfield":{"id":"https://openalex.org/subfields/3312","display_name":"Sociology and Political Science"},"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/T10883","display_name":"Ethics and Social Impacts of AI","score":0.18299999833106995,"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/T12380","display_name":"Authorship Attribution and Profiling","score":0.1525000035762787,"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/interpretability","display_name":"Interpretability","score":0.8414000272750854},{"id":"https://openalex.org/keywords/inference","display_name":"Inference","score":0.6128000020980835},{"id":"https://openalex.org/keywords/context","display_name":"Context (archaeology)","score":0.59579998254776},{"id":"https://openalex.org/keywords/race","display_name":"Race (biology)","score":0.41620001196861267},{"id":"https://openalex.org/keywords/intervention","display_name":"Intervention (counseling)","score":0.4104999899864197},{"id":"https://openalex.org/keywords/motivated-reasoning","display_name":"Motivated reasoning","score":0.3450999855995178},{"id":"https://openalex.org/keywords/simple","display_name":"Simple (philosophy)","score":0.3434000015258789},{"id":"https://openalex.org/keywords/relevance","display_name":"Relevance (law)","score":0.33970001339912415},{"id":"https://openalex.org/keywords/diversity","display_name":"Diversity (politics)","score":0.32919999957084656}],"concepts":[{"id":"https://openalex.org/C2781067378","wikidata":"https://www.wikidata.org/wiki/Q17027399","display_name":"Interpretability","level":2,"score":0.8414000272750854},{"id":"https://openalex.org/C2776214188","wikidata":"https://www.wikidata.org/wiki/Q408386","display_name":"Inference","level":2,"score":0.6128000020980835},{"id":"https://openalex.org/C2779343474","wikidata":"https://www.wikidata.org/wiki/Q3109175","display_name":"Context (archaeology)","level":2,"score":0.59579998254776},{"id":"https://openalex.org/C41008148","wikidata":"https://www.wikidata.org/wiki/Q21198","display_name":"Computer science","level":0,"score":0.5735999941825867},{"id":"https://openalex.org/C180747234","wikidata":"https://www.wikidata.org/wiki/Q23373","display_name":"Cognitive psychology","level":1,"score":0.44940000772476196},{"id":"https://openalex.org/C76509639","wikidata":"https://www.wikidata.org/wiki/Q918036","display_name":"Race (biology)","level":2,"score":0.41620001196861267},{"id":"https://openalex.org/C2780665704","wikidata":"https://www.wikidata.org/wiki/Q959298","display_name":"Intervention (counseling)","level":2,"score":0.4104999899864197},{"id":"https://openalex.org/C154945302","wikidata":"https://www.wikidata.org/wiki/Q11660","display_name":"Artificial intelligence","level":1,"score":0.376800000667572},{"id":"https://openalex.org/C119857082","wikidata":"https://www.wikidata.org/wiki/Q2539","display_name":"Machine learning","level":1,"score":0.37450000643730164},{"id":"https://openalex.org/C2776325391","wikidata":"https://www.wikidata.org/wiki/Q6917865","display_name":"Motivated reasoning","level":3,"score":0.3450999855995178},{"id":"https://openalex.org/C15744967","wikidata":"https://www.wikidata.org/wiki/Q9418","display_name":"Psychology","level":0,"score":0.3440000116825104},{"id":"https://openalex.org/C2780586882","wikidata":"https://www.wikidata.org/wiki/Q7520643","display_name":"Simple (philosophy)","level":2,"score":0.3434000015258789},{"id":"https://openalex.org/C158154518","wikidata":"https://www.wikidata.org/wiki/Q7310970","display_name":"Relevance (law)","level":2,"score":0.33970001339912415},{"id":"https://openalex.org/C2781316041","wikidata":"https://www.wikidata.org/wiki/Q1230584","display_name":"Diversity (politics)","level":2,"score":0.32919999957084656},{"id":"https://openalex.org/C40423286","wikidata":"https://www.wikidata.org/wiki/Q284172","display_name":"Selection bias","level":2,"score":0.3231000006198883},{"id":"https://openalex.org/C197947376","wikidata":"https://www.wikidata.org/wiki/Q5155608","display_name":"Comparability","level":2,"score":0.3163999915122986},{"id":"https://openalex.org/C2991991027","wikidata":"https://www.wikidata.org/wiki/Q6007314","display_name":"Implicit bias","level":2,"score":0.30979999899864197},{"id":"https://openalex.org/C2777877512","wikidata":"https://www.wikidata.org/wiki/Q1116097","display_name":"Common ground","level":2,"score":0.3091999888420105},{"id":"https://openalex.org/C2992700788","wikidata":"https://www.wikidata.org/wiki/Q8461","display_name":"Racial bias","level":3,"score":0.3086000084877014},{"id":"https://openalex.org/C2780084366","wikidata":"https://www.wikidata.org/wiki/Q37732","display_name":"Demographics","level":2,"score":0.30809998512268066},{"id":"https://openalex.org/C2522767166","wikidata":"https://www.wikidata.org/wiki/Q2374463","display_name":"Data science","level":1,"score":0.3068000078201294},{"id":"https://openalex.org/C77805123","wikidata":"https://www.wikidata.org/wiki/Q161272","display_name":"Social psychology","level":1,"score":0.3012000024318695},{"id":"https://openalex.org/C26760741","wikidata":"https://www.wikidata.org/wiki/Q160402","display_name":"Perception","level":2,"score":0.2782999873161316},{"id":"https://openalex.org/C2776889888","wikidata":"https://www.wikidata.org/wiki/Q1135789","display_name":"Unintended consequences","level":2,"score":0.2773999869823456},{"id":"https://openalex.org/C2983427547","wikidata":"https://www.wikidata.org/wiki/Q93200","display_name":"Gender bias","level":2,"score":0.275299996137619},{"id":"https://openalex.org/C158600405","wikidata":"https://www.wikidata.org/wiki/Q5054566","display_name":"Causal inference","level":2,"score":0.2712000012397766},{"id":"https://openalex.org/C189216375","wikidata":"https://www.wikidata.org/wiki/Q1127759","display_name":"Cognitive bias","level":3,"score":0.2660999894142151},{"id":"https://openalex.org/C79585631","wikidata":"https://www.wikidata.org/wiki/Q431498","display_name":"Confirmation bias","level":2,"score":0.2526000142097473},{"id":"https://openalex.org/C112930515","wikidata":"https://www.wikidata.org/wiki/Q4389547","display_name":"Risk analysis (engineering)","level":1,"score":0.25099998712539673}],"mesh":[],"locations_count":2,"locations":[{"id":"pmh:oai:arXiv.org:2506.10922","is_oa":true,"landing_page_url":"http://arxiv.org/abs/2506.10922","pdf_url":"https://arxiv.org/pdf/2506.10922","source":{"id":"https://openalex.org/S4393918464","display_name":"ArXiv.org","issn_l":"2331-8422","issn":["2331-8422"],"is_oa":true,"is_in_doaj":false,"is_core":false,"host_organization":null,"host_organization_name":null,"host_organization_lineage":[],"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":"text"},{"id":"doi:10.48550/arxiv.2506.10922","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2506.10922","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":"pmh:oai:arXiv.org:2506.10922","is_oa":true,"landing_page_url":"http://arxiv.org/abs/2506.10922","pdf_url":"https://arxiv.org/pdf/2506.10922","source":{"id":"https://openalex.org/S4393918464","display_name":"ArXiv.org","issn_l":"2331-8422","issn":["2331-8422"],"is_oa":true,"is_in_doaj":false,"is_core":false,"host_organization":null,"host_organization_name":null,"host_organization_lineage":[],"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":"text"},"sustainable_development_goals":[],"awards":[],"funders":[],"has_content":{"grobid_xml":false,"pdf":false},"content_urls":null,"referenced_works_count":0,"referenced_works":[],"related_works":[],"abstract_inverted_index":{"Large":[0],"language":[1],"models":[2,84,148,152],"(LLMs)":[3],"are":[4,43],"increasingly":[5],"deployed":[6],"in":[7,31,112,126],"high-stakes":[8],"hiring":[9,107,236],"applications,":[10],"making":[11],"decisions":[12],"that":[13,91,231],"directly":[14],"impact":[15],"people's":[16],"careers":[17,103],"and":[18,55,82,105,119,140,149,156,187,190,243],"livelihoods.":[19],"While":[20],"prior":[21],"studies":[22],"suggest":[23,230],"simple":[24,203],"anti-bias":[25],"prompts":[26],"can":[27,153],"eliminate":[28],"demographic":[29],"biases":[30,121,131,167],"controlled":[32],"evaluations,":[33],"we":[34,63,89],"find":[35,90],"these":[36,47,130,166,179],"mitigations":[37],"fail":[38],"when":[39,171],"realistic":[40,93,240],"contextual":[41],"details":[42],"introduced.":[44],"We":[45],"address":[46,178],"failures":[48],"through":[49],"internal":[50,182,245],"bias":[51,66,183,212],"mitigation:":[52],"by":[53],"identifying":[54],"neutralizing":[56],"sensitive":[57],"attribute":[58],"directions":[59,189,200],"within":[60],"model":[61,226],"activations,":[62],"achieve":[64],"robust":[65],"reduction":[67],"across":[68,145],"all":[69,146],"tested":[70,147],"scenarios.":[71,150],"Across":[72],"leading":[73],"commercial":[74],"(GPT-4o,":[75],"Claude":[76],"4":[77],"Sonnet,":[78],"Gemini":[79],"2.5":[80],"Flash)":[81],"open-source":[83],"(Gemma-2":[85],"27B,":[86],"Gemma-3,":[87],"Mistral-24B),":[88],"adding":[92],"context":[94],"such":[95],"as":[96],"company":[97],"names,":[98],"culture":[99],"descriptions":[100],"from":[101,159,201],"public":[102],"pages,":[104],"selective":[106],"constraints":[108],"(e.g.,``only":[109],"accept":[110],"candidates":[111,139,144],"the":[113,173,206],"top":[114],"10\\%\")":[115],"induces":[116],"significant":[117],"racial":[118],"gender":[120],"(up":[122],"to":[123,213],"12\\%":[124],"differences":[125],"interview":[127],"rates).":[128],"When":[129],"emerge,":[132],"they":[133],"consistently":[134,210],"favor":[135],"Black":[136],"over":[137,142],"White":[138],"female":[141],"male":[143],"Moreover,":[151],"infer":[154],"demographics":[155],"become":[157],"biased":[158],"subtle":[160],"cues":[161],"like":[162],"college":[163],"affiliations,":[164],"with":[165],"remaining":[168],"invisible":[169],"even":[170],"inspecting":[172],"model's":[174],"chain-of-thought":[175],"reasoning.":[176],"To":[177],"limitations,":[180],"our":[181],"mitigation":[184,246],"identifies":[185],"race":[186],"gender-correlated":[188],"applies":[191],"affine":[192],"concept":[193],"editing":[194],"at":[195],"inference":[196],"time.":[197],"Despite":[198],"using":[199],"a":[202],"synthetic":[204],"dataset,":[205],"intervention":[207],"generalizes":[208],"robustly,":[209],"reducing":[211],"very":[214],"low":[215],"levels":[216],"(typically":[217],"under":[218],"1\\%,":[219],"always":[220],"below":[221],"2.5\\%)":[222],"while":[223],"largely":[224],"maintaining":[225],"performance.":[227],"Our":[228],"findings":[229],"practitioners":[232],"deploying":[233],"LLMs":[234],"for":[235,248],"should":[237],"adopt":[238],"more":[239],"evaluation":[241],"methodologies":[242],"consider":[244],"strategies":[247],"equitable":[249],"outcomes.":[250]},"counts_by_year":[],"updated_date":"2026-03-20T23:20:44.827607","created_date":"2025-10-10T00:00:00"}