{"id":"https://openalex.org/W7166807385","doi":"https://doi.org/10.18653/v1/2026.acl-long.1029","title":"Debiasing Reward Models via Causally Motivated Inference-Time Intervention","display_name":"Debiasing Reward Models via Causally Motivated Inference-Time Intervention","publication_year":2026,"publication_date":"2026-01-01","ids":{"openalex":"https://openalex.org/W7166807385","doi":"https://doi.org/10.18653/v1/2026.acl-long.1029"},"language":"en","primary_location":{"id":"doi:10.18653/v1/2026.acl-long.1029","doi":"https://doi.org/10.18653/v1/2026.acl-long.1029","is_oa":true,"landing_page_url":"https://doi.org/10.18653/v1/2026.acl-long.1029","pdf_url":"https://aclanthology.org/2026.acl-long.1029.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":"cc-by","license_id":"https://openalex.org/licenses/cc-by","version":"publishedVersion","is_accepted":true,"is_published":true,"raw_source_name":"Proceedings of the 64th 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":"hybrid","oa_url":"https://aclanthology.org/2026.acl-long.1029.pdf","any_repository_has_fulltext":false},"authorships":[{"author_position":"first","author":{"id":"https://openalex.org/A5078963513","display_name":"Kazutoshi Shinoda","orcid":"https://orcid.org/0000-0003-1941-6172","observed_orcids":["https://orcid.org/0000-0003-1941-6172"]},"institutions":[{"id":"https://openalex.org/I2251713219","display_name":"NTT (Japan)","ror":"https://ror.org/00berct97","country_code":"JP","type":"company","lineage":["https://openalex.org/I2251713219"]}],"countries":["JP"],"is_corresponding":false,"raw_author_name":"Kazutoshi Shinoda","raw_affiliation_strings":["Human Informatics Labs. , NTT, Inc"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"Human Informatics Labs. , NTT, Inc","institution_ids":["https://openalex.org/I2251713219"]}]},{"author_position":"middle","author":{"id":"https://openalex.org/A5053698849","display_name":"Kosuke Nishida","orcid":null,"observed_orcids":[]},"institutions":[{"id":"https://openalex.org/I2251713219","display_name":"NTT (Japan)","ror":"https://ror.org/00berct97","country_code":"JP","type":"company","lineage":["https://openalex.org/I2251713219"]}],"countries":["JP"],"is_corresponding":false,"raw_author_name":"Kosuke Nishida","raw_affiliation_strings":["Human Informatics Labs. , NTT, Inc"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"Human Informatics Labs. , NTT, Inc","institution_ids":["https://openalex.org/I2251713219"]}]},{"author_position":"last","author":{"id":"https://openalex.org/A5110780218","display_name":"Kyosuke Nishida","orcid":"https://orcid.org/0000-0002-8443-0651","observed_orcids":["https://orcid.org/0000-0002-8443-0651"]},"institutions":[{"id":"https://openalex.org/I2251713219","display_name":"NTT (Japan)","ror":"https://ror.org/00berct97","country_code":"JP","type":"company","lineage":["https://openalex.org/I2251713219"]}],"countries":["JP"],"is_corresponding":false,"raw_author_name":"Kyosuke Nishida","raw_affiliation_strings":["Human Informatics Labs. , NTT, Inc"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"Human Informatics Labs. , NTT, Inc","institution_ids":["https://openalex.org/I2251713219"]}]}],"institutions":[],"countries_distinct_count":1,"institutions_distinct_count":1,"corresponding_author_ids":[],"corresponding_institution_ids":["https://openalex.org/I2251713219"],"apc_list":null,"apc_paid":null,"fwci":0.0,"has_fulltext":true,"cited_by_count":0,"citation_normalized_percentile":{"value":0.40934312,"is_in_top_1_percent":false,"is_in_top_10_percent":false},"cited_by_percentile_year":null,"biblio":{"volume":null,"issue":null,"first_page":"22474","last_page":"22489"},"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":0.295199990272522,"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":0.295199990272522,"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/T10028","display_name":"Topic Modeling","score":0.2312999963760376,"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/T10181","display_name":"Natural Language Processing Techniques","score":0.1476999968290329,"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/reward-models","display_name":"reward models","score":0.9049999713897705},{"id":"https://openalex.org/keywords/bias-mitigation","display_name":"bias mitigation","score":0.972000002861023},{"id":"https://openalex.org/keywords/inference-time-intervention","display_name":"inference-time intervention","score":0.75},{"id":"https://openalex.org/keywords/causal-inference","display_name":"causal inference","score":0.7139999866485596},{"id":"https://openalex.org/keywords/spurious-features","display_name":"spurious features","score":0.7070000171661377},{"id":"https://openalex.org/keywords/large-language-model-alignment","display_name":"large language model alignment","score":0.8259999752044678}],"concepts":[{"id":"https://openalex.org/C2779458634","wikidata":"https://www.wikidata.org/wiki/Q24963715","display_name":"Debiasing","level":2,"score":0.875},{"id":"https://openalex.org/C15744967","wikidata":"https://www.wikidata.org/wiki/Q9418","display_name":"Psychology","level":0,"score":0.626800000667572},{"id":"https://openalex.org/C2780665704","wikidata":"https://www.wikidata.org/wiki/Q959298","display_name":"Intervention (counseling)","level":2,"score":0.5640000104904175},{"id":"https://openalex.org/C180747234","wikidata":"https://www.wikidata.org/wiki/Q23373","display_name":"Cognitive psychology","level":1,"score":0.558899998664856},{"id":"https://openalex.org/C77805123","wikidata":"https://www.wikidata.org/wiki/Q161272","display_name":"Social psychology","level":1,"score":0.4115999937057495},{"id":"https://openalex.org/C41008148","wikidata":"https://www.wikidata.org/wiki/Q21198","display_name":"Computer science","level":0,"score":0.3562999963760376},{"id":"https://openalex.org/C177264268","wikidata":"https://www.wikidata.org/wiki/Q1514741","display_name":"Set (abstract data type)","level":2,"score":0.298799991607666},{"id":"https://openalex.org/C138496976","wikidata":"https://www.wikidata.org/wiki/Q175002","display_name":"Developmental psychology","level":1,"score":0.28769999742507935},{"id":"https://openalex.org/C148220186","wikidata":"https://www.wikidata.org/wiki/Q7111912","display_name":"Outcome (game theory)","level":2,"score":0.28290000557899475},{"id":"https://openalex.org/C154945302","wikidata":"https://www.wikidata.org/wiki/Q11660","display_name":"Artificial intelligence","level":1,"score":0.26980000734329224},{"id":"https://openalex.org/C2776035688","wikidata":"https://www.wikidata.org/wiki/Q1606558","display_name":"Affect (linguistics)","level":2,"score":0.2635999917984009},{"id":"https://openalex.org/C2775924081","wikidata":"https://www.wikidata.org/wiki/Q55608371","display_name":"Control (management)","level":2,"score":0.25}],"mesh":[],"locations_count":1,"locations":[{"id":"doi:10.18653/v1/2026.acl-long.1029","doi":"https://doi.org/10.18653/v1/2026.acl-long.1029","is_oa":true,"landing_page_url":"https://doi.org/10.18653/v1/2026.acl-long.1029","pdf_url":"https://aclanthology.org/2026.acl-long.1029.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":"cc-by","license_id":"https://openalex.org/licenses/cc-by","version":"publishedVersion","is_accepted":true,"is_published":true,"raw_source_name":"Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)","raw_type":"proceedings-article"}],"best_oa_location":{"id":"doi:10.18653/v1/2026.acl-long.1029","doi":"https://doi.org/10.18653/v1/2026.acl-long.1029","is_oa":true,"landing_page_url":"https://doi.org/10.18653/v1/2026.acl-long.1029","pdf_url":"https://aclanthology.org/2026.acl-long.1029.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":"cc-by","license_id":"https://openalex.org/licenses/cc-by","version":"publishedVersion","is_accepted":true,"is_published":true,"raw_source_name":"Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)","raw_type":"proceedings-article"},"sustainable_development_goals":[],"sustainable_development_goals_aurora":[],"study_designs":[],"awards":[],"funders":[],"has_content":{"pdf":true,"grobid_xml":true},"content_urls":{"pdf":"https://content.openalex.org/works/W7166807385.pdf","grobid_xml":"https://content.openalex.org/works/W7166807385.grobid-xml"},"referenced_works_count":0,"referenced_works":[],"related_works":["https://openalex.org/W7133860152","https://openalex.org/W7130342037","https://openalex.org/W7214809027","https://openalex.org/W7117692447","https://openalex.org/W7102392001","https://openalex.org/W7137203184","https://openalex.org/W4406549323","https://openalex.org/W7128700567","https://openalex.org/W7154564540","https://openalex.org/W4403795687"],"abstract_inverted_index":{"Reward":[0],"models":[1,11],"(RMs)":[2],"play":[3],"a":[4,139],"central":[5],"role":[6],"in":[7,40,56,91,126,157,169],"aligning":[8],"large":[9],"language":[10],"(LLMs)":[12],"with":[13,70,114],"human":[14],"preferences.However,":[15],"RMs":[16,57,110],"are":[17,67,152],"often":[18],"sensitive":[19],"to":[20,29,93,130,136],"spurious":[21,94],"features":[22,95],"such":[23],"as":[24],"response":[25,37],"length.Existing":[26],"inference-time":[27],"approaches":[28],"mitigating":[30,51],"these":[31,80],"biases":[32,55],"typically":[33],"focus":[34],"exclusively":[35],"on":[36,85,143,162],"length,":[38],"resulting":[39],"performance":[41,102,134],"trade-offs.In":[42],"this":[43],"paper,":[44],"we":[45],"propose":[46],"causally":[47],"motivated":[48],"intervention":[49,77],"for":[50,106],"multiple":[52],"types":[53],"of":[54,122,138,166],"at":[58],"inference":[59],"time.Our":[60],"method":[61,84],"first":[62],"identifies":[63],"neurons":[64,125,156],"whose":[65],"activations":[66],"strongly":[68],"correlated":[69],"predefined":[71],"bias":[72,98,150,167],"attributes,":[73],"and":[74,88,112,145],"applies":[75],"neuron-level":[76],"that":[78,137,149],"suppresses":[79],"signals.We":[81],"evaluate":[82],"our":[83,115],"RM":[86,142],"benchmarks":[87],"observe":[89],"reductions":[90],"sensitivity":[92],"across":[96],"diverse":[97],"types,":[99],"without":[100],"inducing":[101],"trade-offs.Moreover,":[103],"when":[104],"used":[105],"preference":[107],"annotation,":[108],"small":[109],"(2B":[111],"7B)":[113],"method,":[116],"which":[117],"edits":[118],"less":[119],"than":[120],"2%":[121],"all":[123],"the":[124,163],"RMs,":[127],"enable":[128],"LLMs":[129],"improve":[131],"alignment,":[132],"achieving":[133],"comparable":[135],"state-of-the-art":[140],"70B":[141],"AlpacaEval":[144],"MT-Bench.Further":[146],"analysis":[147],"reveals":[148],"signals":[151],"primarily":[153],"encoded":[154],"by":[155],"early":[158],"layers,":[159],"shedding":[160],"light":[161],"internal":[164],"mechanisms":[165],"exploitation":[168],"RMs.":[170]},"counts_by_year":[],"updated_date":"2026-10-29T09:56:32.521200","created_date":"2026-07-02T00:00:00"}