{"id":"https://openalex.org/W7102392001","doi":"https://doi.org/10.48550/arxiv.2510.23751","title":"Debiasing Reward Models by Representation Learning with Guarantees","display_name":"Debiasing Reward Models by Representation Learning with Guarantees","publication_year":2025,"publication_date":"2025-10-27","ids":{"openalex":"https://openalex.org/W7102392001","doi":"https://doi.org/10.48550/arxiv.2510.23751"},"language":null,"primary_location":{"id":"pmh:oai:arXiv.org:2510.23751","doi":null,"is_oa":true,"landing_page_url":"http://arxiv.org/abs/2510.23751","pdf_url":"https://arxiv.org/pdf/2510.23751","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,"listed_in":[],"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":"text"},"volume":null,"type":"preprint","indexed_in":["arxiv","datacite"],"open_access":{"is_oa":true,"oa_status":"green","oa_url":"https://arxiv.org/pdf/2510.23751","any_repository_has_fulltext":true},"authorships":[{"author_position":"first","author":{"id":"https://openalex.org/A5067738651","display_name":"Ignavier Ng","orcid":null,"observed_orcids":[]},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Ng, Ignavier","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5086464851","display_name":"Patrick Bl\u00f6baum","orcid":null,"observed_orcids":[]},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Bl\u00f6baum, Patrick","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":null,"display_name":"Bhandari, Siddharth","orcid":null,"observed_orcids":[]},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Bhandari, Siddharth","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5136630040","display_name":"Kun Zhang","orcid":"https://orcid.org/0000-0003-4617-6258","observed_orcids":["https://orcid.org/0000-0003-4617-6258"]},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Zhang, Kun","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"last","author":{"id":null,"display_name":"Kasiviswanathan, Shiva","orcid":null,"observed_orcids":[]},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Kasiviswanathan, Shiva","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":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":true,"primary_topic":{"id":"https://openalex.org/T10883","display_name":"Ethics and Social Impacts of AI","score":0.5292999744415283,"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":0.5292999744415283,"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":0.15719999372959137,"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/T10462","display_name":"Reinforcement Learning in Robotics","score":0.0625,"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/spurious-correlations","display_name":"spurious correlations","score":0.8190000057220459},{"id":"https://openalex.org/keywords/representation-learning","display_name":"representation learning","score":0.6299999952316284},{"id":"https://openalex.org/keywords/large-language-model-alignment","display_name":"large language model alignment","score":0.7350000143051147},{"id":"https://openalex.org/keywords/reinforcement-learning-from-human-feedback","display_name":"reinforcement learning from human feedback","score":0.9670000076293945},{"id":"https://openalex.org/keywords/conceptual-bias","display_name":"conceptual bias","score":0.6930000185966492},{"id":"https://openalex.org/keywords/variational-bayesian-inference","display_name":"variational inference","score":0.9639999866485596},{"id":"https://openalex.org/keywords/preference-modeling","display_name":"preference modeling","score":0.3880000114440918}],"concepts":[{"id":"https://openalex.org/C97256817","wikidata":"https://www.wikidata.org/wiki/Q1462316","display_name":"Spurious relationship","level":2,"score":0.9248999953269958},{"id":"https://openalex.org/C2779458634","wikidata":"https://www.wikidata.org/wiki/Q24963715","display_name":"Debiasing","level":2,"score":0.8716999888420105},{"id":"https://openalex.org/C41008148","wikidata":"https://www.wikidata.org/wiki/Q21198","display_name":"Computer science","level":0,"score":0.738099992275238},{"id":"https://openalex.org/C2776214188","wikidata":"https://www.wikidata.org/wiki/Q408386","display_name":"Inference","level":2,"score":0.6938999891281128},{"id":"https://openalex.org/C51167844","wikidata":"https://www.wikidata.org/wiki/Q4422623","display_name":"Latent variable","level":2,"score":0.6543999910354614},{"id":"https://openalex.org/C154945302","wikidata":"https://www.wikidata.org/wiki/Q11660","display_name":"Artificial intelligence","level":1,"score":0.6248000264167786},{"id":"https://openalex.org/C119857082","wikidata":"https://www.wikidata.org/wiki/Q2539","display_name":"Machine learning","level":1,"score":0.6247000098228455},{"id":"https://openalex.org/C97541855","wikidata":"https://www.wikidata.org/wiki/Q830687","display_name":"Reinforcement learning","level":2,"score":0.5726000070571899},{"id":"https://openalex.org/C2776359362","wikidata":"https://www.wikidata.org/wiki/Q2145286","display_name":"Representation (politics)","level":3,"score":0.5430999994277954},{"id":"https://openalex.org/C165696696","wikidata":"https://www.wikidata.org/wiki/Q11287","display_name":"Exploit","level":2,"score":0.48489999771118164},{"id":"https://openalex.org/C59404180","wikidata":"https://www.wikidata.org/wiki/Q17013334","display_name":"Feature learning","level":2,"score":0.39640000462532043},{"id":"https://openalex.org/C2780451532","wikidata":"https://www.wikidata.org/wiki/Q759676","display_name":"Task (project management)","level":2,"score":0.365200012922287},{"id":"https://openalex.org/C28006648","wikidata":"https://www.wikidata.org/wiki/Q6934509","display_name":"Multi-task learning","level":3,"score":0.3208000063896179},{"id":"https://openalex.org/C117220453","wikidata":"https://www.wikidata.org/wiki/Q5172842","display_name":"Correlation","level":2,"score":0.2842000126838684},{"id":"https://openalex.org/C65965080","wikidata":"https://www.wikidata.org/wiki/Q1806885","display_name":"Latent variable model","level":3,"score":0.2816999852657318},{"id":"https://openalex.org/C137002209","wikidata":"https://www.wikidata.org/wiki/Q898521","display_name":"Hidden variable theory","level":3,"score":0.2741999924182892},{"id":"https://openalex.org/C67186912","wikidata":"https://www.wikidata.org/wiki/Q367664","display_name":"Data modeling","level":2,"score":0.25380000472068787}],"mesh":[],"locations_count":2,"locations":[{"id":"pmh:oai:arXiv.org:2510.23751","doi":null,"is_oa":true,"landing_page_url":"http://arxiv.org/abs/2510.23751","pdf_url":"https://arxiv.org/pdf/2510.23751","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,"listed_in":[],"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":"text"},{"id":"doi:10.48550/arxiv.2510.23751","doi":"https://doi.org/10.48550/arxiv.2510.23751","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2510.23751","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,"listed_in":[],"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":"pmh:oai:arXiv.org:2510.23751","doi":null,"is_oa":true,"landing_page_url":"http://arxiv.org/abs/2510.23751","pdf_url":"https://arxiv.org/pdf/2510.23751","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,"listed_in":[],"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":"text"},"sustainable_development_goals":[],"sustainable_development_goals_aurora":[{"id":"https://metadata.un.org/sdg/10","display_name":"Reduced inequalities","score":0.6937137246131897}],"study_designs":[],"awards":[],"funders":[],"has_content":{"grobid_xml":false,"pdf":false},"content_urls":null,"referenced_works_count":0,"referenced_works":[],"related_works":[],"abstract_inverted_index":{"Recent":[0],"alignment":[1],"techniques,":[2],"such":[3],"as":[4],"reinforcement":[5],"learning":[6,23],"from":[7,96,115],"human":[8,20],"feedback,":[9],"have":[10],"been":[11],"widely":[12],"adopted":[13],"to":[14,139,146],"align":[15],"large":[16],"language":[17],"models":[18,31,68],"with":[19],"preferences":[21],"by":[22],"and":[24,42,99,143,153,165],"leveraging":[25],"reward":[26,67,148,169],"models.":[27,149,170],"In":[28,54],"practice,":[29],"these":[30,64,107,141],"often":[32],"exploit":[33],"spurious":[34,98,124,162],"correlations,":[35],"involving,":[36],"e.g.,":[37],"response":[38],"length,":[39],"discrimination,":[40],"sycophancy,":[41],"conceptual":[43],"bias,":[44],"which":[45],"is":[46,94,127],"a":[47,59,81,120,132],"problem":[48],"that":[49,62,74,88,135,157],"has":[50],"received":[51],"increasing":[52],"attention.":[53],"this":[55],"work,":[56],"we":[57],"propose":[58],"principled":[60],"framework":[61],"mitigates":[63,161],"biases":[65],"in":[66],"while":[69],"preserving":[70],"the":[71,84,89,123],"underlying":[72],"factors":[73],"reflect":[75],"intended":[76],"preferences.":[77],"We":[78,103],"first":[79],"provide":[80],"formulation":[82],"of":[83,118],"data-generating":[85],"process,":[86],"assuming":[87],"observed":[90],"data":[91],"(e.g.,":[92],"text)":[93],"generated":[95],"both":[97],"non-spurious":[100,108],"latent":[101,109,125],"variables.":[102],"show":[104],"that,":[105],"interestingly,":[106],"variables":[110,126,142],"can":[111],"be":[112],"theoretically":[113],"identified":[114],"data,":[116],"regardless":[117],"whether":[119],"surrogate":[121],"for":[122],"available.":[128],"This":[129],"further":[130],"inspires":[131],"practical":[133],"method":[134,159],"uses":[136],"variational":[137],"inference":[138],"recover":[140],"leverages":[144],"them":[145],"train":[147],"Experiments":[150],"on":[151],"synthetic":[152],"real-world":[154],"datasets":[155],"demonstrate":[156],"our":[158],"effectively":[160],"correlation":[163],"issues":[164],"yields":[166],"more":[167],"robust":[168]},"counts_by_year":[],"updated_date":"2026-10-13T10:08:12.033832","created_date":"2025-10-30T00:00:00"}