{"id":"https://openalex.org/W4415329439","doi":"https://doi.org/10.52202/085713-3388","title":"Reverse Engineering Human Preferences with Reinforcement Learning","display_name":"Reverse Engineering Human Preferences with Reinforcement Learning","publication_year":2025,"publication_date":"2025-01-01","ids":{"openalex":"https://openalex.org/W4415329439","doi":"https://doi.org/10.52202/085713-3388"},"language":"en","primary_location":{"id":"doi:10.52202/085713-3388","doi":"https://doi.org/10.52202/085713-3388","is_oa":false,"landing_page_url":"https://doi.org/10.52202/085713-3388","pdf_url":null,"source":{"id":"https://openalex.org/S4363606243","display_name":"neural information processing systems","issn_l":"1049-5258","issn":["1049-5258"],"is_oa":false,"is_in_doaj":false,"is_core":true,"listed_in":["cwts-core","norway-2"],"host_organization":null,"host_organization_name":null,"host_organization_lineage":[],"host_organization_lineage_names":[],"type":"conference"},"license":null,"license_id":null,"version":"publishedVersion","is_accepted":true,"is_published":true,"raw_source_name":"Advances in Neural Information Processing Systems 38","raw_type":"proceedings-article"},"volume":null,"type":"conference-paper","indexed_in":["arxiv","crossref","datacite"],"open_access":{"is_oa":true,"oa_status":"green","oa_url":"https://arxiv.org/pdf/2505.15795","any_repository_has_fulltext":true},"authorships":[{"author_position":"first","author":{"id":"https://openalex.org/A5001337463","display_name":"Lisa Alazraki","orcid":null,"observed_orcids":[]},"institutions":[{"id":"https://openalex.org/I47508984","display_name":"Imperial College London","ror":"https://ror.org/041kmwe10","country_code":"GB","type":"education","lineage":["https://openalex.org/I47508984"]}],"countries":["GB"],"is_corresponding":false,"raw_author_name":"Lisa Alazraki","raw_affiliation_strings":["Imperial College London"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"Imperial College London","institution_ids":["https://openalex.org/I47508984"]}]},{"author_position":"middle","author":{"id":"https://openalex.org/A5120054888","display_name":"Tan Yi-Chern","orcid":null,"observed_orcids":[]},"institutions":[{"id":"https://openalex.org/I4401726847","display_name":"Cohere (Canada)","ror":"https://ror.org/050yr0479","country_code":"US","type":"company","lineage":["https://openalex.org/I4401726847"]}],"countries":["US"],"is_corresponding":false,"raw_author_name":"Yi-Chern Tan","raw_affiliation_strings":["Cohere"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"Cohere","institution_ids":["https://openalex.org/I4401726847"]}]},{"author_position":"middle","author":{"id":"https://openalex.org/A5099010529","display_name":"Jon Ander Campos","orcid":null,"observed_orcids":[]},"institutions":[{"id":"https://openalex.org/I4401726847","display_name":"Cohere (Canada)","ror":"https://ror.org/050yr0479","country_code":"US","type":"company","lineage":["https://openalex.org/I4401726847"]}],"countries":["US"],"is_corresponding":false,"raw_author_name":"Jon Ander Campos","raw_affiliation_strings":["Cohere"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"Cohere","institution_ids":["https://openalex.org/I4401726847"]}]},{"author_position":"middle","author":{"id":"https://openalex.org/A5038732160","display_name":"Maximilian Mozes","orcid":"https://orcid.org/0000-0001-8138-3792","observed_orcids":["https://orcid.org/0000-0001-8138-3792"]},"institutions":[{"id":"https://openalex.org/I4401726847","display_name":"Cohere (Canada)","ror":"https://ror.org/050yr0479","country_code":"US","type":"company","lineage":["https://openalex.org/I4401726847"]}],"countries":["US"],"is_corresponding":false,"raw_author_name":"Maximilian Mozes","raw_affiliation_strings":["Cohere"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"Cohere","institution_ids":["https://openalex.org/I4401726847"]}]},{"author_position":"middle","author":{"id":"https://openalex.org/A5062209641","display_name":"Marek Rei","orcid":null,"observed_orcids":[]},"institutions":[{"id":"https://openalex.org/I47508984","display_name":"Imperial College London","ror":"https://ror.org/041kmwe10","country_code":"GB","type":"education","lineage":["https://openalex.org/I47508984"]}],"countries":["GB"],"is_corresponding":false,"raw_author_name":"Marek Rei","raw_affiliation_strings":["Imperial College London"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"Imperial College London","institution_ids":["https://openalex.org/I47508984"]}]},{"author_position":"last","author":{"id":"https://openalex.org/A5018864006","display_name":"Max Bartolo","orcid":"https://orcid.org/0009-0007-3301-7895","observed_orcids":["https://orcid.org/0009-0007-3301-7895"]},"institutions":[{"id":"https://openalex.org/I4210090411","display_name":"Google DeepMind (United Kingdom)","ror":"https://ror.org/00971b260","country_code":"GB","type":"company","lineage":["https://openalex.org/I4210090411","https://openalex.org/I4210128969"]}],"countries":["GB"],"is_corresponding":false,"raw_author_name":"Max Bartolo","raw_affiliation_strings":["Google DeepMind"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"Google DeepMind","institution_ids":["https://openalex.org/I4210090411"]}]}],"institutions":[],"countries_distinct_count":2,"institutions_distinct_count":3,"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":0.26413988,"is_in_top_1_percent":false,"is_in_top_10_percent":false},"cited_by_percentile_year":null,"biblio":{"volume":null,"issue":null,"first_page":"112220","last_page":"112257"},"is_retracted":false,"is_paratext":false,"is_xpac":false,"primary_topic":{"id":"https://openalex.org/T11689","display_name":"Adversarial Robustness in Machine Learning","score":0.9897000193595886,"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"}},"topics":[{"id":"https://openalex.org/T11689","display_name":"Adversarial Robustness in Machine Learning","score":0.9897000193595886,"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":0.00570000009611249,"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":0.00139999995008111,"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/llm-as-a-judge","display_name":"LLM-as-a-judge","score":0.9210000038146973},{"id":"https://openalex.org/keywords/llm-evaluation","display_name":"LLM evaluation","score":0.9200000166893005},{"id":"https://openalex.org/keywords/adversarial-fine-tuning","display_name":"adversarial fine-tuning","score":0.7179999947547913},{"id":"https://openalex.org/keywords/reward-engineering","display_name":"reward engineering","score":0.3580000102519989},{"id":"https://openalex.org/keywords/reinforcement-learning","display_name":"reinforcement learning","score":0.46700000762939453},{"id":"https://openalex.org/keywords/model-transferability","display_name":"model transferability","score":0.5180000066757202},{"id":"https://openalex.org/keywords/undetectable-attacks","display_name":"undetectable attacks","score":0.6290000081062317}],"concepts":[{"id":"https://openalex.org/C22019652","wikidata":"https://www.wikidata.org/wiki/Q331309","display_name":"Overfitting","level":3,"score":0.7851999998092651},{"id":"https://openalex.org/C41008148","wikidata":"https://www.wikidata.org/wiki/Q21198","display_name":"Computer science","level":0,"score":0.671999990940094},{"id":"https://openalex.org/C97541855","wikidata":"https://www.wikidata.org/wiki/Q830687","display_name":"Reinforcement learning","level":2,"score":0.6050999760627747},{"id":"https://openalex.org/C2780691134","wikidata":"https://www.wikidata.org/wiki/Q1129448","display_name":"Preamble","level":3,"score":0.4927999973297119},{"id":"https://openalex.org/C2776207758","wikidata":"https://www.wikidata.org/wiki/Q5303302","display_name":"Downstream (manufacturing)","level":2,"score":0.49000000953674316},{"id":"https://openalex.org/C154945302","wikidata":"https://www.wikidata.org/wiki/Q11660","display_name":"Artificial intelligence","level":1,"score":0.48350000381469727},{"id":"https://openalex.org/C119857082","wikidata":"https://www.wikidata.org/wiki/Q2539","display_name":"Machine learning","level":1,"score":0.4661000072956085},{"id":"https://openalex.org/C48044578","wikidata":"https://www.wikidata.org/wiki/Q727490","display_name":"Scalability","level":2,"score":0.45419999957084656},{"id":"https://openalex.org/C37736160","wikidata":"https://www.wikidata.org/wiki/Q1801315","display_name":"Adversarial system","level":2,"score":0.421999990940094},{"id":"https://openalex.org/C191172861","wikidata":"https://www.wikidata.org/wiki/Q7899321","display_name":"Upstream (networking)","level":2,"score":0.39089998602867126},{"id":"https://openalex.org/C2780992000","wikidata":"https://www.wikidata.org/wiki/Q17016113","display_name":"Generator (circuit theory)","level":3,"score":0.365200012922287},{"id":"https://openalex.org/C207850805","wikidata":"https://www.wikidata.org/wiki/Q269608","display_name":"Reverse engineering","level":2,"score":0.33899998664855957},{"id":"https://openalex.org/C18762648","wikidata":"https://www.wikidata.org/wiki/Q42213","display_name":"Work (physics)","level":2,"score":0.3197999894618988},{"id":"https://openalex.org/C112930515","wikidata":"https://www.wikidata.org/wiki/Q4389547","display_name":"Risk analysis (engineering)","level":1,"score":0.3100000023841858},{"id":"https://openalex.org/C173801870","wikidata":"https://www.wikidata.org/wiki/Q201413","display_name":"Heuristic","level":2,"score":0.302700012922287},{"id":"https://openalex.org/C2779843651","wikidata":"https://www.wikidata.org/wiki/Q7390335","display_name":"SIGNAL (programming language)","level":2,"score":0.28679999709129333},{"id":"https://openalex.org/C137293760","wikidata":"https://www.wikidata.org/wiki/Q3621696","display_name":"Language model","level":2,"score":0.28290000557899475},{"id":"https://openalex.org/C127705205","wikidata":"https://www.wikidata.org/wiki/Q5748245","display_name":"Heuristics","level":2,"score":0.2685000002384186},{"id":"https://openalex.org/C554936623","wikidata":"https://www.wikidata.org/wiki/Q199657","display_name":"Reading (process)","level":2,"score":0.2662999927997589},{"id":"https://openalex.org/C90559484","wikidata":"https://www.wikidata.org/wiki/Q778379","display_name":"Expression (computer science)","level":2,"score":0.25279998779296875}],"mesh":[],"locations_count":3,"locations":[{"id":"doi:10.52202/085713-3388","doi":"https://doi.org/10.52202/085713-3388","is_oa":false,"landing_page_url":"https://doi.org/10.52202/085713-3388","pdf_url":null,"source":{"id":"https://openalex.org/S4363606243","display_name":"neural information processing systems","issn_l":"1049-5258","issn":["1049-5258"],"is_oa":false,"is_in_doaj":false,"is_core":true,"listed_in":["cwts-core","norway-2"],"host_organization":null,"host_organization_name":null,"host_organization_lineage":[],"host_organization_lineage_names":[],"type":"conference"},"license":null,"license_id":null,"version":"publishedVersion","is_accepted":true,"is_published":true,"raw_source_name":"Advances in Neural Information Processing Systems 38","raw_type":"proceedings-article"},{"id":"pmh:oai:arXiv.org:2505.15795","doi":null,"is_oa":true,"landing_page_url":"http://arxiv.org/abs/2505.15795","pdf_url":"https://arxiv.org/pdf/2505.15795","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":"other-oa","license_id":"https://openalex.org/licenses/other-oa","version":"submittedVersion","is_accepted":false,"is_published":false,"raw_source_name":null,"raw_type":"text"},{"id":"doi:10.48550/arxiv.2505.15795","doi":"https://doi.org/10.48550/arxiv.2505.15795","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2505.15795","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":false,"raw_source_name":null,"raw_type":"Preprint"}],"best_oa_location":{"id":"pmh:oai:arXiv.org:2505.15795","doi":null,"is_oa":true,"landing_page_url":"http://arxiv.org/abs/2505.15795","pdf_url":"https://arxiv.org/pdf/2505.15795","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":"other-oa","license_id":"https://openalex.org/licenses/other-oa","version":"submittedVersion","is_accepted":false,"is_published":false,"raw_source_name":null,"raw_type":"text"},"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/W4415329439.pdf","grobid_xml":"https://content.openalex.org/works/W4415329439.grobid-xml"},"referenced_works_count":0,"referenced_works":[],"related_works":["https://openalex.org/W4417287945","https://openalex.org/W4415330383","https://openalex.org/W4416036315","https://openalex.org/W7165165314","https://openalex.org/W4403662541","https://openalex.org/W4404782878","https://openalex.org/W7162562640","https://openalex.org/W7129719721","https://openalex.org/W7221833195","https://openalex.org/W4402671833"],"abstract_inverted_index":{"The":[0],"capabilities":[1],"of":[2,46,143,173],"Large":[3],"Language":[4],"Models":[5],"(LLMs)":[6],"are":[7,155,160],"routinely":[8],"evaluated":[9],"by":[10,56,71,87,190],"other":[11,123],"LLMs":[12,109,192],"trained":[13],"to":[14,33,42,64,69,92,101,193],"predict":[15],"human":[16,183],"preferences.":[17],"This":[18],"framework--known":[19],"as":[20,36,89],"LLM-as-a-judge--is":[21],"highly":[22],"scalable":[23],"and":[24,82,152,209],"relatively":[25],"low":[26],"cost.":[27],"However,":[28],"it":[29],"is":[30,134],"also":[31,138,180],"vulnerable":[32],"malicious":[34],"exploitation,":[35],"LLM":[37],"responses":[38],"can":[39,59,185],"be":[40,60,186],"tuned":[41,145],"overfit":[43],"the":[44,47,53,66,84,129,141,144,150,153,171],"preferences":[45,184],"judge.":[48],"Previous":[49],"work":[50],"shows":[51],"that":[52,96,107,140,159,182,201],"answers":[54],"generated":[55],"a":[57,72,79,90],"candidate-LLM":[58,151],"edited":[61],"post":[62],"hoc":[63],"maximise":[65],"score":[67],"assigned":[68],"them":[70],"judge-LLM.":[73],"In":[74],"this":[75],"study,":[76],"we":[77],"adopt":[78],"different":[80],"approach":[81,200],"use":[83],"signal":[85],"provided":[86],"judge-LLMs":[88],"reward":[91],"adversarially":[93],"tune":[94],"models":[95,113,158],"generate":[97],"text":[98],"preambles":[99,196],"designed":[100],"boost":[102],"downstream":[103],"performance.":[104],"We":[105,137],"find":[106,203],"frozen":[108],"pipelined":[110],"with":[111,157],"these":[112],"attain":[114],"higher":[115],"LLM-evaluation":[116],"scores":[117],"than":[118],"existing":[119],"frameworks.":[120],"Crucially,":[121],"unlike":[122],"frameworks":[124],"which":[125],"intervene":[126],"directly":[127],"on":[128],"model's":[130],"response,":[131],"our":[132],"method":[133],"virtually":[135],"undetectable.":[136],"demonstrate":[139,181],"effectiveness":[142],"preamble":[146],"generator":[147],"transfers":[148],"when":[149],"judge-LLM":[154],"replaced":[156],"not":[161],"used":[162],"during":[163],"training.":[164],"These":[165],"findings":[166],"raise":[167],"important":[168],"questions":[169],"about":[170],"design":[172],"more":[174],"reliable":[175],"LLM-as-a-judge":[176],"evaluation":[177],"settings.":[178],"They":[179],"reverse":[187],"engineered":[188],"effectively,":[189],"pipelining":[191],"optimise":[194],"upstream":[195],"via":[197],"reinforcement":[198],"learning--an":[199],"could":[202],"future":[204],"applications":[205],"in":[206],"diverse":[207],"tasks":[208],"domains":[210],"beyond":[211],"adversarial":[212],"attacks.":[213]},"counts_by_year":[],"updated_date":"2026-10-27T15:51:10.671819","created_date":"2025-10-19T00:00:00"}