{"id":"https://openalex.org/W4412889688","doi":"https://doi.org/10.18653/v1/2025.acl-long.1504","title":"Efficient Safety Alignment of Large Language Models via Preference Re-ranking and Representation-based Reward Modeling","display_name":"Efficient Safety Alignment of Large Language Models via Preference Re-ranking and Representation-based Reward Modeling","publication_year":2025,"publication_date":"2025-01-01","ids":{"openalex":"https://openalex.org/W4412889688","doi":"https://doi.org/10.18653/v1/2025.acl-long.1504"},"language":"en","primary_location":{"id":"doi:10.18653/v1/2025.acl-long.1504","is_oa":true,"landing_page_url":"https://doi.org/10.18653/v1/2025.acl-long.1504","pdf_url":"https://aclanthology.org/2025.acl-long.1504.pdf","source":null,"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 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)","raw_type":"proceedings-article"},"type":"conference-paper","indexed_in":["crossref"],"open_access":{"is_oa":true,"oa_status":"gold","oa_url":"https://aclanthology.org/2025.acl-long.1504.pdf","any_repository_has_fulltext":null},"authorships":[{"author_position":"first","author":{"id":"https://openalex.org/A5059591124","display_name":"Deng Qi-yuan","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Deng Qiyuan","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5101905617","display_name":"Xuefeng Bai","orcid":"https://orcid.org/0000-0001-7044-0683"},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Xuefeng Bai","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5006323375","display_name":"Kehai Chen","orcid":"https://orcid.org/0000-0002-4346-7618"},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Kehai Chen","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5100631214","display_name":"Yaowei Wang","orcid":"https://orcid.org/0000-0002-7853-1587"},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Yaowei Wang","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5038612499","display_name":"Liqiang Nie","orcid":"https://orcid.org/0000-0003-1476-0273"},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Liqiang Nie","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"last","author":{"id":"https://openalex.org/A5100618695","display_name":"Min Zhang","orcid":"https://orcid.org/0009-0003-4063-0082"},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Min Zhang","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":true,"cited_by_count":0,"citation_normalized_percentile":null,"cited_by_percentile_year":null,"biblio":{"volume":null,"issue":null,"first_page":"31156","last_page":"31171"},"is_retracted":false,"is_paratext":false,"is_xpac":false,"primary_topic":{"id":"https://openalex.org/T10181","display_name":"Natural Language Processing Techniques","score":0.9740999937057495,"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/T10181","display_name":"Natural Language Processing Techniques","score":0.9740999937057495,"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":0.9596999883651733,"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/T11719","display_name":"Data Quality and Management","score":0.9132999777793884,"subfield":{"id":"https://openalex.org/subfields/1803","display_name":"Management Science and Operations Research"},"field":{"id":"https://openalex.org/fields/18","display_name":"Decision Sciences"},"domain":{"id":"https://openalex.org/domains/2","display_name":"Social Sciences"}}],"keywords":[{"id":"https://openalex.org/keywords/computer-science","display_name":"Computer science","score":0.6816499829292297},{"id":"https://openalex.org/keywords/preference","display_name":"Preference","score":0.6516062021255493},{"id":"https://openalex.org/keywords/ranking","display_name":"Ranking (information retrieval)","score":0.6358224153518677},{"id":"https://openalex.org/keywords/representation","display_name":"Representation (politics)","score":0.5850588083267212},{"id":"https://openalex.org/keywords/natural-language-processing","display_name":"Natural language processing","score":0.4886029362678528},{"id":"https://openalex.org/keywords/artificial-intelligence","display_name":"Artificial intelligence","score":0.4625280499458313},{"id":"https://openalex.org/keywords/statistics","display_name":"Statistics","score":0.16891586780548096},{"id":"https://openalex.org/keywords/mathematics","display_name":"Mathematics","score":0.14178940653800964}],"concepts":[{"id":"https://openalex.org/C41008148","wikidata":"https://www.wikidata.org/wiki/Q21198","display_name":"Computer science","level":0,"score":0.6816499829292297},{"id":"https://openalex.org/C2781249084","wikidata":"https://www.wikidata.org/wiki/Q908656","display_name":"Preference","level":2,"score":0.6516062021255493},{"id":"https://openalex.org/C189430467","wikidata":"https://www.wikidata.org/wiki/Q7293293","display_name":"Ranking (information retrieval)","level":2,"score":0.6358224153518677},{"id":"https://openalex.org/C2776359362","wikidata":"https://www.wikidata.org/wiki/Q2145286","display_name":"Representation (politics)","level":3,"score":0.5850588083267212},{"id":"https://openalex.org/C204321447","wikidata":"https://www.wikidata.org/wiki/Q30642","display_name":"Natural language processing","level":1,"score":0.4886029362678528},{"id":"https://openalex.org/C154945302","wikidata":"https://www.wikidata.org/wiki/Q11660","display_name":"Artificial intelligence","level":1,"score":0.4625280499458313},{"id":"https://openalex.org/C105795698","wikidata":"https://www.wikidata.org/wiki/Q12483","display_name":"Statistics","level":1,"score":0.16891586780548096},{"id":"https://openalex.org/C33923547","wikidata":"https://www.wikidata.org/wiki/Q395","display_name":"Mathematics","level":0,"score":0.14178940653800964},{"id":"https://openalex.org/C94625758","wikidata":"https://www.wikidata.org/wiki/Q7163","display_name":"Politics","level":2,"score":0.0},{"id":"https://openalex.org/C199539241","wikidata":"https://www.wikidata.org/wiki/Q7748","display_name":"Law","level":1,"score":0.0},{"id":"https://openalex.org/C17744445","wikidata":"https://www.wikidata.org/wiki/Q36442","display_name":"Political science","level":0,"score":0.0}],"mesh":[],"locations_count":1,"locations":[{"id":"doi:10.18653/v1/2025.acl-long.1504","is_oa":true,"landing_page_url":"https://doi.org/10.18653/v1/2025.acl-long.1504","pdf_url":"https://aclanthology.org/2025.acl-long.1504.pdf","source":null,"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 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)","raw_type":"proceedings-article"}],"best_oa_location":{"id":"doi:10.18653/v1/2025.acl-long.1504","is_oa":true,"landing_page_url":"https://doi.org/10.18653/v1/2025.acl-long.1504","pdf_url":"https://aclanthology.org/2025.acl-long.1504.pdf","source":null,"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 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)","raw_type":"proceedings-article"},"sustainable_development_goals":[],"awards":[{"id":"https://openalex.org/G1289754401","display_name":null,"funder_award_id":"U24A20328","funder_id":"https://openalex.org/F4320321001","funder_display_name":"National Natural Science Foundation of China"},{"id":"https://openalex.org/G3537643952","display_name":null,"funder_award_id":"62350710797","funder_id":"https://openalex.org/F4320321001","funder_display_name":"National Natural Science Foundation of China"},{"id":"https://openalex.org/G7320519987","display_name":null,"funder_award_id":"62276077","funder_id":"https://openalex.org/F4320321001","funder_display_name":"National Natural Science Foundation of China"}],"funders":[{"id":"https://openalex.org/F4320321001","display_name":"National Natural Science Foundation of China","ror":"https://ror.org/01h0zpd94"}],"has_content":{"pdf":true,"grobid_xml":true},"content_urls":{"pdf":"https://content.openalex.org/works/W4412889688.pdf","grobid_xml":"https://content.openalex.org/works/W4412889688.grobid-xml"},"referenced_works_count":0,"referenced_works":[],"related_works":["https://openalex.org/W2188500270","https://openalex.org/W2303858293","https://openalex.org/W2915512527","https://openalex.org/W51364034","https://openalex.org/W2793336762","https://openalex.org/W2091548507","https://openalex.org/W2368816706","https://openalex.org/W2931602588","https://openalex.org/W1986265453","https://openalex.org/W3204019825"],"abstract_inverted_index":{"Reinforcement":[0],"Learning":[1],"(RL)":[2],"algorithms":[3],"for":[4,113],"safety":[5,98,134],"alignment":[6],"of":[7,21,53,69,82],"Large":[8],"Language":[9],"Models":[10],"(LLMs),":[11],"such":[12],"as":[13],"Direct":[14],"Preference":[15],"Optimization":[16],"(DPO),":[17],"encounter":[18],"the":[19,33,50,67,70,74,95,122,127,133],"challenge":[20],"distribution":[22,61,128],"shift.Current":[23],"approaches":[24],"typically":[25],"address":[26],"this":[27,41,86],"issue":[28],"through":[29],"online":[30],"sampling":[31,71],"from":[32,73],"target":[34,75],"policy,":[35],"which":[36,105],"requires":[37],"significant":[38],"computational":[39,140],"resources.In":[40],"paper,":[42],"we":[43,88],"hypothesize":[44],"that":[45,93,121],"during":[46],"off-policy":[47],"training,":[48],"while":[49,136],"ranking":[51],"order":[52],"output":[54],"generated":[55],"by":[56],"policy":[57,76],"changes,":[58],"their":[59],"overall":[60],"remains":[62],"relatively":[63],"stable.This":[64],"stability":[65],"allows":[66],"conversion":[68],"process":[72],"into":[77],"a":[78,90],"computationally":[79],"efficient":[80],"reranking":[81],"preference":[83,114],"data.Building":[84],"on":[85],"hypothesis,":[87],"propose":[89],"new":[91],"framework":[92],"leverages":[94],"model's":[96],"intrinsic":[97],"judgment":[99],"capability":[100],"to":[101,109],"extract":[102],"reward":[103],"signals,":[104],"are":[106],"then":[107],"used":[108],"calculate":[110],"label":[111],"confidence":[112],"reordering.Extensive":[115],"experiments":[116],"and":[117],"theoretical":[118],"analysis":[119],"demonstrate":[120],"proposed":[123],"method":[124],"effectively":[125],"addresses":[126],"shift":[129],"issue,":[130],"remarkably":[131],"enhancing":[132],"performance":[135],"avoiding":[137],"about":[138],"300x":[139],"overheads.":[141]},"counts_by_year":[],"updated_date":"2026-07-14T23:27:15.235271","created_date":"2025-10-10T00:00:00"}