{"id":"https://openalex.org/W4412945154","doi":"https://doi.org/10.18653/v1/2025.acl-long.1173","title":"Representation Bending for Large Language Model Safety","display_name":"Representation Bending for Large Language Model Safety","publication_year":2025,"publication_date":"2025-01-01","ids":{"openalex":"https://openalex.org/W4412945154","doi":"https://doi.org/10.18653/v1/2025.acl-long.1173"},"language":"en","primary_location":{"id":"doi:10.18653/v1/2025.acl-long.1173","is_oa":false,"landing_page_url":"https://doi.org/10.18653/v1/2025.acl-long.1173","pdf_url":null,"source":null,"license":null,"license_id":null,"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":false,"oa_status":"closed","oa_url":null,"any_repository_has_fulltext":false},"authorships":[{"author_position":"first","author":{"id":"https://openalex.org/A5044916571","display_name":"Ashkan Yousefpour","orcid":"https://orcid.org/0000-0003-4869-9183"},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Ashkan Yousefpour","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5062099224","display_name":"Taeheon Kim","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Taeheon Kim","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5119204463","display_name":"Ryan Sungmo Kwon","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Ryan Sungmo Kwon","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5070650987","display_name":"Seungbeen Lee","orcid":"https://orcid.org/0009-0005-0728-7547"},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Seungbeen Lee","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5114619009","display_name":"Wonje Jeung","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Wonje Jeung","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5053165370","display_name":"Seungju Han","orcid":"https://orcid.org/0000-0001-7293-1419"},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Seungju Han","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5035953329","display_name":"Alvin Wan","orcid":"https://orcid.org/0000-0003-2625-5899"},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Alvin Wan","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":null,"display_name":"Harrison Ngan","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Harrison Ngan","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5101881857","display_name":"Youngjae Yu","orcid":"https://orcid.org/0000-0002-5867-0782"},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Youngjae Yu","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"last","author":{"id":"https://openalex.org/A5073483751","display_name":"Jonghyun Choi","orcid":"https://orcid.org/0000-0002-7934-8434"},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Jonghyun Choi","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":3,"citation_normalized_percentile":null,"cited_by_percentile_year":null,"biblio":{"volume":null,"issue":null,"first_page":"24073","last_page":"24098"},"is_retracted":false,"is_paratext":false,"is_xpac":false,"primary_topic":{"id":"https://openalex.org/T11689","display_name":"Adversarial Robustness in Machine Learning","score":0.8848999738693237,"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":0.8848999738693237,"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/T10743","display_name":"Software Testing and Debugging Techniques","score":0.8515999913215637,"subfield":{"id":"https://openalex.org/subfields/1712","display_name":"Software"},"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/T11357","display_name":"Risk and Safety Analysis","score":0.8499000072479248,"subfield":{"id":"https://openalex.org/subfields/1804","display_name":"Statistics, Probability and Uncertainty"},"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":0.6153576374053955},{"id":"https://openalex.org/keywords/representation","display_name":"Representation (politics)","score":0.5812459588050842},{"id":"https://openalex.org/keywords/bending","display_name":"Bending","score":0.45102518796920776},{"id":"https://openalex.org/keywords/natural-language-processing","display_name":"Natural language processing","score":0.406355082988739},{"id":"https://openalex.org/keywords/structural-engineering","display_name":"Structural engineering","score":0.22870832681655884},{"id":"https://openalex.org/keywords/engineering","display_name":"Engineering","score":0.17517772316932678},{"id":"https://openalex.org/keywords/political-science","display_name":"Political science","score":0.07545086741447449}],"concepts":[{"id":"https://openalex.org/C41008148","wikidata":"https://www.wikidata.org/wiki/Q21198","display_name":"Computer science","level":0,"score":0.6153576374053955},{"id":"https://openalex.org/C2776359362","wikidata":"https://www.wikidata.org/wiki/Q2145286","display_name":"Representation (politics)","level":3,"score":0.5812459588050842},{"id":"https://openalex.org/C87210426","wikidata":"https://www.wikidata.org/wiki/Q1072476","display_name":"Bending","level":2,"score":0.45102518796920776},{"id":"https://openalex.org/C204321447","wikidata":"https://www.wikidata.org/wiki/Q30642","display_name":"Natural language processing","level":1,"score":0.406355082988739},{"id":"https://openalex.org/C66938386","wikidata":"https://www.wikidata.org/wiki/Q633538","display_name":"Structural engineering","level":1,"score":0.22870832681655884},{"id":"https://openalex.org/C127413603","wikidata":"https://www.wikidata.org/wiki/Q11023","display_name":"Engineering","level":0,"score":0.17517772316932678},{"id":"https://openalex.org/C17744445","wikidata":"https://www.wikidata.org/wiki/Q36442","display_name":"Political science","level":0,"score":0.07545086741447449},{"id":"https://openalex.org/C199539241","wikidata":"https://www.wikidata.org/wiki/Q7748","display_name":"Law","level":1,"score":0.0},{"id":"https://openalex.org/C94625758","wikidata":"https://www.wikidata.org/wiki/Q7163","display_name":"Politics","level":2,"score":0.0}],"mesh":[],"locations_count":2,"locations":[{"id":"doi:10.18653/v1/2025.acl-long.1173","is_oa":false,"landing_page_url":"https://doi.org/10.18653/v1/2025.acl-long.1173","pdf_url":null,"source":null,"license":null,"license_id":null,"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"},{"id":"pmh:oai:s-space.snu.ac.kr:10371/230305","is_oa":false,"landing_page_url":"https://hdl.handle.net/10371/230305","pdf_url":null,"source":{"id":"https://openalex.org/S4306401345","display_name":"Seoul National University Open Repository (Seoul National University)","issn_l":null,"issn":null,"is_oa":false,"is_in_doaj":false,"is_core":false,"host_organization":"https://openalex.org/I139264467","host_organization_name":"Seoul National University","host_organization_lineage":["https://openalex.org/I139264467"],"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":"Conference Paper"}],"best_oa_location":null,"sustainable_development_goals":[],"awards":[],"funders":[],"has_content":{"pdf":false,"grobid_xml":false},"content_urls":null,"referenced_works_count":0,"referenced_works":[],"related_works":["https://openalex.org/W4391375266","https://openalex.org/W2899084033","https://openalex.org/W2748952813","https://openalex.org/W2390279801","https://openalex.org/W4391913857","https://openalex.org/W2358668433","https://openalex.org/W4396701345","https://openalex.org/W2376932109","https://openalex.org/W2001405890","https://openalex.org/W4396696052"],"abstract_inverted_index":{"Large":[0],"Language":[1],"Models":[2],"(LLMs)":[3],"have":[4],"emerged":[5],"as":[6,53,64,140],"powerful":[7],"tools,":[8],"but":[9],"their":[10],"inherent":[11],"safety":[12],"risks":[13,29],"-":[14,24,115,125],"ranging":[15],"from":[16],"harmful":[17,95],"content":[18],"generation":[19],"to":[20,72,103,126,148],"broader":[21],"societal":[22],"harms":[23],"pose":[25],"significant":[26],"challenges.":[27],"These":[28],"can":[30],"be":[31],"amplified":[32],"by":[33],"the":[34,41,92,110],"recent":[35],"adversarial":[36,59],"attacks,":[37,76],"fine-tuning":[38,54],"vulnerabilities,":[39],"and":[40,69,144,166],"increasing":[42],"deployment":[43],"of":[44,112],"LLMs":[45],"in":[46,97,151,163],"high-stakes":[47],"environments.":[48],"Existing":[49],"safety-enhancing":[50],"techniques,":[51],"such":[52,139],"with":[55,146,160],"human":[56],"feedback":[57],"or":[58,77],"training,":[60],"are":[61],"still":[62],"vulnerable":[63],"they":[65],"address":[66],"specific":[67],"threats":[68],"often":[70],"fail":[71],"generalize":[73],"across":[74,155],"unseen":[75],"require":[78],"manual":[79],"system-level":[80],"defenses.":[81],"This":[82],"paper":[83],"introduces":[84],"REPBEND,":[85],"a":[86,100],"novel":[87],"approach":[88],"that":[89],"fundamentally":[90],"disrupts":[91],"representations":[93],"underlying":[94],"behaviors":[96],"LLMs,":[98],"offering":[99],"scalable":[101],"solution":[102],"enhance":[104],"(potentially":[105],"inherent)":[106],"safety.":[107],"REPBEND":[108,132],"brings":[109],"idea":[111],"activation":[113],"steering":[114,120],"simple":[116],"vector":[117],"arithmetic":[118],"for":[119],"model's":[121],"behavior":[122],"during":[123],"inference":[124],"loss-based":[127],"fine-tuning.":[128],"Through":[129],"extensive":[130],"evaluation,":[131],"achieves":[133],"state-of-the-art":[134],"performance,":[135],"outperforming":[136],"prior":[137],"methods":[138],"Circuit":[141],"Breaker,":[142],"RMU,":[143],"NPO,":[145],"up":[147],"95%":[149],"reduction":[150,162],"attack":[152],"success":[153],"rates":[154],"diverse":[156],"jailbreak":[157],"benchmarks,":[158],"all":[159],"negligible":[161],"model":[164],"usability":[165],"general":[167],"capabilities.":[168]},"counts_by_year":[{"year":2026,"cited_by_count":2},{"year":2025,"cited_by_count":1}],"updated_date":"2026-07-22T07:51:19.307946","created_date":"2025-10-10T00:00:00"}