{"id":"https://openalex.org/W4416035194","doi":"https://doi.org/10.18653/v1/2025.findings-emnlp.366","title":"SafeSwitch: Steering Unsafe LLM Behavior via Internal Activation Signals","display_name":"SafeSwitch: Steering Unsafe LLM Behavior via Internal Activation Signals","publication_year":2025,"publication_date":"2025-01-01","ids":{"openalex":"https://openalex.org/W4416035194","doi":"https://doi.org/10.18653/v1/2025.findings-emnlp.366"},"language":null,"primary_location":{"id":"doi:10.18653/v1/2025.findings-emnlp.366","is_oa":true,"landing_page_url":"https://doi.org/10.18653/v1/2025.findings-emnlp.366","pdf_url":"https://aclanthology.org/2025.findings-emnlp.366.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":"Findings of the Association for Computational Linguistics: EMNLP 2025","raw_type":"proceedings-article"},"type":"conference-paper","indexed_in":["crossref"],"open_access":{"is_oa":true,"oa_status":"gold","oa_url":"https://aclanthology.org/2025.findings-emnlp.366.pdf","any_repository_has_fulltext":null},"authorships":[{"author_position":"first","author":{"id":null,"display_name":"Peixuan Han","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Peixuan Han","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5100651524","display_name":"Cheng Qian","orcid":"https://orcid.org/0000-0003-1500-9991"},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Cheng Qian","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5114125839","display_name":"Xiusi Chen","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Xiusi Chen","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5000223629","display_name":"Yuji Zhang","orcid":"https://orcid.org/0000-0002-5429-6762"},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Yuji Zhang","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5007960166","display_name":"Heng Ji","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Heng Ji","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"last","author":{"id":"https://openalex.org/A5101638182","display_name":"Denghui Zhang","orcid":"https://orcid.org/0000-0003-4492-3643"},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Denghui 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":2,"citation_normalized_percentile":null,"cited_by_percentile_year":null,"biblio":{"volume":null,"issue":null,"first_page":"6936","last_page":"6955"},"is_retracted":false,"is_paratext":false,"is_xpac":false,"primary_topic":{"id":"https://openalex.org/T10142","display_name":"Formal Methods in Verification","score":0.09589999914169312,"subfield":{"id":"https://openalex.org/subfields/1703","display_name":"Computational Theory and Mathematics"},"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/T10142","display_name":"Formal Methods in Verification","score":0.09589999914169312,"subfield":{"id":"https://openalex.org/subfields/1703","display_name":"Computational Theory and Mathematics"},"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/T10639","display_name":"Advanced Software Engineering Methodologies","score":0.04749999940395355,"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/T11424","display_name":"Security and Verification in Computing","score":0.04670000076293945,"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/control-theory","display_name":"Control theory (sociology)","score":0.37389999628067017},{"id":"https://openalex.org/keywords/noise","display_name":"Noise (video)","score":0.29429998993873596},{"id":"https://openalex.org/keywords/component","display_name":"Component (thermodynamics)","score":0.2816999852657318},{"id":"https://openalex.org/keywords/work","display_name":"Work (physics)","score":0.267300009727478},{"id":"https://openalex.org/keywords/control-system","display_name":"Control system","score":0.25110000371932983}],"concepts":[{"id":"https://openalex.org/C41008148","wikidata":"https://www.wikidata.org/wiki/Q21198","display_name":"Computer science","level":0,"score":0.4893999993801117},{"id":"https://openalex.org/C47446073","wikidata":"https://www.wikidata.org/wiki/Q5165890","display_name":"Control theory (sociology)","level":3,"score":0.37389999628067017},{"id":"https://openalex.org/C127413603","wikidata":"https://www.wikidata.org/wiki/Q11023","display_name":"Engineering","level":0,"score":0.32179999351501465},{"id":"https://openalex.org/C99498987","wikidata":"https://www.wikidata.org/wiki/Q2210247","display_name":"Noise (video)","level":3,"score":0.29429998993873596},{"id":"https://openalex.org/C168167062","wikidata":"https://www.wikidata.org/wiki/Q1117970","display_name":"Component (thermodynamics)","level":2,"score":0.2816999852657318},{"id":"https://openalex.org/C133731056","wikidata":"https://www.wikidata.org/wiki/Q4917288","display_name":"Control engineering","level":1,"score":0.2750999927520752},{"id":"https://openalex.org/C18762648","wikidata":"https://www.wikidata.org/wiki/Q42213","display_name":"Work (physics)","level":2,"score":0.267300009727478},{"id":"https://openalex.org/C17500928","wikidata":"https://www.wikidata.org/wiki/Q959968","display_name":"Control system","level":2,"score":0.25110000371932983},{"id":"https://openalex.org/C112972136","wikidata":"https://www.wikidata.org/wiki/Q7595718","display_name":"Stability (learning theory)","level":2,"score":0.24619999527931213},{"id":"https://openalex.org/C2780791683","wikidata":"https://www.wikidata.org/wiki/Q846785","display_name":"Action (physics)","level":2,"score":0.24469999969005585}],"mesh":[],"locations_count":1,"locations":[{"id":"doi:10.18653/v1/2025.findings-emnlp.366","is_oa":true,"landing_page_url":"https://doi.org/10.18653/v1/2025.findings-emnlp.366","pdf_url":"https://aclanthology.org/2025.findings-emnlp.366.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":"Findings of the Association for Computational Linguistics: EMNLP 2025","raw_type":"proceedings-article"}],"best_oa_location":{"id":"doi:10.18653/v1/2025.findings-emnlp.366","is_oa":true,"landing_page_url":"https://doi.org/10.18653/v1/2025.findings-emnlp.366","pdf_url":"https://aclanthology.org/2025.findings-emnlp.366.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":"Findings of the Association for Computational Linguistics: EMNLP 2025","raw_type":"proceedings-article"},"sustainable_development_goals":[],"awards":[],"funders":[],"has_content":{"pdf":true,"grobid_xml":true},"content_urls":{"pdf":"https://content.openalex.org/works/W4416035194.pdf","grobid_xml":"https://content.openalex.org/works/W4416035194.grobid-xml"},"referenced_works_count":0,"referenced_works":[],"related_works":[],"abstract_inverted_index":{"Large":[0],"language":[1,150],"models":[2],"(LLMs)":[3],"exhibit":[4],"exceptional":[5],"capabilities":[6],"across":[7],"various":[8],"tasks":[9],"but":[10],"also":[11,123],"pose":[12],"risks":[13],"by":[14,39,73,104],"generating":[15],"harmful":[16,83,102,108],"content.Existing":[17],"safety":[18,55,88,168],"mechanisms,":[19],"while":[20,110,138],"improving":[21],"model":[22],"safety,":[23,158],"often":[24],"lead":[25],"to":[26,32,92,163],"overly":[27],"cautious":[28],"behavior":[29],"and":[30,85,94,134,155,166],"fail":[31],"fully":[33],"leverage":[34],"LLMs'":[35],"internal":[36,52,58,77],"cognitive":[37],"processes.Inspired":[38],"humans'":[40],"reflective":[41],"thinking":[42],"capability,":[43],"we":[44,63],"first":[45],"show":[46],"that":[47,69,80,90],"LLMs":[48],"can":[49],"similarly":[50],"perform":[51],"assessments":[53],"about":[54],"in":[56,128],"their":[57],"states.Building":[59],"on":[60,107],"this":[61,171],"insight,":[62],"propose":[64],"SafeSwitch,":[65],"a":[66,87,115,160],"dynamic":[67],"framework":[68],"regulates":[70],"unsafe":[71],"outputs":[72,103],"utilizing":[74],"the":[75,145],"prober-based":[76],"state":[78],"monitor":[79],"actively":[81],"detects":[82],"intentions,":[84],"activates":[86],"head":[89],"leads":[91],"safer":[93],"more":[95,130,164],"conservative":[96],"responses":[97],"only":[98,139],"when":[99],"necessary.SafeSwitch":[100],"reduces":[101],"approximately":[105],"80%":[106],"queries":[109],"maintaining":[111],"strong":[112],"utility,":[113],"reaching":[114],"Pareto":[116],"optimal":[117],"among":[118],"several":[119],"methods.Our":[120],"method":[121],"is":[122],"advantageous":[124],"over":[125],"traditional":[126],"methods":[127],"offering":[129,159],"informative,":[131],"context-aware":[132],"refusals,":[133],"achieves":[135],"these":[136],"benefits":[137],"tuning":[140],"less":[141],"than":[142],"6%":[143],"of":[144],"original":[146],"parameters.SafeSwitch":[147],"demonstrates":[148],"large":[149],"models'":[151],"capacity":[152],"for":[153,170],"self-awareness":[154],"reflection":[156],"regarding":[157],"promising":[161],"approach":[162],"nuanced":[165],"effective":[167],"controls.Codes":[169],"work":[172],"are":[173],"available":[174],"at":[175],"https://":[176],"github.com/Hanpx20/SafeSwitch.":[177]},"counts_by_year":[{"year":2025,"cited_by_count":2}],"updated_date":"2026-07-14T23:27:15.235271","created_date":"2025-11-08T00:00:00"}
