{"id":"https://openalex.org/W4407632011","doi":"https://doi.org/10.48550/arxiv.2502.09674","title":"The Hidden Dimensions of LLM Alignment: A Multi-Dimensional Analysis of Orthogonal Safety Directions","display_name":"The Hidden Dimensions of LLM Alignment: A Multi-Dimensional Analysis of Orthogonal Safety Directions","publication_year":2025,"publication_date":"2025-02-13","ids":{"openalex":"https://openalex.org/W4407632011","doi":"https://doi.org/10.48550/arxiv.2502.09674"},"language":"en","primary_location":{"id":"pmh:oai:arXiv.org:2502.09674","doi":null,"is_oa":true,"landing_page_url":"http://arxiv.org/abs/2502.09674","pdf_url":"https://arxiv.org/pdf/2502.09674","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":"cc-by","license_id":"https://openalex.org/licenses/cc-by","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/2502.09674","any_repository_has_fulltext":true},"authorships":[{"author_position":"first","author":{"id":"https://openalex.org/A5133632806","display_name":"Wenbo Pan","orcid":null,"observed_orcids":[]},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Pan, Wenbo","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5116292997","display_name":"Zhichao Liu","orcid":null,"observed_orcids":[]},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Liu, Zhichao","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5116292998","display_name":"Qiguang Chen","orcid":null,"observed_orcids":[]},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Chen, Qiguang","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5137924235","display_name":"Xiangyang Zhou","orcid":null,"observed_orcids":[]},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Zhou, Xiangyang","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5116293000","display_name":"Haining Yu","orcid":null,"observed_orcids":[]},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Yu, Haining","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"last","author":{"id":"https://openalex.org/A5013643572","display_name":"Xiaohua Jia","orcid":"https://orcid.org/0000-0001-8702-8302","observed_orcids":["https://orcid.org/0000-0001-8702-8302"]},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Jia, Xiaohua","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":1,"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":false,"primary_topic":{"id":"https://openalex.org/T11689","display_name":"Adversarial Robustness in Machine Learning","score":0.9136000275611877,"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":0.9136000275611877,"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/T10883","display_name":"Ethics and Social Impacts of AI","score":0.02810000069439411,"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/T12026","display_name":"Explainable Artificial Intelligence (XAI)","score":0.018799999728798866,"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/activation-space","display_name":"activation space","score":0.5899999737739563},{"id":"https://openalex.org/keywords/representation-shift","display_name":"representation shift","score":0.5789999961853027},{"id":"https://openalex.org/keywords/refusal-behavior","display_name":"refusal behavior","score":0.7839999794960022},{"id":"https://openalex.org/keywords/llama-3","display_name":"Llama 3","score":0.718999981880188},{"id":"https://openalex.org/keywords/mechanistic-interpretability","display_name":"mechanistic interpretability","score":0.5960000157356262},{"id":"https://openalex.org/keywords/role-play","display_name":"role play","score":0.652999997138977}],"concepts":[{"id":"https://openalex.org/C41008148","wikidata":"https://www.wikidata.org/wiki/Q21198","display_name":"Computer science","level":0,"score":0.3818095922470093}],"mesh":[],"locations_count":2,"locations":[{"id":"pmh:oai:arXiv.org:2502.09674","doi":null,"is_oa":true,"landing_page_url":"http://arxiv.org/abs/2502.09674","pdf_url":"https://arxiv.org/pdf/2502.09674","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":"cc-by","license_id":"https://openalex.org/licenses/cc-by","version":"submittedVersion","is_accepted":false,"is_published":false,"raw_source_name":null,"raw_type":"text"},{"id":"doi:10.48550/arxiv.2502.09674","doi":"https://doi.org/10.48550/arxiv.2502.09674","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2502.09674","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":"cc-by","license_id":"https://openalex.org/licenses/cc-by","version":null,"is_accepted":false,"is_published":false,"raw_source_name":null,"raw_type":"Preprint"}],"best_oa_location":{"id":"pmh:oai:arXiv.org:2502.09674","doi":null,"is_oa":true,"landing_page_url":"http://arxiv.org/abs/2502.09674","pdf_url":"https://arxiv.org/pdf/2502.09674","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":"cc-by","license_id":"https://openalex.org/licenses/cc-by","version":"submittedVersion","is_accepted":false,"is_published":false,"raw_source_name":null,"raw_type":"text"},"sustainable_development_goals":[],"study_designs":[],"awards":[],"funders":[{"id":"https://openalex.org/F4320307285","display_name":"Impact Fund","ror":"https://ror.org/00jb20j87"},{"id":"https://openalex.org/F4320309893","display_name":"City University of Hong Kong","ror":"https://ror.org/03q8dnn23"}],"has_content":{"grobid_xml":true,"pdf":true},"content_urls":{"pdf":"https://content.openalex.org/works/W4407632011.pdf","grobid_xml":"https://content.openalex.org/works/W4407632011.grobid-xml"},"referenced_works_count":0,"referenced_works":[],"related_works":["https://openalex.org/W7105055559","https://openalex.org/W4414957397","https://openalex.org/W7164846010","https://openalex.org/W7167063237","https://openalex.org/W7165818020","https://openalex.org/W4414850569","https://openalex.org/W7163174745","https://openalex.org/W4403573019","https://openalex.org/W7166733414","https://openalex.org/W7157694911"],"abstract_inverted_index":{"Large":[0],"Language":[1],"Models'":[2],"safety-aligned":[3,42],"behaviors,":[4],"such":[5],"as":[6],"refusing":[7,67],"harmful":[8,136],"queries,":[9],"can":[10,138],"be":[11],"represented":[12],"by":[13,47],"linear":[14],"directions":[15,72,91,107,120,141],"in":[16,73,121,135],"activation":[17],"space.":[18],"Previous":[19],"research":[20],"modeled":[21],"safety":[22,34,60,146,153],"behavior":[23,43],"with":[24],"a":[25,80,157],"single":[26],"direction,":[27,113],"limiting":[28],"mechanistic":[29],"understanding":[30,152],"to":[31,142],"an":[32],"isolated":[33],"feature.":[35],"In":[36],"this":[37],"work,":[38],"we":[39,51,76,128],"discover":[40],"that":[41,79,130],"is":[44],"jointly":[45],"controlled":[46],"multi-dimensional":[48,158],"directions.":[49],"Namely,":[50],"study":[52],"the":[53,74,84,111,115,123,144],"vector":[54],"space":[55],"of":[56,118],"representation":[57],"shifts":[58],"during":[59],"fine-tuning":[61],"on":[62,151],"Llama":[63],"3":[64],"8B":[65],"for":[66],"jailbreaks.":[68],"By":[69],"studying":[70],"orthogonal":[71],"space,":[75],"first":[77],"find":[78],"dominant":[81,112],"direction":[82],"governs":[83],"model's":[85,124],"refusal":[86,125],"behavior,":[87],"while":[88],"multiple":[89],"smaller":[90],"represent":[92],"distinct":[93],"and":[94,100,161],"interpretable":[95],"features":[96],"like":[97],"hypothetical":[98],"narrative":[99],"role-playing.":[101],"We":[102],"then":[103],"measure":[104],"how":[105],"different":[106],"promote":[108],"or":[109],"suppress":[110],"showing":[114],"important":[116],"role":[117],"secondary":[119],"shaping":[122],"representation.":[126],"Finally,":[127],"demonstrate":[129],"removing":[131],"certain":[132],"trigger":[133],"tokens":[134],"queries":[137],"mitigate":[139],"these":[140],"bypass":[143],"learned":[145],"capability,":[147],"providing":[148],"new":[149],"insights":[150],"alignment":[154],"vulnerability":[155],"from":[156],"perspective.":[159],"Code":[160],"artifacts":[162],"are":[163],"available":[164],"at":[165],"https://github.com/BMPixel/safety-residual-space.":[166]},"counts_by_year":[{"year":2025,"cited_by_count":1}],"updated_date":"2026-10-14T16:39:01.820549","created_date":"2025-02-18T00:00:00"}