{"id":"https://openalex.org/W7166809691","doi":"https://doi.org/10.18653/v1/2026.acl-long.1380","title":"Embracing Anisotropy: Turning Massive Activations into Interpretable Control Knobs for Large Language Models","display_name":"Embracing Anisotropy: Turning Massive Activations into Interpretable Control Knobs for Large Language Models","publication_year":2026,"publication_date":"2026-01-01","ids":{"openalex":"https://openalex.org/W7166809691","doi":"https://doi.org/10.18653/v1/2026.acl-long.1380"},"language":"en","primary_location":{"id":"doi:10.18653/v1/2026.acl-long.1380","doi":"https://doi.org/10.18653/v1/2026.acl-long.1380","is_oa":true,"landing_page_url":"https://doi.org/10.18653/v1/2026.acl-long.1380","pdf_url":"https://aclanthology.org/2026.acl-long.1380.pdf","source":{"id":"https://openalex.org/S4393915740","display_name":"Proceedings of the conference - Association for Computational Linguistics. Meeting","issn_l":"0736-587X","issn":["0736-587X"],"is_oa":false,"is_in_doaj":false,"is_core":false,"listed_in":["norway-1"],"host_organization":null,"host_organization_name":null,"host_organization_lineage":[],"host_organization_lineage_names":[],"type":"conference"},"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 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)","raw_type":"proceedings-article"},"volume":null,"type":"conference-paper","indexed_in":["crossref"],"open_access":{"is_oa":true,"oa_status":"hybrid","oa_url":"https://aclanthology.org/2026.acl-long.1380.pdf","any_repository_has_fulltext":false},"authorships":[{"author_position":"first","author":{"id":"https://openalex.org/A5127914347","display_name":"Youngji Roh","orcid":null,"observed_orcids":[]},"institutions":[{"id":"https://openalex.org/I193775966","display_name":"Yonsei University","ror":"https://ror.org/01wjejq96","country_code":"KR","type":"education","lineage":["https://openalex.org/I193775966"]}],"countries":["KR"],"is_corresponding":false,"raw_author_name":"Youngji Roh","raw_affiliation_strings":["Yonsei University"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"Yonsei University","institution_ids":["https://openalex.org/I193775966"]}]},{"author_position":"middle","author":{"id":"https://openalex.org/A5104254030","display_name":"Hyunjin Cho","orcid":"https://orcid.org/0009-0004-6849-3216","observed_orcids":["https://orcid.org/0009-0004-6849-3216"]},"institutions":[{"id":"https://openalex.org/I193775966","display_name":"Yonsei University","ror":"https://ror.org/01wjejq96","country_code":"KR","type":"education","lineage":["https://openalex.org/I193775966"]}],"countries":["KR"],"is_corresponding":false,"raw_author_name":"Hyunjin Cho","raw_affiliation_strings":["Yonsei University"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"Yonsei University","institution_ids":["https://openalex.org/I193775966"]}]},{"author_position":"last","author":{"id":"https://openalex.org/A5139807853","display_name":"Jaehyung Kim","orcid":null,"observed_orcids":[]},"institutions":[{"id":"https://openalex.org/I193775966","display_name":"Yonsei University","ror":"https://ror.org/01wjejq96","country_code":"KR","type":"education","lineage":["https://openalex.org/I193775966"]}],"countries":["KR"],"is_corresponding":false,"raw_author_name":"Jaehyung Kim","raw_affiliation_strings":["Yonsei University"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"Yonsei University","institution_ids":["https://openalex.org/I193775966"]}]}],"institutions":[],"countries_distinct_count":1,"institutions_distinct_count":1,"corresponding_author_ids":[],"corresponding_institution_ids":["https://openalex.org/I193775966"],"apc_list":null,"apc_paid":null,"fwci":0.0,"has_fulltext":true,"cited_by_count":0,"citation_normalized_percentile":{"value":0.43467912,"is_in_top_1_percent":false,"is_in_top_10_percent":false},"cited_by_percentile_year":null,"biblio":{"volume":null,"issue":null,"first_page":"29930","last_page":"29956"},"is_retracted":false,"is_paratext":false,"is_xpac":false,"primary_topic":{"id":"https://openalex.org/T12026","display_name":"Explainable Artificial Intelligence (XAI)","score":0.9926999807357788,"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/T12026","display_name":"Explainable Artificial Intelligence (XAI)","score":0.9926999807357788,"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":0.002899999963119626,"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/T11689","display_name":"Adversarial Robustness in Machine Learning","score":0.0015999999595806003,"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/large-language-models","display_name":"large language models","score":0.921999990940094},{"id":"https://openalex.org/keywords/activation-steering","display_name":"activation steering","score":0.8309999704360962},{"id":"https://openalex.org/keywords/interpretability","display_name":"interpretability","score":0.7319999933242798},{"id":"https://openalex.org/keywords/domain-adaptation","display_name":"domain adaptation","score":0.7680000066757202},{"id":"https://openalex.org/keywords/jailbreaking","display_name":"jailbreaking","score":0.9599999785423279}],"concepts":[{"id":"https://openalex.org/C41008148","wikidata":"https://www.wikidata.org/wiki/Q21198","display_name":"Computer science","level":0,"score":0.6141999959945679},{"id":"https://openalex.org/C154945302","wikidata":"https://www.wikidata.org/wiki/Q11660","display_name":"Artificial intelligence","level":1,"score":0.5174000263214111},{"id":"https://openalex.org/C2775924081","wikidata":"https://www.wikidata.org/wiki/Q55608371","display_name":"Control (management)","level":2,"score":0.4952999949455261},{"id":"https://openalex.org/C204321447","wikidata":"https://www.wikidata.org/wiki/Q30642","display_name":"Natural language processing","level":1,"score":0.36959999799728394},{"id":"https://openalex.org/C2776401178","wikidata":"https://www.wikidata.org/wiki/Q12050496","display_name":"Feature (linguistics)","level":2,"score":0.3659000098705292},{"id":"https://openalex.org/C2781067378","wikidata":"https://www.wikidata.org/wiki/Q17027399","display_name":"Interpretability","level":2,"score":0.3361999988555908},{"id":"https://openalex.org/C12713177","wikidata":"https://www.wikidata.org/wiki/Q1900281","display_name":"Perspective (graphical)","level":2,"score":0.3118000030517578},{"id":"https://openalex.org/C2777212361","wikidata":"https://www.wikidata.org/wiki/Q5127848","display_name":"Class (philosophy)","level":2,"score":0.2953999936580658},{"id":"https://openalex.org/C116834253","wikidata":"https://www.wikidata.org/wiki/Q2039217","display_name":"Identification (biology)","level":2,"score":0.28630000352859497},{"id":"https://openalex.org/C137293760","wikidata":"https://www.wikidata.org/wiki/Q3621696","display_name":"Language model","level":2,"score":0.2849999964237213},{"id":"https://openalex.org/C36503486","wikidata":"https://www.wikidata.org/wiki/Q11235244","display_name":"Domain (mathematical analysis)","level":2,"score":0.2840999960899353},{"id":"https://openalex.org/C31972630","wikidata":"https://www.wikidata.org/wiki/Q844240","display_name":"Computer vision","level":1,"score":0.27799999713897705},{"id":"https://openalex.org/C26517878","wikidata":"https://www.wikidata.org/wiki/Q228039","display_name":"Key (lock)","level":2,"score":0.2685999870300293},{"id":"https://openalex.org/C2780791683","wikidata":"https://www.wikidata.org/wiki/Q846785","display_name":"Action (physics)","level":2,"score":0.2644999921321869},{"id":"https://openalex.org/C17500928","wikidata":"https://www.wikidata.org/wiki/Q959968","display_name":"Control system","level":2,"score":0.26339998841285706},{"id":"https://openalex.org/C184337299","wikidata":"https://www.wikidata.org/wiki/Q1437428","display_name":"Semantics (computer science)","level":2,"score":0.25769999623298645},{"id":"https://openalex.org/C2779530757","wikidata":"https://www.wikidata.org/wiki/Q1207505","display_name":"Quality (philosophy)","level":2,"score":0.25540000200271606}],"mesh":[],"locations_count":1,"locations":[{"id":"doi:10.18653/v1/2026.acl-long.1380","doi":"https://doi.org/10.18653/v1/2026.acl-long.1380","is_oa":true,"landing_page_url":"https://doi.org/10.18653/v1/2026.acl-long.1380","pdf_url":"https://aclanthology.org/2026.acl-long.1380.pdf","source":{"id":"https://openalex.org/S4393915740","display_name":"Proceedings of the conference - Association for Computational Linguistics. Meeting","issn_l":"0736-587X","issn":["0736-587X"],"is_oa":false,"is_in_doaj":false,"is_core":false,"listed_in":["norway-1"],"host_organization":null,"host_organization_name":null,"host_organization_lineage":[],"host_organization_lineage_names":[],"type":"conference"},"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 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)","raw_type":"proceedings-article"}],"best_oa_location":{"id":"doi:10.18653/v1/2026.acl-long.1380","doi":"https://doi.org/10.18653/v1/2026.acl-long.1380","is_oa":true,"landing_page_url":"https://doi.org/10.18653/v1/2026.acl-long.1380","pdf_url":"https://aclanthology.org/2026.acl-long.1380.pdf","source":{"id":"https://openalex.org/S4393915740","display_name":"Proceedings of the conference - Association for Computational Linguistics. Meeting","issn_l":"0736-587X","issn":["0736-587X"],"is_oa":false,"is_in_doaj":false,"is_core":false,"listed_in":["norway-1"],"host_organization":null,"host_organization_name":null,"host_organization_lineage":[],"host_organization_lineage_names":[],"type":"conference"},"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 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)","raw_type":"proceedings-article"},"sustainable_development_goals":[],"study_designs":[],"awards":[{"id":"https://openalex.org/G2257911995","display_name":null,"funder_award_id":"RS-2020-II201361","funder_id":"https://openalex.org/F4320328359","funder_display_name":"Ministry of Science and ICT, South Korea"},{"id":"https://openalex.org/G3182541678","display_name":null,"funder_award_id":"RS-2020-II201361","funder_id":"https://openalex.org/F4320321314","funder_display_name":"Yonsei University"},{"id":"https://openalex.org/G3212097187","display_name":null,"funder_award_id":"RS-2020-II201361","funder_id":"https://openalex.org/F4320335489","funder_display_name":"Institute for Information and Communications Technology Promotion"},{"id":"https://openalex.org/G3369352815","display_name":null,"funder_award_id":"RS-2025-02215344","funder_id":"https://openalex.org/F4320335489","funder_display_name":"Institute for Information and Communications Technology Promotion"}],"funders":[{"id":"https://openalex.org/F4320321314","display_name":"Yonsei University","ror":"https://ror.org/01wjejq96"},{"id":"https://openalex.org/F4320328359","display_name":"Ministry of Science and ICT, South Korea","ror":"https://ror.org/01wpjm123"},{"id":"https://openalex.org/F4320335489","display_name":"Institute for Information and Communications Technology Promotion","ror":"https://ror.org/01g0hqq23"},{"id":"https://openalex.org/F4320337041","display_name":"Department of Artificial Intelligence, Korea University","ror":null}],"has_content":{"grobid_xml":true,"pdf":true},"content_urls":{"pdf":"https://content.openalex.org/works/W7166809691.pdf","grobid_xml":"https://content.openalex.org/works/W7166809691.grobid-xml"},"referenced_works_count":0,"referenced_works":[],"related_works":["https://openalex.org/W7092295448","https://openalex.org/W4392271820","https://openalex.org/W7161099702","https://openalex.org/W4417046140","https://openalex.org/W7126058550","https://openalex.org/W4417129507","https://openalex.org/W4414586248","https://openalex.org/W4415035773","https://openalex.org/W7164883594","https://openalex.org/W4403323964"],"abstract_inverted_index":{"Large":[0],"Language":[1],"Models":[2],"(LLMs)":[3],"exhibit":[4],"highly":[5],"anisotropic":[6],"internal":[7],"representations,":[8],"often":[9],"characterized":[10],"by":[11],"massive":[12],"activations,":[13],"a":[14,17,44,61,70,134],"phenomenon":[15],"where":[16],"small":[18],"subset":[19],"of":[20,145,152,190,207],"feature":[21],"dimensions":[22,35,48,77],"possesses":[23],"magnitudes":[24],"significantly":[25],"larger":[26],"than":[27],"the":[28,101,146,156,160,170,205],"rest.While":[29],"prior":[30],"works":[31],"view":[32],"these":[33,47],"extreme":[34],"primarily":[36],"as":[37,50,79],"artifacts":[38],"to":[39,65,100,204],"be":[40],"managed,":[41],"we":[42,59,90],"propose":[43,60],"distinct":[45],"perspective:":[46],"serve":[49],"intrinsic":[51],"interpretable":[52,80],"functional":[53],"units":[54],"arising":[55],"from":[56],"domain":[57,114],"specialization.Specifically,":[58],"simple":[62],"magnitude-based":[63],"criterion":[64],"identify":[66],"Domain-Critical":[67],"Dimensions":[68],"in":[69,113,143,202],"trainingfree":[71],"manner.Our":[72],"analyses":[73],"reveal":[74],"that":[75,106,177],"such":[76],"behave":[78],"semantic":[81],"detectors":[82],"for":[83],"symbolic/quantitative":[84],"patterns":[85],"or":[86],"domain-specific":[87],"terms.In":[88],"addition,":[89],"introduce":[91],"Critical":[92],"Dimension":[93],"Steering,":[94],"which":[95,150],"applies":[96],"activation":[97],"steering":[98,112],"exclusively":[99],"identified":[102],"dimensions.Empirical":[103],"results":[104],"show":[105],"this":[107],"approach":[108],"outperforms":[109],"conventional":[110],"wholedimension":[111],"adaptation":[115],"and":[116,159,186,226],"jailbreaking":[117],"scenarios.1":[118],"The":[119,164,247],"following":[120,147],"are":[121,188],"multiple":[122],"choice":[123],"questions":[124],"(with":[125],"answers)":[126],"about":[127],"high":[128],"school":[129],"biology.You":[130],"must":[131],"respond":[132],"with":[133],"single":[135],"alphabet":[136],"character.Energy":[137],"is":[138,167,175,179,219],"harvested":[139],"during":[140,169,214],"cellular":[141,153],"respiration":[142,154],"stages.Which":[144],"correctly":[148],"states":[149],"phase":[151],"harvests":[155],"most":[157,165],"energy":[158,166],"correct":[161],"explanation":[162],"why?A.":[163],"released":[168],"Krebs":[171],"cycle":[172],"because":[173,201,217],"it":[174],"here":[176],"pyruvate":[178],"completely":[180,220],"...":[181,183,197,199,211,213],"B.":[182],"both":[184],"FADH2":[185],"NADH":[187],"produced.Each":[189],"those":[191],"molecules":[192],"will":[193],"release":[194],"2":[195],"ATPs":[196],"C.":[198],"phosphorylation":[200,206,216],"addition":[203],"ADP":[208],"into":[209,223],"ATP":[210],"D.":[212],"oxidative":[215],"H2O":[218],"broken":[221],"down":[222],"H":[224,245],"+":[225],"O2.Answer:":[227],"334":[228],"(Top":[229,232,235],"1)":[230],"1068":[231],"4)":[233],"2106":[234],"7)":[236],"Act.Token":[237,238,239],"2048.0":[240],"<bos>":[241],"65.0":[242],"_the":[243,249,273,279],"43.8":[244],"324.0":[246],"64.0":[248],"43.0":[250],"_ATP":[251,257,275],"147.0":[252],"_biology":[253],"62.0":[254],"_is":[255],"42.2":[256],"139.0":[258],"Energy":[259],"61.0":[260],"_to":[261,267],"41.5":[262],"_respiration":[263,295],"138.0":[264],"_Krebs":[265],"59.5":[266],"41.2":[268,274,280],"_energy":[269],"137.0":[270],"_cellular":[271,289],"58.8":[272],"132.0":[276],"_school":[277],"57.5":[278],"_phosphorylation":[281,299],"130.0":[282,288,294],"_harvested":[283],"55.8":[284],"_that":[285],"40.8":[286],"_ADP":[287],"55.0":[290],"_The":[291],"38.5":[292],"_NAD":[293],"54.8":[296],"_in":[297],"38.2":[298]},"counts_by_year":[],"updated_date":"2026-10-30T12:39:25.321706","created_date":"2026-07-02T00:00:00"}