{"id":"https://openalex.org/W7131410766","doi":"https://doi.org/10.1109/ccwc67433.2026.11393646","title":"Interpreting Negation in GPT-2: Layer-and Head-Level Causal Analysis","display_name":"Interpreting Negation in GPT-2: Layer-and Head-Level Causal Analysis","publication_year":2026,"publication_date":"2026-01-05","ids":{"openalex":"https://openalex.org/W7131410766","doi":"https://doi.org/10.1109/ccwc67433.2026.11393646"},"language":null,"primary_location":{"id":"doi:10.1109/ccwc67433.2026.11393646","is_oa":false,"landing_page_url":"https://doi.org/10.1109/ccwc67433.2026.11393646","pdf_url":null,"source":null,"license":null,"license_id":null,"version":"publishedVersion","is_accepted":true,"is_published":true,"raw_source_name":"2026 IEEE 16th Annual Computing and Communication Workshop and Conference (CCWC)","raw_type":"proceedings-article"},"type":"conference-paper","indexed_in":["arxiv","crossref"],"open_access":{"is_oa":true,"oa_status":"green","oa_url":"https://arxiv.org/pdf/2603.12423","any_repository_has_fulltext":true},"authorships":[{"author_position":"first","author":{"id":"https://openalex.org/A5093917709","display_name":"Abdullah Al Mofael","orcid":null},"institutions":[{"id":"https://openalex.org/I197173980","display_name":"Southeastern Louisiana University","ror":"https://ror.org/0531xck41","country_code":"US","type":"education","lineage":["https://openalex.org/I197173980","https://openalex.org/I2799628689"]}],"countries":["US"],"is_corresponding":false,"raw_author_name":"Abdullah Al Mofael","raw_affiliation_strings":["Southeastern Louisiana University,*Department of Computer Science,Hammond,LA,USA"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"Southeastern Louisiana University,*Department of Computer Science,Hammond,LA,USA","institution_ids":["https://openalex.org/I197173980"]}]},{"author_position":"middle","author":{"id":"https://openalex.org/A5034023337","display_name":"Lisa M. Kuhn","orcid":null},"institutions":[{"id":"https://openalex.org/I197173980","display_name":"Southeastern Louisiana University","ror":"https://ror.org/0531xck41","country_code":"US","type":"education","lineage":["https://openalex.org/I197173980","https://openalex.org/I2799628689"]}],"countries":["US"],"is_corresponding":false,"raw_author_name":"Lisa M. Kuhn","raw_affiliation_strings":["Southeastern Louisiana University,Department of Mathematics,Hammond,LA,USA"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"Southeastern Louisiana University,Department of Mathematics,Hammond,LA,USA","institution_ids":["https://openalex.org/I197173980"]}]},{"author_position":"middle","author":{"id":"https://openalex.org/A5000302669","display_name":"Ghassan Alkadi","orcid":null},"institutions":[{"id":"https://openalex.org/I197173980","display_name":"Southeastern Louisiana University","ror":"https://ror.org/0531xck41","country_code":"US","type":"education","lineage":["https://openalex.org/I197173980","https://openalex.org/I2799628689"]}],"countries":["US"],"is_corresponding":false,"raw_author_name":"Ghassan Alkadi","raw_affiliation_strings":["Southeastern Louisiana University,*Department of Computer Science,Hammond,LA,USA"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"Southeastern Louisiana University,*Department of Computer Science,Hammond,LA,USA","institution_ids":["https://openalex.org/I197173980"]}]},{"author_position":"last","author":{"id":"https://openalex.org/A5111403097","display_name":"Kuo-pao Yang","orcid":null},"institutions":[{"id":"https://openalex.org/I197173980","display_name":"Southeastern Louisiana University","ror":"https://ror.org/0531xck41","country_code":"US","type":"education","lineage":["https://openalex.org/I197173980","https://openalex.org/I2799628689"]}],"countries":["US"],"is_corresponding":false,"raw_author_name":"Kuo-Pao Yang","raw_affiliation_strings":["Southeastern Louisiana University,*Department of Computer Science,Hammond,LA,USA"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"Southeastern Louisiana University,*Department of Computer Science,Hammond,LA,USA","institution_ids":["https://openalex.org/I197173980"]}]}],"institutions":[],"countries_distinct_count":1,"institutions_distinct_count":1,"corresponding_author_ids":[],"corresponding_institution_ids":["https://openalex.org/I197173980"],"apc_list":null,"apc_paid":null,"fwci":0.0,"has_fulltext":false,"cited_by_count":0,"citation_normalized_percentile":{"value":0.24001493,"is_in_top_1_percent":false,"is_in_top_10_percent":false},"cited_by_percentile_year":null,"biblio":{"volume":null,"issue":null,"first_page":"0042","last_page":"0050"},"is_retracted":false,"is_paratext":false,"is_xpac":false,"primary_topic":{"id":"https://openalex.org/T10142","display_name":"Formal Methods in Verification","score":0.14069999754428864,"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":0.14069999754428864,"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/T10181","display_name":"Natural Language Processing Techniques","score":0.07590000331401825,"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/T12072","display_name":"Machine Learning and Algorithms","score":0.05829999968409538,"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/negation","display_name":"Negation","score":0.9646999835968018},{"id":"https://openalex.org/keywords/meaning","display_name":"Meaning (existential)","score":0.5512999892234802},{"id":"https://openalex.org/keywords/polarity","display_name":"Polarity (international relations)","score":0.40049999952316284},{"id":"https://openalex.org/keywords/causal-chain","display_name":"Causal chain","score":0.38440001010894775},{"id":"https://openalex.org/keywords/key","display_name":"Key (lock)","score":0.32350000739097595}],"concepts":[{"id":"https://openalex.org/C2185349","wikidata":"https://www.wikidata.org/wiki/Q190558","display_name":"Negation","level":2,"score":0.9646999835968018},{"id":"https://openalex.org/C2780876879","wikidata":"https://www.wikidata.org/wiki/Q3054749","display_name":"Meaning (existential)","level":2,"score":0.5512999892234802},{"id":"https://openalex.org/C41895202","wikidata":"https://www.wikidata.org/wiki/Q8162","display_name":"Linguistics","level":1,"score":0.5468999743461609},{"id":"https://openalex.org/C15744967","wikidata":"https://www.wikidata.org/wiki/Q9418","display_name":"Psychology","level":0,"score":0.48080000281333923},{"id":"https://openalex.org/C2777361361","wikidata":"https://www.wikidata.org/wiki/Q1112585","display_name":"Polarity (international relations)","level":3,"score":0.40049999952316284},{"id":"https://openalex.org/C79897977","wikidata":"https://www.wikidata.org/wiki/Q5054568","display_name":"Causal chain","level":2,"score":0.38440001010894775},{"id":"https://openalex.org/C180747234","wikidata":"https://www.wikidata.org/wiki/Q23373","display_name":"Cognitive psychology","level":1,"score":0.36629998683929443},{"id":"https://openalex.org/C26517878","wikidata":"https://www.wikidata.org/wiki/Q228039","display_name":"Key (lock)","level":2,"score":0.32350000739097595},{"id":"https://openalex.org/C41008148","wikidata":"https://www.wikidata.org/wiki/Q21198","display_name":"Computer science","level":0,"score":0.3005000054836273},{"id":"https://openalex.org/C77805123","wikidata":"https://www.wikidata.org/wiki/Q161272","display_name":"Social psychology","level":1,"score":0.2849999964237213},{"id":"https://openalex.org/C2987525970","wikidata":"https://www.wikidata.org/wiki/Q96374569","display_name":"Causal analysis","level":2,"score":0.2827000021934509},{"id":"https://openalex.org/C204321447","wikidata":"https://www.wikidata.org/wiki/Q30642","display_name":"Natural language processing","level":1,"score":0.28139999508857727},{"id":"https://openalex.org/C12725497","wikidata":"https://www.wikidata.org/wiki/Q810247","display_name":"Baseline (sea)","level":2,"score":0.27000001072883606}],"mesh":[],"locations_count":2,"locations":[{"id":"doi:10.1109/ccwc67433.2026.11393646","is_oa":false,"landing_page_url":"https://doi.org/10.1109/ccwc67433.2026.11393646","pdf_url":null,"source":null,"license":null,"license_id":null,"version":"publishedVersion","is_accepted":true,"is_published":true,"raw_source_name":"2026 IEEE 16th Annual Computing and Communication Workshop and Conference (CCWC)","raw_type":"proceedings-article"},{"id":"pmh:oai:arXiv.org:2603.12423","is_oa":true,"landing_page_url":"http://arxiv.org/abs/2603.12423","pdf_url":"https://arxiv.org/pdf/2603.12423","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,"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":null,"license_id":null,"version":"submittedVersion","is_accepted":false,"is_published":false,"raw_source_name":null,"raw_type":"text"}],"best_oa_location":{"id":"pmh:oai:arXiv.org:2603.12423","is_oa":true,"landing_page_url":"http://arxiv.org/abs/2603.12423","pdf_url":"https://arxiv.org/pdf/2603.12423","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,"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":null,"license_id":null,"version":"submittedVersion","is_accepted":false,"is_published":false,"raw_source_name":null,"raw_type":"text"},"sustainable_development_goals":[],"awards":[],"funders":[],"has_content":{"grobid_xml":false,"pdf":false},"content_urls":null,"referenced_works_count":5,"referenced_works":["https://openalex.org/W2123402141","https://openalex.org/W3034995113","https://openalex.org/W4205460703","https://openalex.org/W4249875616","https://openalex.org/W4385718075"],"related_works":[],"abstract_inverted_index":{"Negation":[0,77],"remains":[1],"a":[2,21,50,74,165,269,278],"persistent":[3],"challenge":[4],"for":[5,274,281],"modern":[6],"language":[7],"models,":[8],"often":[9],"causing":[10],"reversed":[11],"meanings":[12],"or":[13],"factual":[14],"errors.":[15],"In":[16,104,123],"this":[17,70,154,283],"work,":[18],"we":[19,72],"conduct":[20],"causal":[22,102,236],"analysis":[23,46],"of":[24,54,66,168,268],"how":[25,120,133,141,282],"GPT-2":[26],"Small":[27],"internally":[28],"processes":[29],"such":[30],"linguistic":[31,62],"transformations.":[32],"We":[33,94],"examine":[34],"its":[35],"hidden":[36],"representations":[37],"at":[38],"both":[39],"the":[40,76,83,184,255],"layer":[41],"and":[42,57,64,91,145,198,225,249,251,271],"head":[43],"level.":[44],"Our":[45,150],"is":[47,156,161],"based":[48],"on":[49,188,254],"self-curated":[51],"12,OOO-pair":[52],"dataset":[53],"matched":[55],"affirmative":[56,89,110,201,212],"negated":[58,116],"sentences,":[59],"covering":[60],"multiple":[61],"templates":[63],"forms":[65,243],"negation.":[67],"To":[68],"quantify":[69],"behavior,":[71],"define":[73],"metric,":[75],"Effect":[78],"Score":[79],"(NES),":[80],"which":[81],"measures":[82],"model's":[84,185],"sensitivity":[85],"in":[86],"distinguishing":[87],"between":[88],"statements":[90],"their":[92,115],"negations.":[93],"carried":[95],"out":[96],"two":[97],"key":[98],"interventions":[99],"to":[100,118,131,176],"probe":[101],"structure.":[103],"activation":[105],"patching,":[106],"internal":[107],"activations":[108,202],"from":[109],"sentences":[111],"were":[112,128],"inserted":[113],"into":[114],"counterparts":[117],"see":[119],"meaning":[121],"shifted.":[122],"ablation,":[124],"specific":[125,180],"attention":[126,170],"heads":[127,210],"temporarily":[129],"disabled":[130],"observe":[132],"logical":[134,275],"polarity":[135],"changed.":[136],"Together,":[137],"these":[138,179,209,235],"steps":[139],"revealed":[140],"negation":[142,186,196,242],"signals":[143],"move":[144],"evolve":[146],"through":[147],"GPT-2's":[148],"layers.":[149],"findings":[151,264],"indicate":[152],"that":[153,208,234],"capability":[155],"not":[157],"widespread;":[158],"instead,":[159],"it":[160],"highly":[162],"concentrated":[163],"within":[164,173],"limited":[166],"number":[167],"mid-layer":[169],"heads,":[171],"primarily":[172],"layers":[174],"4":[175],"6.":[177],"Ablating":[178],"components":[181],"directly":[182],"disrupts":[183],"sensitivity:":[187],"our":[189],"in-domain,":[190],"ablation":[191,221],"increased":[192,204],"NES":[193,205,224],"(indicating":[194],"weaker":[195],"sensitivity),":[197],"re-introducing":[199],"cached":[200],"(rescue)":[203],"further,":[206],"confirming":[207],"carry":[211],"signal":[213],"rather":[214],"than":[215],"restoring":[216],"baseline":[217],"behavior.":[218],"On":[219],"xNot360,":[220],"slightly":[222],"decreased":[223],"rescue":[226],"restored":[227],"performance":[228],"above":[229],"baseline.":[230],"This":[231],"pattern":[232],"demonstrates":[233],"patterns":[237],"are":[238],"consistent":[239],"across":[240],"various":[241],"(\u2018never,\u2019":[244],"\u2018does":[245],"not,\u2019":[246],"\u2018doesn't,\u2019":[247],"\u2018cannot,\u2019":[248],"\u2018can\u2018t\u2019)":[250],"remain":[252],"detectable":[253],"external":[256],"xNot360":[257],"benchmark,":[258],"though":[259],"with":[260],"smaller":[261],"magnitude.":[262],"These":[263],"offer":[265],"direct":[266],"evidence":[267],"localized":[270],"interpretable":[272],"circuit":[273],"polarity,":[276],"providing":[277],"mechanistic":[279],"explanation":[280],"transformer":[284],"architecture":[285],"encodes":[286],"meaning.":[287]},"counts_by_year":[],"updated_date":"2026-08-05T07:39:15.569665","created_date":"2026-02-26T00:00:00"}