{"id":"https://openalex.org/W4387559252","doi":"https://doi.org/10.48550/arxiv.2310.04625","title":"Copy Suppression: Comprehensively Understanding an Attention Head","display_name":"Copy Suppression: Comprehensively Understanding an Attention Head","publication_year":2023,"publication_date":"2023-10-06","ids":{"openalex":"https://openalex.org/W4387559252","doi":"https://doi.org/10.48550/arxiv.2310.04625"},"language":"en","primary_location":{"id":"pmh:oai:arXiv.org:2310.04625","doi":null,"is_oa":true,"landing_page_url":"http://arxiv.org/abs/2310.04625","pdf_url":"https://arxiv.org/pdf/2310.04625","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/2310.04625","any_repository_has_fulltext":true},"authorships":[{"author_position":"first","author":{"id":"https://openalex.org/A5108869919","display_name":"Callum Stuart McDougall","orcid":null,"observed_orcids":[]},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"McDougall, Callum","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5071070679","display_name":"Arthur Conmy","orcid":null,"observed_orcids":[]},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Conmy, Arthur","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5093047611","display_name":"Cody Rushing","orcid":null,"observed_orcids":[]},"institutions":[{"id":"https://openalex.org/I86519309","display_name":"The University of Texas at Austin","ror":"https://ror.org/00hj54h04","country_code":"US","type":"education","lineage":["https://openalex.org/I86519309"]}],"countries":["US"],"is_corresponding":false,"raw_author_name":"Rushing, Cody","raw_affiliation_strings":["University of Texas at Austin "],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"University of Texas at Austin ","institution_ids":["https://openalex.org/I86519309"]}]},{"author_position":"middle","author":{"id":null,"display_name":"McGrath, Thomas","orcid":null,"observed_orcids":[]},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"McGrath, Thomas","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"last","author":{"id":"https://openalex.org/A5044489746","display_name":"Neel Nanda","orcid":null,"observed_orcids":[]},"institutions":[{"id":"https://openalex.org/I4210090411","display_name":"Google DeepMind (United Kingdom)","ror":"https://ror.org/00971b260","country_code":"GB","type":"company","lineage":["https://openalex.org/I4210090411","https://openalex.org/I4210128969"]}],"countries":["GB"],"is_corresponding":false,"raw_author_name":"Nanda, Neel","raw_affiliation_strings":["Google DeepMind "],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"Google DeepMind ","institution_ids":["https://openalex.org/I4210090411"]}]}],"institutions":[],"countries_distinct_count":2,"institutions_distinct_count":2,"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":null,"last_page":null},"is_retracted":false,"is_paratext":false,"is_xpac":false,"primary_topic":{"id":"https://openalex.org/T10028","display_name":"Topic Modeling","score":0.5551999807357788,"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/T10028","display_name":"Topic Modeling","score":0.5551999807357788,"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/T12026","display_name":"Explainable Artificial Intelligence (XAI)","score":0.37059998512268066,"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/T10181","display_name":"Natural Language Processing Techniques","score":0.030500000342726707,"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/attention-heads","display_name":"attention heads","score":0.8809999823570251},{"id":"https://openalex.org/keywords/model-calibration","display_name":"model calibration","score":0.7459999918937683},{"id":"https://openalex.org/keywords/self-healing","display_name":"self-healing","score":0.8920000195503235},{"id":"https://openalex.org/keywords/ablation","display_name":"ablation","score":0.5070000290870667},{"id":"https://openalex.org/keywords/language-model-interpretability","display_name":"language model interpretability","score":0.4740000069141388},{"id":"https://openalex.org/keywords/overconfidence","display_name":"overconfidence","score":0.4359999895095825}],"concepts":[{"id":"https://openalex.org/C48145219","wikidata":"https://www.wikidata.org/wiki/Q1335365","display_name":"Security token","level":2,"score":0.6884354948997498},{"id":"https://openalex.org/C41008148","wikidata":"https://www.wikidata.org/wiki/Q21198","display_name":"Computer science","level":0,"score":0.679397463798523},{"id":"https://openalex.org/C2779151265","wikidata":"https://www.wikidata.org/wiki/Q1156791","display_name":"Copying","level":2,"score":0.6629629135131836},{"id":"https://openalex.org/C2779343474","wikidata":"https://www.wikidata.org/wiki/Q3109175","display_name":"Context (archaeology)","level":2,"score":0.5949440598487854},{"id":"https://openalex.org/C154945302","wikidata":"https://www.wikidata.org/wiki/Q11660","display_name":"Artificial intelligence","level":1,"score":0.3693963885307312},{"id":"https://openalex.org/C180747234","wikidata":"https://www.wikidata.org/wiki/Q23373","display_name":"Cognitive psychology","level":1,"score":0.35504430532455444},{"id":"https://openalex.org/C169760540","wikidata":"https://www.wikidata.org/wiki/Q207011","display_name":"Neuroscience","level":1,"score":0.3370070457458496},{"id":"https://openalex.org/C15744967","wikidata":"https://www.wikidata.org/wiki/Q9418","display_name":"Psychology","level":0,"score":0.29997360706329346},{"id":"https://openalex.org/C86803240","wikidata":"https://www.wikidata.org/wiki/Q420","display_name":"Biology","level":0,"score":0.18129804730415344},{"id":"https://openalex.org/C54355233","wikidata":"https://www.wikidata.org/wiki/Q7162","display_name":"Genetics","level":1,"score":0.09228900074958801},{"id":"https://openalex.org/C38652104","wikidata":"https://www.wikidata.org/wiki/Q3510521","display_name":"Computer security","level":1,"score":0.08213165402412415},{"id":"https://openalex.org/C151730666","wikidata":"https://www.wikidata.org/wiki/Q7205","display_name":"Paleontology","level":1,"score":0.0}],"mesh":[],"locations_count":2,"locations":[{"id":"pmh:oai:arXiv.org:2310.04625","doi":null,"is_oa":true,"landing_page_url":"http://arxiv.org/abs/2310.04625","pdf_url":"https://arxiv.org/pdf/2310.04625","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.2310.04625","doi":"https://doi.org/10.48550/arxiv.2310.04625","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2310.04625","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:2310.04625","doi":null,"is_oa":true,"landing_page_url":"http://arxiv.org/abs/2310.04625","pdf_url":"https://arxiv.org/pdf/2310.04625","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":[],"has_content":{"grobid_xml":true,"pdf":true},"content_urls":{"pdf":"https://content.openalex.org/works/W4387559252.pdf","grobid_xml":"https://content.openalex.org/works/W4387559252.grobid-xml"},"referenced_works_count":0,"referenced_works":[],"related_works":["https://openalex.org/W4404783649","https://openalex.org/W7220702684","https://openalex.org/W7216323833","https://openalex.org/W4392182427","https://openalex.org/W7162215185","https://openalex.org/W7219155387","https://openalex.org/W7171323353","https://openalex.org/W4414968658","https://openalex.org/W4385473442","https://openalex.org/W7172501068"],"abstract_inverted_index":{"We":[0,77,178],"present":[1],"a":[2,25,123,126,200],"single":[3],"attention":[4],"head":[5,37],"in":[6,21,33,103,125,140,151,199],"GPT-2":[7,104],"Small":[8],"that":[9,71,81,180],"has":[10],"one":[11,187],"main":[12],"role":[13,121,139],"across":[14],"the":[15,34,36,74,79,82,99,107,114,119,197,206],"entire":[16],"training":[17],"distribution.":[18],"If":[19],"components":[20,149],"earlier":[22,32],"layers":[23],"predict":[24],"certain":[26,65],"token,":[27],"and":[28,92],"this":[29,42,112,158],"token":[30],"appears":[31],"context,":[35],"suppresses":[38,49],"it:":[39],"we":[40],"call":[41],"copy":[43,87,135,191,207],"suppression.":[44],"Attention":[45],"Head":[46],"10.7":[47],"(L10H7)":[48],"naive":[50],"copying":[51],"behavior":[52,198],"which":[53,189,193],"improves":[54],"overall":[55],"model":[56,128,148],"calibration.":[57],"This":[58],"explains":[59,194],"why":[60],"multiple":[61],"prior":[62],"works":[63],"studying":[64],"narrow":[66,201],"tasks":[67],"found":[68],"negative":[69],"heads":[70],"systematically":[72],"favored":[73],"wrong":[75],"answer.":[76],"uncover":[78],"mechanism":[80],"Negative":[83],"Heads":[84],"use":[85],"for":[86,157],"suppression":[88,136,161,208],"with":[89],"weights-based":[90],"evidence":[91],"are":[93],"able":[94],"to":[95,129,144,163,176],"explain":[96],"76.9%":[97],"of":[98,101,109,118,122,134,188,196,205],"impact":[100],"L10H7":[102],"Small.":[105],"To":[106],"best":[108],"our":[110,214],"knowledge,":[111],"is":[113,137,170,174,182,190],"most":[115],"comprehensive":[116],"description":[117],"complete":[120],"component":[124],"language":[127],"date.":[130],"One":[131],"major":[132],"effect":[133],"its":[138],"self-repair.":[141],"Self-repair":[142],"refers":[143],"how":[145],"ablating":[146],"crucial":[147],"results":[150],"downstream":[152],"neural":[153],"network":[154],"parts":[155],"compensating":[156],"ablation.":[159],"Copy":[160],"leads":[162],"self-repair:":[164],"if":[165],"an":[166],"initial":[167],"overconfident":[168],"copier":[169],"ablated,":[171],"then":[172],"there":[173],"nothing":[175],"suppress.":[177],"show":[179],"self-repair":[181],"implemented":[183],"by":[184],"several":[185],"mechanisms,":[186],"suppression,":[192],"39%":[195],"task.":[202],"Interactive":[203],"visualisations":[204],"phenomena":[209],"may":[210],"be":[211],"seen":[212],"at":[213],"web":[215],"app":[216],"https://copy-suppression.streamlit.app/":[217]},"counts_by_year":[{"year":2025,"cited_by_count":1},{"year":2024,"cited_by_count":1}],"updated_date":"2026-09-30T09:27:27.532063","created_date":"2023-10-12T00:00:00"}