{"id":"https://openalex.org/W4308023630","doi":"https://doi.org/10.48550/arxiv.2211.00593","title":"Interpretability in the Wild: a Circuit for Indirect Object Identification in GPT-2 small","display_name":"Interpretability in the Wild: a Circuit for Indirect Object Identification in GPT-2 small","publication_year":2022,"publication_date":"2022-11-01","ids":{"openalex":"https://openalex.org/W4308023630","doi":"https://doi.org/10.48550/arxiv.2211.00593"},"language":"en","primary_location":{"id":"pmh:oai:arXiv.org:2211.00593","doi":null,"is_oa":true,"landing_page_url":"http://arxiv.org/abs/2211.00593","pdf_url":"https://arxiv.org/pdf/2211.00593","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":null,"license_id":null,"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/2211.00593","any_repository_has_fulltext":true},"authorships":[{"author_position":"first","author":{"id":"https://openalex.org/A5100413965","display_name":"Kevin Wang","orcid":"https://orcid.org/0009-0003-4227-7745","observed_orcids":["https://orcid.org/0009-0003-4227-7745"]},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Wang, Kevin","raw_affiliation_strings":["Redwood Research"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"Redwood Research","institution_ids":[]}]},{"author_position":"middle","author":{"id":"https://openalex.org/A5017761207","display_name":"Alexandre Variengien","orcid":null,"observed_orcids":[]},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Variengien, Alexandre","raw_affiliation_strings":["Redwood Research"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"Redwood Research","institution_ids":[]}]},{"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":["Redwood Research"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"Redwood Research","institution_ids":[]}]},{"author_position":"middle","author":{"id":"https://openalex.org/A5050408969","display_name":"Buck Shlegeris","orcid":null,"observed_orcids":[]},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Shlegeris, Buck","raw_affiliation_strings":["Redwood Research"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"Redwood Research","institution_ids":[]}]},{"author_position":"last","author":{"id":"https://openalex.org/A5060196069","display_name":"Jacob Steinhardt","orcid":"https://orcid.org/0000-0002-0257-3860","observed_orcids":["https://orcid.org/0000-0002-0257-3860"]},"institutions":[{"id":"https://openalex.org/I95457486","display_name":"University of California, Berkeley","ror":"https://ror.org/01an7q238","country_code":"US","type":"education","lineage":["https://openalex.org/I95457486"]}],"countries":["US"],"is_corresponding":false,"raw_author_name":"Steinhardt, Jacob","raw_affiliation_strings":["Redwood Research","UC Berkeley"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"Redwood Research","institution_ids":[]},{"raw_affiliation_string":"UC Berkeley","institution_ids":["https://openalex.org/I95457486"]}]}],"institutions":[],"countries_distinct_count":1,"institutions_distinct_count":1,"corresponding_author_ids":[],"corresponding_institution_ids":[],"apc_list":null,"apc_paid":null,"fwci":null,"has_fulltext":true,"cited_by_count":55,"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/T12026","display_name":"Explainable Artificial Intelligence (XAI)","score":1.0,"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":1.0,"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":0.0,"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":0.0,"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/mechanistic-interpretability","display_name":"mechanistic interpretability","score":0.9470000267028809},{"id":"https://openalex.org/keywords/attention-heads","display_name":"attention heads","score":0.8220000267028809},{"id":"https://openalex.org/keywords/causal-intervention","display_name":"causal intervention","score":0.8759999871253967},{"id":"https://openalex.org/keywords/faithfulness","display_name":"faithfulness","score":0.625},{"id":"https://openalex.org/keywords/completeness","display_name":"completeness","score":0.7110000252723694},{"id":"https://openalex.org/keywords/minimality","display_name":"minimality","score":0.925000011920929},{"id":"https://openalex.org/keywords/natural-language-understanding","display_name":"natural language understanding","score":0.49900001287460327}],"concepts":[{"id":"https://openalex.org/C2781067378","wikidata":"https://www.wikidata.org/wiki/Q17027399","display_name":"Interpretability","level":2,"score":0.9731392860412598},{"id":"https://openalex.org/C116834253","wikidata":"https://www.wikidata.org/wiki/Q2039217","display_name":"Identification (biology)","level":2,"score":0.6559686064720154},{"id":"https://openalex.org/C41008148","wikidata":"https://www.wikidata.org/wiki/Q21198","display_name":"Computer science","level":0,"score":0.6384496688842773},{"id":"https://openalex.org/C2781238097","wikidata":"https://www.wikidata.org/wiki/Q175026","display_name":"Object (grammar)","level":2,"score":0.5242217779159546},{"id":"https://openalex.org/C100776233","wikidata":"https://www.wikidata.org/wiki/Q2532492","display_name":"Bridge (graph theory)","level":2,"score":0.5126435160636902},{"id":"https://openalex.org/C119857082","wikidata":"https://www.wikidata.org/wiki/Q2539","display_name":"Machine learning","level":1,"score":0.5031408667564392},{"id":"https://openalex.org/C154945302","wikidata":"https://www.wikidata.org/wiki/Q11660","display_name":"Artificial intelligence","level":1,"score":0.4966593384742737},{"id":"https://openalex.org/C2780586882","wikidata":"https://www.wikidata.org/wiki/Q7520643","display_name":"Simple (philosophy)","level":2,"score":0.44994157552719116},{"id":"https://openalex.org/C195324797","wikidata":"https://www.wikidata.org/wiki/Q33742","display_name":"Natural language","level":2,"score":0.41954505443573},{"id":"https://openalex.org/C111472728","wikidata":"https://www.wikidata.org/wiki/Q9471","display_name":"Epistemology","level":1,"score":0.13221478462219238},{"id":"https://openalex.org/C18903297","wikidata":"https://www.wikidata.org/wiki/Q7150","display_name":"Ecology","level":1,"score":0.08195847272872925},{"id":"https://openalex.org/C86803240","wikidata":"https://www.wikidata.org/wiki/Q420","display_name":"Biology","level":0,"score":0.0},{"id":"https://openalex.org/C71924100","wikidata":"https://www.wikidata.org/wiki/Q11190","display_name":"Medicine","level":0,"score":0.0},{"id":"https://openalex.org/C126322002","wikidata":"https://www.wikidata.org/wiki/Q11180","display_name":"Internal medicine","level":1,"score":0.0},{"id":"https://openalex.org/C138885662","wikidata":"https://www.wikidata.org/wiki/Q5891","display_name":"Philosophy","level":0,"score":0.0}],"mesh":[],"locations_count":2,"locations":[{"id":"pmh:oai:arXiv.org:2211.00593","doi":null,"is_oa":true,"landing_page_url":"http://arxiv.org/abs/2211.00593","pdf_url":"https://arxiv.org/pdf/2211.00593","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":null,"license_id":null,"version":"submittedVersion","is_accepted":false,"is_published":false,"raw_source_name":null,"raw_type":"text"},{"id":"doi:10.48550/arxiv.2211.00593","doi":"https://doi.org/10.48550/arxiv.2211.00593","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2211.00593","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:2211.00593","doi":null,"is_oa":true,"landing_page_url":"http://arxiv.org/abs/2211.00593","pdf_url":"https://arxiv.org/pdf/2211.00593","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":null,"license_id":null,"version":"submittedVersion","is_accepted":false,"is_published":false,"raw_source_name":null,"raw_type":"text"},"sustainable_development_goals":[],"study_designs":[],"awards":[],"funders":[],"has_content":{"pdf":false,"grobid_xml":false},"content_urls":null,"referenced_works_count":0,"referenced_works":[],"related_works":["https://openalex.org/W7135020037","https://openalex.org/W4404987111","https://openalex.org/W4396814526","https://openalex.org/W7125680013","https://openalex.org/W7077507337","https://openalex.org/W4387687067","https://openalex.org/W7166798189","https://openalex.org/W7162135667","https://openalex.org/W7166846670","https://openalex.org/W7163726677"],"abstract_inverted_index":{"Research":[0],"in":[1,12,27,34,107,137],"mechanistic":[2,146],"interpretability":[3,83],"seeks":[4],"to":[5,134,156,160],"explain":[6],"behaviors":[7,26,33],"of":[8,14,82,115,148],"machine":[9],"learning":[10],"models":[11,36,151,163],"terms":[13],"their":[15],"internal":[16],"components.":[17],"However,":[18],"most":[19],"previous":[20],"work":[21,141],"either":[22],"focuses":[23],"on":[24,86],"simple":[25],"small":[28,54],"models,":[29],"or":[30],"describes":[31],"complicated":[32],"larger":[35,162],"with":[37],"broad":[38],"strokes.":[39],"In":[40],"this":[41,45,92],"work,":[42],"we":[43,77],"bridge":[44],"gap":[46],"by":[47],"presenting":[48],"an":[49],"explanation":[50,66,117],"for":[51],"how":[52],"GPT-2":[53],"performs":[55],"a":[56,80,101,108,145],"natural":[57,102],"language":[58,109],"task":[59],"called":[60],"indirect":[61],"object":[62],"identification":[63],"(IOI).":[64],"Our":[65,140],"encompasses":[67],"26":[68],"attention":[69],"heads":[70],"grouped":[71],"into":[72],"7":[73],"main":[74],"classes,":[75],"which":[76],"discovered":[78],"using":[79,118],"combination":[81],"approaches":[84],"relying":[85],"causal":[87],"interventions.":[88],"To":[89],"our":[90,116,129,138,158],"knowledge,":[91],"investigation":[93],"is":[94,152],"the":[95,105,113],"largest":[96],"end-to-end":[97],"attempt":[98],"at":[99],"reverse-engineering":[100],"behavior":[103],"\"in":[104],"wild\"":[106],"model.":[110],"We":[111],"evaluate":[112],"reliability":[114],"three":[119],"quantitative":[120],"criteria--faithfulness,":[121],"completeness":[122],"and":[123,164],"minimality.":[124],"Though":[125],"these":[126],"criteria":[127],"support":[128],"explanation,":[130],"they":[131],"also":[132],"point":[133],"remaining":[135],"gaps":[136],"understanding.":[139],"provides":[142],"evidence":[143],"that":[144],"understanding":[147,159],"large":[149],"ML":[150],"feasible,":[153],"opening":[154],"opportunities":[155],"scale":[157],"both":[161],"more":[165],"complex":[166],"tasks.":[167]},"counts_by_year":[{"year":2026,"cited_by_count":8},{"year":2025,"cited_by_count":12},{"year":2024,"cited_by_count":21},{"year":2023,"cited_by_count":14}],"updated_date":"2026-11-03T10:50:19.383937","created_date":"2022-11-07T00:00:00"}