{"id":"https://openalex.org/W4392428679","doi":"https://doi.org/10.48550/arxiv.2403.00745","title":"AtP*: An efficient and scalable method for localizing LLM behaviour to components","display_name":"AtP*: An efficient and scalable method for localizing LLM behaviour to components","publication_year":2024,"publication_date":"2024-03-01","ids":{"openalex":"https://openalex.org/W4392428679","doi":"https://doi.org/10.48550/arxiv.2403.00745"},"language":"en","primary_location":{"id":"pmh:oai:arXiv.org:2403.00745","doi":null,"is_oa":true,"landing_page_url":"http://arxiv.org/abs/2403.00745","pdf_url":"https://arxiv.org/pdf/2403.00745","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/2403.00745","any_repository_has_fulltext":true},"authorships":[{"author_position":"first","author":{"id":"https://openalex.org/A5027979882","display_name":"J\u00e1nos Kram\u00e1r","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":"Kram\u00e1r, J\u00e1nos","raw_affiliation_strings":["Google DeepMind"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"Google DeepMind","institution_ids":["https://openalex.org/I4210090411"]}]},{"author_position":"middle","author":{"id":"https://openalex.org/A5026856356","display_name":"Tom Lieberum","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":"Lieberum, Tom","raw_affiliation_strings":["Google DeepMind"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"Google DeepMind","institution_ids":["https://openalex.org/I4210090411"]}]},{"author_position":"middle","author":{"id":"https://openalex.org/A5012971694","display_name":"Rohin Shah","orcid":"https://orcid.org/0000-0002-0656-2800","observed_orcids":["https://orcid.org/0000-0002-0656-2800"]},"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":"Shah, Rohin","raw_affiliation_strings":["Google DeepMind"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"Google DeepMind","institution_ids":["https://openalex.org/I4210090411"]}]},{"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":1,"institutions_distinct_count":1,"corresponding_author_ids":[],"corresponding_institution_ids":["https://openalex.org/I4210090411"],"apc_list":null,"apc_paid":null,"fwci":null,"has_fulltext":false,"cited_by_count":0,"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":0.9990000128746033,"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":0.9990000128746033,"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/T10743","display_name":"Software Testing and Debugging Techniques","score":0.0003000000142492354,"subfield":{"id":"https://openalex.org/subfields/1712","display_name":"Software"},"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":9.999999747378752e-05,"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-patching","display_name":"activation patching","score":0.9679999947547913},{"id":"https://openalex.org/keywords/large-language-models","display_name":"large language models","score":0.8109999895095825},{"id":"https://openalex.org/keywords/causal-attribution","display_name":"causal attribution","score":0.6100000143051147},{"id":"https://openalex.org/keywords/false-negative-results","display_name":"false-negative results","score":0.6380000114440918},{"id":"https://openalex.org/keywords/scalability","display_name":"scalability","score":0.5830000042915344},{"id":"https://openalex.org/keywords/failure-modes","display_name":"failure modes","score":0.546999990940094}],"concepts":[{"id":"https://openalex.org/C48044578","wikidata":"https://www.wikidata.org/wiki/Q727490","display_name":"Scalability","level":2,"score":0.6638692617416382},{"id":"https://openalex.org/C41008148","wikidata":"https://www.wikidata.org/wiki/Q21198","display_name":"Computer science","level":0,"score":0.45764607191085815},{"id":"https://openalex.org/C185592680","wikidata":"https://www.wikidata.org/wiki/Q2329","display_name":"Chemistry","level":0,"score":0.4567597508430481},{"id":"https://openalex.org/C70721500","wikidata":"https://www.wikidata.org/wiki/Q177005","display_name":"Computational biology","level":1,"score":0.3636455833911896},{"id":"https://openalex.org/C77088390","wikidata":"https://www.wikidata.org/wiki/Q8513","display_name":"Database","level":1,"score":0.19148603081703186},{"id":"https://openalex.org/C86803240","wikidata":"https://www.wikidata.org/wiki/Q420","display_name":"Biology","level":0,"score":0.18411248922348022}],"mesh":[],"locations_count":2,"locations":[{"id":"pmh:oai:arXiv.org:2403.00745","doi":null,"is_oa":true,"landing_page_url":"http://arxiv.org/abs/2403.00745","pdf_url":"https://arxiv.org/pdf/2403.00745","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.2403.00745","doi":"https://doi.org/10.48550/arxiv.2403.00745","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2403.00745","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:2403.00745","doi":null,"is_oa":true,"landing_page_url":"http://arxiv.org/abs/2403.00745","pdf_url":"https://arxiv.org/pdf/2403.00745","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":{"pdf":false,"grobid_xml":false},"content_urls":null,"referenced_works_count":0,"referenced_works":[],"related_works":["https://openalex.org/W7223113814","https://openalex.org/W7164208903","https://openalex.org/W7079293072","https://openalex.org/W7214106579","https://openalex.org/W4404987730","https://openalex.org/W7212459010","https://openalex.org/W7140127889","https://openalex.org/W4415109308","https://openalex.org/W7164919854","https://openalex.org/W7134130596"],"abstract_inverted_index":{"Activation":[0,53],"Patching":[1,46,54],"is":[2],"a":[3,20,48,72,123],"method":[4,124],"of":[5,10,29,59,62,74,95,129,133],"directly":[6],"computing":[7],"causal":[8],"attributions":[9],"behavior":[11],"to":[12,52,66,81,125],"model":[13,30],"components.":[14],"However,":[15],"applying":[16],"it":[17],"exhaustively":[18],"requires":[19],"sweep":[21],"with":[22,78,114],"cost":[23],"scaling":[24],"linearly":[25],"in":[26],"the":[27,91,127],"number":[28],"components,":[31],"which":[32,64],"can":[33],"be":[34],"prohibitively":[35],"expensive":[36],"for":[37,100],"SoTA":[38],"Large":[39],"Language":[40],"Models":[41],"(LLMs).":[42],"We":[43,70,89],"investigate":[44],"Attribution":[45],"(AtP),":[47],"fast":[49],"gradient-based":[50],"approximation":[51],"and":[55,97,104],"find":[56],"two":[57,79],"classes":[58],"failure":[60,84],"modes":[61,85],"AtP":[63,75,96,107],"lead":[65],"significant":[67,118],"false":[68,131],"negatives.":[69],"propose":[71],"variant":[73],"called":[76],"AtP*,":[77],"changes":[80],"address":[82],"these":[83],"while":[86],"retaining":[87],"scalability.":[88],"present":[90],"first":[92],"systematic":[93],"study":[94],"alternative":[98],"methods":[99],"faster":[101],"activation":[102],"patching":[103],"show":[105],"that":[106],"significantly":[108],"outperforms":[109],"all":[110],"other":[111],"investigated":[112],"methods,":[113],"AtP*":[115,134],"providing":[116],"further":[117],"improvement.":[119],"Finally,":[120],"we":[121],"provide":[122],"bound":[126],"probability":[128],"remaining":[130],"negatives":[132],"estimates.":[135]},"counts_by_year":[],"updated_date":"2026-10-01T08:54:57.152042","created_date":"2024-03-05T00:00:00"}