{"id":"https://openalex.org/W4403578326","doi":"https://doi.org/10.48550/arxiv.2410.12555","title":"Investigating Sensitive Directions in GPT-2: An Improved Baseline and Comparative Analysis of SAEs","display_name":"Investigating Sensitive Directions in GPT-2: An Improved Baseline and Comparative Analysis of SAEs","publication_year":2024,"publication_date":"2024-10-16","ids":{"openalex":"https://openalex.org/W4403578326","doi":"https://doi.org/10.48550/arxiv.2410.12555"},"language":"en","primary_location":{"id":"pmh:oai:arXiv.org:2410.12555","doi":null,"is_oa":true,"landing_page_url":"http://arxiv.org/abs/2410.12555","pdf_url":"https://arxiv.org/pdf/2410.12555","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","is_traditional":true,"indexed_in":["arxiv","datacite"],"open_access":{"is_oa":true,"oa_status":"green","oa_url":"https://arxiv.org/pdf/2410.12555","any_repository_has_fulltext":true},"authorships":[{"author_position":"first","author":{"id":"https://openalex.org/A5105670161","display_name":"Daniel J. Lee","orcid":"https://orcid.org/0000-0002-6837-3474","observed_orcids":["https://orcid.org/0000-0002-6837-3474"]},"institutions":[{"id":"https://openalex.org/I136199984","display_name":"Harvard University","ror":"https://ror.org/03vek6s52","country_code":"US","type":"education","lineage":["https://openalex.org/I136199984"]}],"countries":["US"],"is_corresponding":false,"raw_author_name":"Lee, Daniel J.","raw_affiliation_strings":["Department of Biomedical Informatics Harvard Medical School Boston , MA"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"Department of Biomedical Informatics Harvard Medical School Boston , MA","institution_ids":["https://openalex.org/I136199984"]}]},{"author_position":"last","author":{"id":"https://openalex.org/A5114412487","display_name":"Stefan Heimersheim","orcid":null,"observed_orcids":[]},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Heimersheim, Stefan","raw_affiliation_strings":["Apollo Research London , UK"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"Apollo Research London , UK","institution_ids":[]}]}],"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":1,"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":0.9458000063896179,"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":0.9458000063896179,"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":0.04309999942779541,"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":0.003599999938160181,"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/sparse-autoencoders","display_name":"sparse autoencoders","score":0.9520000219345093},{"id":"https://openalex.org/keywords/gpt-2","display_name":"GPT-2","score":0.9190000295639038},{"id":"https://openalex.org/keywords/language-model-interpretability","display_name":"language model interpretability","score":0.628000020980835},{"id":"https://openalex.org/keywords/next-token-prediction","display_name":"next-token prediction","score":0.8489999771118164}],"concepts":[{"id":"https://openalex.org/C12725497","wikidata":"https://www.wikidata.org/wiki/Q810247","display_name":"Baseline (sea)","level":2,"score":0.7829174995422363},{"id":"https://openalex.org/C197934379","wikidata":"https://www.wikidata.org/wiki/Q2047938","display_name":"Adverse effect","level":2,"score":0.42172834277153015},{"id":"https://openalex.org/C71924100","wikidata":"https://www.wikidata.org/wiki/Q11190","display_name":"Medicine","level":0,"score":0.35039079189300537},{"id":"https://openalex.org/C126322002","wikidata":"https://www.wikidata.org/wiki/Q11180","display_name":"Internal medicine","level":1,"score":0.32447659969329834},{"id":"https://openalex.org/C86803240","wikidata":"https://www.wikidata.org/wiki/Q420","display_name":"Biology","level":0,"score":0.1517116129398346},{"id":"https://openalex.org/C505870484","wikidata":"https://www.wikidata.org/wiki/Q180538","display_name":"Fishery","level":1,"score":0.0}],"mesh":[],"locations_count":2,"locations":[{"id":"pmh:oai:arXiv.org:2410.12555","doi":null,"is_oa":true,"landing_page_url":"http://arxiv.org/abs/2410.12555","pdf_url":"https://arxiv.org/pdf/2410.12555","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.2410.12555","doi":"https://doi.org/10.48550/arxiv.2410.12555","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2410.12555","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":null,"license_id":null,"version":null,"is_accepted":false,"is_published":false,"raw_source_name":null,"raw_type":"Preprint"}],"best_oa_location":{"id":"pmh:oai:arXiv.org:2410.12555","doi":null,"is_oa":true,"landing_page_url":"http://arxiv.org/abs/2410.12555","pdf_url":"https://arxiv.org/pdf/2410.12555","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":{"grobid_xml":true,"pdf":true},"content_urls":{"pdf":"https://content.openalex.org/works/W4403578326.pdf","grobid_xml":"https://content.openalex.org/works/W4403578326.grobid-xml"},"referenced_works_count":0,"referenced_works":[],"related_works":["https://openalex.org/W4414849786","https://openalex.org/W7164023407","https://openalex.org/W4416037078","https://openalex.org/W7170560400","https://openalex.org/W7151244895","https://openalex.org/W7222852349","https://openalex.org/W4403780485","https://openalex.org/W7084392038","https://openalex.org/W7223054405","https://openalex.org/W4406950770"],"abstract_inverted_index":{"Sensitive":[0],"directions":[1,33,69,89],"experiments":[2],"attempt":[3],"to":[4,60,110],"understand":[5],"the":[6,17,31,61,81],"computational":[7],"features":[8,100],"of":[9],"Language":[10],"Models":[11],"(LMs)":[12],"by":[13,23,35,71],"measuring":[14],"how":[15],"much":[16],"next":[18],"token":[19],"prediction":[20],"probabilities":[21],"change":[22],"perturbing":[24],"activations":[25],"along":[26],"specific":[27],"directions.":[28,42],"We":[29,43,64],"extend":[30],"sensitive":[32],"work":[34],"introducing":[36],"an":[37],"improved":[38,62],"baseline":[39],"for":[40,48],"perturbation":[41],"demonstrate":[44],"that":[45,67,97],"KL":[46],"divergence":[47],"Sparse":[49],"Autoencoder":[50],"(SAE)":[51],"reconstruction":[52],"errors":[53],"are":[54],"no":[55],"longer":[56],"pathologically":[57],"high":[58],"compared":[59,109],"baseline.":[63],"also":[65],"show":[66],"feature":[68,88],"uncovered":[70],"SAEs":[72],"have":[73],"varying":[74],"impacts":[75],"on":[76,80,106],"model":[77,107],"outputs":[78,108],"depending":[79],"SAE's":[82],"sparsity,":[83],"with":[84],"lower":[85],"L0":[86],"SAE":[87,99],"exerting":[90],"a":[91],"greater":[92],"influence.":[93],"Additionally,":[94],"we":[95],"find":[96],"end-to-end":[98],"do":[101],"not":[102],"exhibit":[103],"stronger":[104],"effects":[105],"traditional":[111],"SAEs.":[112]},"counts_by_year":[{"year":2026,"cited_by_count":1}],"updated_date":"2026-11-04T10:54:05.112534","created_date":"2024-10-20T00:00:00"}