{"id":"https://openalex.org/W4414992203","doi":"https://doi.org/10.18653/v1/2026.acl-long.82","title":"CRISP: Persistent Concept Unlearning via Sparse Autoencoders","display_name":"CRISP: Persistent Concept Unlearning via Sparse Autoencoders","publication_year":2026,"publication_date":"2026-01-01","ids":{"openalex":"https://openalex.org/W4414992203","doi":"https://doi.org/10.18653/v1/2026.acl-long.82"},"language":"en","primary_location":{"id":"doi:10.18653/v1/2026.acl-long.82","is_oa":true,"landing_page_url":"https://doi.org/10.18653/v1/2026.acl-long.82","pdf_url":null,"source":null,"license":"cc-by","license_id":"https://openalex.org/licenses/cc-by","version":"publishedVersion","is_accepted":true,"is_published":true,"raw_source_name":"Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)","raw_type":"proceedings-article"},"type":"conference-paper","indexed_in":["arxiv","crossref","datacite"],"open_access":{"is_oa":true,"oa_status":"gold","oa_url":"https://doi.org/10.18653/v1/2026.acl-long.82","any_repository_has_fulltext":true},"authorships":[{"author_position":"first","author":{"id":"https://openalex.org/A5099137410","display_name":"Tomer Ashuach","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Tomer Ashuach","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5023607289","display_name":"Dana Arad","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Dana Arad","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5020998070","display_name":"Aaron Mueller","orcid":"https://orcid.org/0009-0005-1148-5001"},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Aaron Mueller","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5020691628","display_name":"Martin Tutek","orcid":"https://orcid.org/0000-0001-5227-5397"},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Martin Tutek","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"last","author":{"id":"https://openalex.org/A5051184573","display_name":"Yonatan Belinkov","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Yonatan Belinkov","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]}],"institutions":[],"countries_distinct_count":0,"institutions_distinct_count":0,"corresponding_author_ids":[],"corresponding_institution_ids":[],"apc_list":null,"apc_paid":null,"fwci":0.0,"has_fulltext":true,"cited_by_count":0,"citation_normalized_percentile":{"value":0.01069263,"is_in_top_1_percent":false,"is_in_top_10_percent":false},"cited_by_percentile_year":null,"biblio":{"volume":null,"issue":null,"first_page":"1806","last_page":"1825"},"is_retracted":false,"is_paratext":false,"is_xpac":false,"primary_topic":{"id":"https://openalex.org/T12535","display_name":"Machine Learning and Data Classification","score":0.9437000155448914,"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/T12535","display_name":"Machine Learning and Data Classification","score":0.9437000155448914,"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/salient","display_name":"Salient","score":0.6959999799728394},{"id":"https://openalex.org/keywords/inference","display_name":"Inference","score":0.6507999897003174},{"id":"https://openalex.org/keywords/work","display_name":"Work (physics)","score":0.30720001459121704},{"id":"https://openalex.org/keywords/language-understanding","display_name":"Language understanding","score":0.29159998893737793},{"id":"https://openalex.org/keywords/autoencoder","display_name":"Autoencoder","score":0.27869999408721924}],"concepts":[{"id":"https://openalex.org/C41008148","wikidata":"https://www.wikidata.org/wiki/Q21198","display_name":"Computer science","level":0,"score":0.741100013256073},{"id":"https://openalex.org/C2780719617","wikidata":"https://www.wikidata.org/wiki/Q1030752","display_name":"Salient","level":2,"score":0.6959999799728394},{"id":"https://openalex.org/C2776214188","wikidata":"https://www.wikidata.org/wiki/Q408386","display_name":"Inference","level":2,"score":0.6507999897003174},{"id":"https://openalex.org/C154945302","wikidata":"https://www.wikidata.org/wiki/Q11660","display_name":"Artificial intelligence","level":1,"score":0.6492999792098999},{"id":"https://openalex.org/C119857082","wikidata":"https://www.wikidata.org/wiki/Q2539","display_name":"Machine learning","level":1,"score":0.43230000138282776},{"id":"https://openalex.org/C204321447","wikidata":"https://www.wikidata.org/wiki/Q30642","display_name":"Natural language processing","level":1,"score":0.4156000018119812},{"id":"https://openalex.org/C18762648","wikidata":"https://www.wikidata.org/wiki/Q42213","display_name":"Work (physics)","level":2,"score":0.30720001459121704},{"id":"https://openalex.org/C2983448237","wikidata":"https://www.wikidata.org/wiki/Q1078276","display_name":"Language understanding","level":2,"score":0.29159998893737793},{"id":"https://openalex.org/C101738243","wikidata":"https://www.wikidata.org/wiki/Q786435","display_name":"Autoencoder","level":3,"score":0.27869999408721924},{"id":"https://openalex.org/C26517878","wikidata":"https://www.wikidata.org/wiki/Q228039","display_name":"Key (lock)","level":2,"score":0.2540000081062317}],"mesh":[],"locations_count":3,"locations":[{"id":"doi:10.18653/v1/2026.acl-long.82","is_oa":true,"landing_page_url":"https://doi.org/10.18653/v1/2026.acl-long.82","pdf_url":null,"source":null,"license":"cc-by","license_id":"https://openalex.org/licenses/cc-by","version":"publishedVersion","is_accepted":true,"is_published":true,"raw_source_name":"Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)","raw_type":"proceedings-article"},{"id":"pmh:oai:arXiv.org:2508.13650","is_oa":true,"landing_page_url":"https://arxiv.org/abs/2508.13650","pdf_url":"https://arxiv.org/pdf/2508.13650","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"},{"id":"doi:10.48550/arxiv.2508.13650","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2508.13650","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,"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":null,"raw_source_name":null,"raw_type":"Preprint"}],"best_oa_location":{"id":"doi:10.18653/v1/2026.acl-long.82","is_oa":true,"landing_page_url":"https://doi.org/10.18653/v1/2026.acl-long.82","pdf_url":null,"source":null,"license":"cc-by","license_id":"https://openalex.org/licenses/cc-by","version":"publishedVersion","is_accepted":true,"is_published":true,"raw_source_name":"Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)","raw_type":"proceedings-article"},"sustainable_development_goals":[],"awards":[],"funders":[{"id":"https://openalex.org/F4320320300","display_name":"European Commission","ror":"https://ror.org/00k4n6c32"}],"has_content":{"pdf":false,"grobid_xml":false},"content_urls":null,"referenced_works_count":0,"referenced_works":[],"related_works":[],"abstract_inverted_index":{"As":[0],"large":[1],"language":[2],"models":[3],"(LLMs)":[4],"are":[5],"increasingly":[6],"deployed":[7],"in":[8,53],"real-world":[9],"applications,":[10],"the":[11,54,113,144],"need":[12],"to":[13,32],"selectively":[14],"remove":[15],"unwanted":[16],"knowledge":[17,119],"while":[18,120],"preserving":[19,121],"model":[20],"utility":[21],"has":[22,27],"become":[23],"paramount.":[24],"Recent":[25],"work":[26],"explored":[28],"sparse":[29],"autoencoders":[30],"(SAEs)":[31],"perform":[33],"precise":[34,141],"interventions":[35,58],"on":[36,108],"monosemantic":[37],"features.":[38,146],"However,":[39],"most":[40],"SAE-based":[41],"methods":[42],"operate":[43],"at":[44],"inference":[45],"time,":[46],"which":[47],"does":[48],"not":[49],"create":[50],"persistent":[51,77],"changes":[52],"model's":[55],"parameters.":[56],"Such":[57],"can":[59],"be":[60],"bypassed":[61],"or":[62],"reversed":[63],"by":[64],"malicious":[65],"actors":[66],"with":[67,97],"parameter":[68],"access.":[69],"We":[70,95],"introduce":[71],"CRISP,":[72],"a":[73],"parameter-efficient":[74],"method":[75,104],"for":[76],"concept":[78],"unlearning":[79,110],"using":[80],"SAEs.":[81],"CRISP":[82,130],"automatically":[83],"identifies":[84],"salient":[85],"SAE":[86],"features":[87],"across":[88],"multiple":[89],"layers":[90],"and":[91,100,123,137],"suppresses":[92],"their":[93],"activations.":[94],"experiment":[96],"two":[98],"LLMs":[99],"show":[101],"that":[102,129],"our":[103],"outperforms":[105],"prior":[106],"approaches":[107],"safety-critical":[109],"tasks":[111],"from":[112],"WMDP":[114],"benchmark,":[115],"successfully":[116],"removing":[117],"harmful":[118],"general":[122],"in-domain":[124],"capabilities.":[125],"Feature-level":[126],"analysis":[127],"reveals":[128],"achieves":[131],"semantically":[132],"coherent":[133],"separation":[134],"between":[135],"target":[136,145],"benign":[138],"concepts,":[139],"allowing":[140],"suppression":[142],"of":[143]},"counts_by_year":[],"updated_date":"2026-07-29T14:22:42.915294","created_date":"2025-10-10T00:00:00"}