{"id":"https://openalex.org/W7166886429","doi":"https://doi.org/10.18653/v1/2026.findings-acl.1298","title":"Towards Understanding the Robustness of Sparse Autoencoders","display_name":"Towards Understanding the Robustness of Sparse Autoencoders","publication_year":2026,"publication_date":"2026-01-01","ids":{"openalex":"https://openalex.org/W7166886429","doi":"https://doi.org/10.18653/v1/2026.findings-acl.1298"},"language":null,"primary_location":{"id":"doi:10.18653/v1/2026.findings-acl.1298","is_oa":true,"landing_page_url":"https://doi.org/10.18653/v1/2026.findings-acl.1298","pdf_url":"https://aclanthology.org/2026.findings-acl.1298.pdf","source":null,"license":"cc-by","license_id":"https://openalex.org/licenses/cc-by","version":"publishedVersion","is_accepted":true,"is_published":true,"raw_source_name":"Findings of the Association for Computational Linguistics: ACL 2026","raw_type":"proceedings-article"},"type":"conference-paper","indexed_in":["crossref"],"open_access":{"is_oa":true,"oa_status":"gold","oa_url":"https://aclanthology.org/2026.findings-acl.1298.pdf","any_repository_has_fulltext":null},"authorships":[{"author_position":"first","author":{"id":"https://openalex.org/A5139831374","display_name":"Ahson Saiyed","orcid":null},"institutions":[{"id":"https://openalex.org/I51556381","display_name":"University of Virginia","ror":"https://ror.org/0153tk833","country_code":"US","type":"education","lineage":["https://openalex.org/I51556381"]}],"countries":["US"],"is_corresponding":false,"raw_author_name":"Ahson Saiyed","raw_affiliation_strings":["University of Virginia"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"University of Virginia","institution_ids":["https://openalex.org/I51556381"]}]},{"author_position":"middle","author":{"id":"https://openalex.org/A5139762868","display_name":"Sabrina Sadiekh","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Sabrina Sadiekh","raw_affiliation_strings":["Independent Researcher"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"Independent Researcher","institution_ids":[]}]},{"author_position":"last","author":{"id":"https://openalex.org/A5139822569","display_name":"Chirag Agarwal","orcid":null},"institutions":[{"id":"https://openalex.org/I51556381","display_name":"University of Virginia","ror":"https://ror.org/0153tk833","country_code":"US","type":"education","lineage":["https://openalex.org/I51556381"]}],"countries":["US"],"is_corresponding":false,"raw_author_name":"Chirag Agarwal","raw_affiliation_strings":["University of Virginia"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"University of Virginia","institution_ids":["https://openalex.org/I51556381"]}]}],"institutions":[],"countries_distinct_count":1,"institutions_distinct_count":1,"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":0.85887516,"is_in_top_1_percent":false,"is_in_top_10_percent":false},"cited_by_percentile_year":null,"biblio":{"volume":null,"issue":null,"first_page":"26066","last_page":"26086"},"is_retracted":false,"is_paratext":false,"is_xpac":false,"primary_topic":{"id":"https://openalex.org/T10775","display_name":"Generative Adversarial Networks and Image Synthesis","score":0.17270000278949738,"subfield":{"id":"https://openalex.org/subfields/1707","display_name":"Computer Vision and Pattern Recognition"},"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/T10775","display_name":"Generative Adversarial Networks and Image Synthesis","score":0.17270000278949738,"subfield":{"id":"https://openalex.org/subfields/1707","display_name":"Computer Vision and Pattern Recognition"},"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/T10057","display_name":"Face and Expression Recognition","score":0.10220000147819519,"subfield":{"id":"https://openalex.org/subfields/1707","display_name":"Computer Vision and Pattern Recognition"},"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/T11612","display_name":"Stochastic Gradient Optimization Techniques","score":0.09929999709129333,"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/robustness","display_name":"Robustness (evolution)","score":0.5788999795913696},{"id":"https://openalex.org/keywords/pattern-recognition","display_name":"Pattern recognition (psychology)","score":0.3517000079154968},{"id":"https://openalex.org/keywords/artificial-neural-network","display_name":"Artificial neural network","score":0.25690001249313354},{"id":"https://openalex.org/keywords/noisy-data","display_name":"Noisy data","score":0.2567000091075897},{"id":"https://openalex.org/keywords/noise","display_name":"Noise (video)","score":0.24660000205039978}],"concepts":[{"id":"https://openalex.org/C41008148","wikidata":"https://www.wikidata.org/wiki/Q21198","display_name":"Computer science","level":0,"score":0.6147000193595886},{"id":"https://openalex.org/C154945302","wikidata":"https://www.wikidata.org/wiki/Q11660","display_name":"Artificial intelligence","level":1,"score":0.6129000186920166},{"id":"https://openalex.org/C63479239","wikidata":"https://www.wikidata.org/wiki/Q7353546","display_name":"Robustness (evolution)","level":3,"score":0.5788999795913696},{"id":"https://openalex.org/C153180895","wikidata":"https://www.wikidata.org/wiki/Q7148389","display_name":"Pattern recognition (psychology)","level":2,"score":0.3517000079154968},{"id":"https://openalex.org/C119857082","wikidata":"https://www.wikidata.org/wiki/Q2539","display_name":"Machine learning","level":1,"score":0.33250001072883606},{"id":"https://openalex.org/C50644808","wikidata":"https://www.wikidata.org/wiki/Q192776","display_name":"Artificial neural network","level":2,"score":0.25690001249313354},{"id":"https://openalex.org/C2781170535","wikidata":"https://www.wikidata.org/wiki/Q30587856","display_name":"Noisy data","level":2,"score":0.2567000091075897},{"id":"https://openalex.org/C99498987","wikidata":"https://www.wikidata.org/wiki/Q2210247","display_name":"Noise (video)","level":3,"score":0.24660000205039978},{"id":"https://openalex.org/C2776401178","wikidata":"https://www.wikidata.org/wiki/Q12050496","display_name":"Feature (linguistics)","level":2,"score":0.24570000171661377},{"id":"https://openalex.org/C124101348","wikidata":"https://www.wikidata.org/wiki/Q172491","display_name":"Data mining","level":1,"score":0.24449999630451202}],"mesh":[],"locations_count":1,"locations":[{"id":"doi:10.18653/v1/2026.findings-acl.1298","is_oa":true,"landing_page_url":"https://doi.org/10.18653/v1/2026.findings-acl.1298","pdf_url":"https://aclanthology.org/2026.findings-acl.1298.pdf","source":null,"license":"cc-by","license_id":"https://openalex.org/licenses/cc-by","version":"publishedVersion","is_accepted":true,"is_published":true,"raw_source_name":"Findings of the Association for Computational Linguistics: ACL 2026","raw_type":"proceedings-article"}],"best_oa_location":{"id":"doi:10.18653/v1/2026.findings-acl.1298","is_oa":true,"landing_page_url":"https://doi.org/10.18653/v1/2026.findings-acl.1298","pdf_url":"https://aclanthology.org/2026.findings-acl.1298.pdf","source":null,"license":"cc-by","license_id":"https://openalex.org/licenses/cc-by","version":"publishedVersion","is_accepted":true,"is_published":true,"raw_source_name":"Findings of the Association for Computational Linguistics: ACL 2026","raw_type":"proceedings-article"},"sustainable_development_goals":[],"awards":[],"funders":[],"has_content":{"pdf":true,"grobid_xml":true},"content_urls":{"pdf":"https://content.openalex.org/works/W7166886429.pdf","grobid_xml":"https://content.openalex.org/works/W7166886429.grobid-xml"},"referenced_works_count":0,"referenced_works":[],"related_works":[],"abstract_inverted_index":{"Large":[0],"Language":[1],"Models":[2],"(LLMs)":[3],"remain":[4,26],"vulnerable":[5],"to":[6,71,80],"optimization-based":[7],"jailbreak":[8,76,134],"attacks":[9,60],"that":[10],"exploit":[11],"internal":[12],"gradient":[13],"structure.While":[14],"Sparse":[15],"Autoencoders":[16],"(SAEs)":[17],"are":[18,119],"widely":[19],"used":[20],"for":[21],"interpretability,":[22],"their":[23],"robustness":[24,114],"implications":[25],"underexplored.We":[27],"present":[28],"a":[29,72,92,106,122],"study":[30],"of":[31],"integrating":[32],"pretrained":[33],"SAEs":[34],"into":[35],"transformer":[36],"residual":[37],"streams":[38],"at":[39],"inference":[40],"time,":[41],"without":[42],"modifying":[43],"model":[44,50],"weights":[45],"or":[46],"blocking":[47],"gradients.Across":[48],"four":[49],"families":[51],"(Gemma,":[52],"LLaMA,":[53],"Mistral,":[54],"Qwen)":[55],"and":[56,84,100,104,115],"two":[57],"strong":[58],"white-box":[59],"(GCG,":[61],"BEAST)":[62],"plus":[63],"three":[64],"black-box":[65],"benchmarks,":[66],"SAE-augmented":[67],"models":[68],"achieve":[69],"up":[70],"5\u00d7":[73],"reduction":[74],"in":[75],"success":[77,102],"rate":[78],"relative":[79],"the":[81,129],"undefended":[82],"baseline":[83],"reduce":[85],"cross-model":[86],"attack":[87,101],"transferability.Parametric":[88],"ablations":[89],"reveal":[90],"(i)":[91],"monotonic":[93],"dose-response":[94],"relationship":[95],"between":[96],"L":[97],"0":[98],"sparsity":[99],"rate,":[103],"(ii)":[105],"layer-dependent":[107],"defense-utility":[108],"tradeoff,":[109],"where":[110],"intermediate":[111],"layers":[112],"balance":[113],"clean":[116],"performance.These":[117],"findings":[118],"consistent":[120],"with":[121],"representational":[123],"bottleneck":[124],"hypothesis:":[125],"sparse":[126],"projection":[127],"reshapes":[128],"optimization":[130],"geometry":[131],"exploited":[132],"by":[133],"attacks.":[135],"1":[136]},"counts_by_year":[],"updated_date":"2026-08-21T09:56:20.448147","created_date":"2026-07-02T00:00:00"}