{"id":"https://openalex.org/W4414787224","doi":"https://doi.org/10.18653/v1/2026.findings-acl.944","title":"Safe-SAIL: Towards a Fine-grained Safety Landscape of Large Language Models via Sparse Autoencoder Interpretation Framework","display_name":"Safe-SAIL: Towards a Fine-grained Safety Landscape of Large Language Models via Sparse Autoencoder Interpretation Framework","publication_year":2026,"publication_date":"2026-01-01","ids":{"openalex":"https://openalex.org/W4414787224","doi":"https://doi.org/10.18653/v1/2026.findings-acl.944"},"language":"en","primary_location":{"id":"doi:10.18653/v1/2026.findings-acl.944","is_oa":true,"landing_page_url":"https://doi.org/10.18653/v1/2026.findings-acl.944","pdf_url":"https://aclanthology.org/2026.findings-acl.944.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":["arxiv","crossref","datacite"],"open_access":{"is_oa":true,"oa_status":"gold","oa_url":"https://aclanthology.org/2026.findings-acl.944.pdf","any_repository_has_fulltext":true},"authorships":[{"author_position":"first","author":{"id":"https://openalex.org/A5108318734","display_name":"Jiaqi Weng","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Jiaqi Weng","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5101712340","display_name":"Han Zheng","orcid":"https://orcid.org/0000-0001-5051-071X"},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Han Zheng","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5100778121","display_name":"Hanyu Zhang","orcid":"https://orcid.org/0000-0002-4728-7702"},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Hanyu Zhang","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":null,"display_name":"Ej Zhou","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Ej Zhou","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5010316915","display_name":"Qinqin He","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Qinqin He","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":null,"display_name":"Jialing Tao","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Jialing Tao","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":null,"display_name":"Hui Xue","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Hui Xue","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":null,"display_name":"Zhixuan Chu","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Zhixuan Chu","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"last","author":{"id":"https://openalex.org/A5018582673","display_name":"Xiting Wang","orcid":"https://orcid.org/0000-0002-1846-1118"},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Xiting Wang","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":0.0082906,"is_in_top_1_percent":false,"is_in_top_10_percent":false},"cited_by_percentile_year":null,"biblio":{"volume":null,"issue":null,"first_page":"18916","last_page":"18935"},"is_retracted":false,"is_paratext":false,"is_xpac":false,"primary_topic":{"id":"https://openalex.org/T10028","display_name":"Topic Modeling","score":0.7864000201225281,"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/T10028","display_name":"Topic Modeling","score":0.7864000201225281,"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/T12423","display_name":"Software Reliability and Analysis Research","score":0.7731000185012817,"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"}}],"keywords":[{"id":"https://openalex.org/keywords/interpretability","display_name":"Interpretability","score":0.9261000156402588},{"id":"https://openalex.org/keywords/limiting","display_name":"Limiting","score":0.5042999982833862},{"id":"https://openalex.org/keywords/set","display_name":"Set (abstract data type)","score":0.4936000108718872},{"id":"https://openalex.org/keywords/software-deployment","display_name":"Software deployment","score":0.4763999879360199},{"id":"https://openalex.org/keywords/interpretation","display_name":"Interpretation (philosophy)","score":0.45980000495910645},{"id":"https://openalex.org/keywords/feature","display_name":"Feature (linguistics)","score":0.4339999854564667},{"id":"https://openalex.org/keywords/autoencoder","display_name":"Autoencoder","score":0.4140999913215637}],"concepts":[{"id":"https://openalex.org/C2781067378","wikidata":"https://www.wikidata.org/wiki/Q17027399","display_name":"Interpretability","level":2,"score":0.9261000156402588},{"id":"https://openalex.org/C41008148","wikidata":"https://www.wikidata.org/wiki/Q21198","display_name":"Computer science","level":0,"score":0.6545000076293945},{"id":"https://openalex.org/C188198153","wikidata":"https://www.wikidata.org/wiki/Q1613840","display_name":"Limiting","level":2,"score":0.5042999982833862},{"id":"https://openalex.org/C177264268","wikidata":"https://www.wikidata.org/wiki/Q1514741","display_name":"Set (abstract data type)","level":2,"score":0.4936000108718872},{"id":"https://openalex.org/C105339364","wikidata":"https://www.wikidata.org/wiki/Q2297740","display_name":"Software deployment","level":2,"score":0.4763999879360199},{"id":"https://openalex.org/C527412718","wikidata":"https://www.wikidata.org/wiki/Q855395","display_name":"Interpretation (philosophy)","level":2,"score":0.45980000495910645},{"id":"https://openalex.org/C2776401178","wikidata":"https://www.wikidata.org/wiki/Q12050496","display_name":"Feature (linguistics)","level":2,"score":0.4339999854564667},{"id":"https://openalex.org/C101738243","wikidata":"https://www.wikidata.org/wiki/Q786435","display_name":"Autoencoder","level":3,"score":0.4140999913215637},{"id":"https://openalex.org/C112930515","wikidata":"https://www.wikidata.org/wiki/Q4389547","display_name":"Risk analysis (engineering)","level":1,"score":0.41350001096725464},{"id":"https://openalex.org/C154945302","wikidata":"https://www.wikidata.org/wiki/Q11660","display_name":"Artificial intelligence","level":1,"score":0.4065000116825104},{"id":"https://openalex.org/C119857082","wikidata":"https://www.wikidata.org/wiki/Q2539","display_name":"Machine learning","level":1,"score":0.38589999079704285},{"id":"https://openalex.org/C168167062","wikidata":"https://www.wikidata.org/wiki/Q1117970","display_name":"Component (thermodynamics)","level":2,"score":0.3628000020980835},{"id":"https://openalex.org/C2522767166","wikidata":"https://www.wikidata.org/wiki/Q2374463","display_name":"Data science","level":1,"score":0.35499998927116394},{"id":"https://openalex.org/C26517878","wikidata":"https://www.wikidata.org/wiki/Q228039","display_name":"Key (lock)","level":2,"score":0.35370001196861267},{"id":"https://openalex.org/C120936955","wikidata":"https://www.wikidata.org/wiki/Q2155640","display_name":"Empirical research","level":2,"score":0.3400000035762787},{"id":"https://openalex.org/C2779304628","wikidata":"https://www.wikidata.org/wiki/Q3503480","display_name":"Face (sociological concept)","level":2,"score":0.335099995136261},{"id":"https://openalex.org/C137293760","wikidata":"https://www.wikidata.org/wiki/Q3621696","display_name":"Language model","level":2,"score":0.29339998960494995},{"id":"https://openalex.org/C2778755073","wikidata":"https://www.wikidata.org/wiki/Q10858537","display_name":"Scale (ratio)","level":2,"score":0.2705000042915344},{"id":"https://openalex.org/C3017944768","wikidata":"https://www.wikidata.org/wiki/Q1450463","display_name":"Poison control","level":2,"score":0.26339998841285706},{"id":"https://openalex.org/C12174686","wikidata":"https://www.wikidata.org/wiki/Q1058438","display_name":"Risk assessment","level":2,"score":0.2547999918460846},{"id":"https://openalex.org/C124101348","wikidata":"https://www.wikidata.org/wiki/Q172491","display_name":"Data mining","level":1,"score":0.2535000145435333}],"mesh":[],"locations_count":3,"locations":[{"id":"doi:10.18653/v1/2026.findings-acl.944","is_oa":true,"landing_page_url":"https://doi.org/10.18653/v1/2026.findings-acl.944","pdf_url":"https://aclanthology.org/2026.findings-acl.944.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"},{"id":"pmh:oai:arXiv.org:2509.18127","is_oa":true,"landing_page_url":"https://arxiv.org/abs/2509.18127","pdf_url":"https://arxiv.org/pdf/2509.18127","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":"cc-by-nc-sa","license_id":"https://openalex.org/licenses/cc-by-nc-sa","version":"submittedVersion","is_accepted":false,"is_published":false,"raw_source_name":null,"raw_type":"text"},{"id":"doi:10.48550/arxiv.2509.18127","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2509.18127","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.findings-acl.944","is_oa":true,"landing_page_url":"https://doi.org/10.18653/v1/2026.findings-acl.944","pdf_url":"https://aclanthology.org/2026.findings-acl.944.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":[{"id":"https://openalex.org/F4320321001","display_name":"National Natural Science Foundation of China","ror":"https://ror.org/01h0zpd94"},{"id":"https://openalex.org/F4320322927","display_name":"Zhejiang University","ror":"https://ror.org/00a2xv884"},{"id":"https://openalex.org/F4320338464","display_name":"Natural Science Foundation of Zhejiang Province","ror":"https://ror.org/01h0zpd94"}],"has_content":{"pdf":true,"grobid_xml":true},"content_urls":{"pdf":"https://content.openalex.org/works/W4414787224.pdf","grobid_xml":"https://content.openalex.org/works/W4414787224.grobid-xml"},"referenced_works_count":0,"referenced_works":[],"related_works":[],"abstract_inverted_index":{"Sparse":[0],"autoencoders":[1],"(SAEs)":[2],"enable":[3],"interpretability":[4],"research":[5],"by":[6,93],"decomposing":[7],"entangled":[8],"model":[9,154],"activations":[10],"into":[11],"monosemantic":[12],"features.However,":[13],"under":[14],"what":[15],"circumstances":[16],"SAEs":[17,33,83,108],"derive":[18],"most":[19],"fine-grained":[20],"latent":[21],"features":[22,63,118],"for":[23,38,60,115,141],"safety-a":[24],"low-frequency":[25],"concept":[26],"domain-remains":[27],"unexplored.Two":[28],"key":[29],"challenges":[30],"exist:":[31],"identifying":[32],"with":[34,84,109],"the":[35,44,137],"greatest":[36],"potential":[37],"generating":[39],"safety":[40,86],"domain-specific":[41,87],"features,":[42],"and":[43,89,112,125,133,145,149,158,168],"prohibitively":[45],"high":[46],"cost":[47,92],"of":[48,71,107,139],"detailed":[49],"feature":[50,143],"explanation.In":[51],"this":[52,127],"paper,":[53],"we":[54,102,129],"propose":[55],"Safe-SAIL,":[56,101],"a":[57,76,96,104],"unified":[58],"framework":[59],"interpreting":[61],"SAE":[62],"in":[64,163],"safety-critical":[65],"domains":[66],"to":[67,80],"advance":[68],"mechanistic":[69],"understanding":[70],"large":[72],"language":[73],"models.Safe-SAIL":[74],"introduces":[75],"pre-explanation":[77],"evaluation":[78],"metric":[79],"efficiently":[81],"identify":[82],"strong":[85],"interpretability,":[88],"reduces":[90],"interpretation":[91],"55%":[94],"through":[95],"segmentlevel":[97],"simulation":[98],"strategy.Building":[99],"on":[100,136],"train":[103],"comprehensive":[105],"suite":[106],"human-readable":[110],"explanations":[111],"systematic":[113],"evaluations":[114],"1,758":[116],"safety-related":[117],"spanning":[119],"four":[120],"domains:":[121],"pornography,":[122],"politics,":[123],"violence,":[124],"terror.Using":[126],"resource,":[128],"conduct":[130],"empirical":[131],"analyses":[132],"provide":[134],"insights":[135],"effectiveness":[138],"Safe-SAIL":[140],"risk":[142],"identification":[144],"how":[146],"safetycritical":[147],"entities":[148],"concepts":[150],"are":[151,160],"encoded":[152],"across":[153],"layers.All":[155],"models,":[156],"explanations,":[157],"tools":[159],"publicly":[161],"released":[162],"our":[164],"open-source":[165],"toolkit":[166],"1":[167],"companion":[169],"product":[170],"2":[171],".":[172]},"counts_by_year":[],"updated_date":"2026-08-01T09:00:35.917206","created_date":"2025-10-10T00:00:00"}