{"id":"https://openalex.org/W4412886691","doi":"https://doi.org/10.18653/v1/2025.acl-long.229","title":"Unveiling Language-Specific Features in Large Language Models via Sparse Autoencoders","display_name":"Unveiling Language-Specific Features in Large Language Models via Sparse Autoencoders","publication_year":2025,"publication_date":"2025-01-01","ids":{"openalex":"https://openalex.org/W4412886691","doi":"https://doi.org/10.18653/v1/2025.acl-long.229"},"language":"en","primary_location":{"id":"doi:10.18653/v1/2025.acl-long.229","is_oa":true,"landing_page_url":"https://doi.org/10.18653/v1/2025.acl-long.229","pdf_url":"https://aclanthology.org/2025.acl-long.229.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":"Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)","raw_type":"proceedings-article"},"type":"conference-paper","indexed_in":["crossref"],"open_access":{"is_oa":true,"oa_status":"gold","oa_url":"https://aclanthology.org/2025.acl-long.229.pdf","any_repository_has_fulltext":null},"authorships":[{"author_position":"first","author":{"id":null,"display_name":"Boyi Deng","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Boyi Deng","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5100377278","display_name":"Yu Wan","orcid":"https://orcid.org/0000-0002-7310-7883"},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Yu Wan","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5028040391","display_name":"Baosong Yang","orcid":"https://orcid.org/0000-0001-5002-2409"},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Baosong Yang","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5100652602","display_name":"Yidan Zhang","orcid":"https://orcid.org/0000-0002-7466-0234"},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Yidan Zhang","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"last","author":{"id":"https://openalex.org/A5051925942","display_name":"Fuli Feng","orcid":"https://orcid.org/0000-0002-5828-9842"},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Fuli Feng","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":null,"has_fulltext":true,"cited_by_count":0,"citation_normalized_percentile":null,"cited_by_percentile_year":null,"biblio":{"volume":null,"issue":null,"first_page":"4563","last_page":"4608"},"is_retracted":false,"is_paratext":false,"is_xpac":false,"primary_topic":{"id":"https://openalex.org/T10181","display_name":"Natural Language Processing Techniques","score":0.9865000247955322,"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/T10181","display_name":"Natural Language Processing Techniques","score":0.9865000247955322,"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":0.9803000092506409,"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/computer-science","display_name":"Computer science","score":0.7470785975456238},{"id":"https://openalex.org/keywords/artificial-intelligence","display_name":"Artificial intelligence","score":0.5874465107917786},{"id":"https://openalex.org/keywords/natural-language-processing","display_name":"Natural language processing","score":0.5460475087165833},{"id":"https://openalex.org/keywords/language-model","display_name":"Language model","score":0.4611224830150604}],"concepts":[{"id":"https://openalex.org/C41008148","wikidata":"https://www.wikidata.org/wiki/Q21198","display_name":"Computer science","level":0,"score":0.7470785975456238},{"id":"https://openalex.org/C154945302","wikidata":"https://www.wikidata.org/wiki/Q11660","display_name":"Artificial intelligence","level":1,"score":0.5874465107917786},{"id":"https://openalex.org/C204321447","wikidata":"https://www.wikidata.org/wiki/Q30642","display_name":"Natural language processing","level":1,"score":0.5460475087165833},{"id":"https://openalex.org/C137293760","wikidata":"https://www.wikidata.org/wiki/Q3621696","display_name":"Language model","level":2,"score":0.4611224830150604}],"mesh":[],"locations_count":1,"locations":[{"id":"doi:10.18653/v1/2025.acl-long.229","is_oa":true,"landing_page_url":"https://doi.org/10.18653/v1/2025.acl-long.229","pdf_url":"https://aclanthology.org/2025.acl-long.229.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":"Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)","raw_type":"proceedings-article"}],"best_oa_location":{"id":"doi:10.18653/v1/2025.acl-long.229","is_oa":true,"landing_page_url":"https://doi.org/10.18653/v1/2025.acl-long.229","pdf_url":"https://aclanthology.org/2025.acl-long.229.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":"Proceedings of the 63rd 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/F4320325599","display_name":"University of Science and Technology of China","ror":"https://ror.org/04c4dkn09"}],"has_content":{"pdf":true,"grobid_xml":true},"content_urls":{"pdf":"https://content.openalex.org/works/W4412886691.pdf","grobid_xml":"https://content.openalex.org/works/W4412886691.grobid-xml"},"referenced_works_count":0,"referenced_works":[],"related_works":["https://openalex.org/W4391375266","https://openalex.org/W2899084033","https://openalex.org/W2748952813","https://openalex.org/W2390279801","https://openalex.org/W4391913857","https://openalex.org/W2358668433","https://openalex.org/W4396701345","https://openalex.org/W2376932109","https://openalex.org/W2001405890","https://openalex.org/W3204019825"],"abstract_inverted_index":{"The":[0],"mechanisms":[1],"behind":[2],"multilingual":[3],"capabilities":[4],"in":[5,89],"Large":[6],"Language":[7],"Models":[8],"(LLMs)":[9],"have":[10,102],"been":[11],"examined":[12],"using":[13],"neuron-based":[14],"or":[15],"internal-activation-based":[16],"methods.However,":[17],"these":[18,82,119],"methods":[19],"often":[20],"face":[21],"challenges":[22],"such":[23],"as":[24],"superposition":[25],"and":[26,107],"layer-wise":[27],"activation":[28],"variance,":[29],"which":[30],"limit":[31],"their":[32],"reliability.Sparse":[33],"Autoencoders":[34],"(SAEs)":[35],"offer":[36],"a":[37,48,56],"more":[38],"nuanced":[39],"analysis":[40],"by":[41,133],"decomposing":[42],"the":[43,61,130],"activations":[44],"of":[45,52,63,92],"LLMs":[46],"into":[47],"sparse":[49],"linear":[50],"combination":[51],"SAE":[53,83,105],"features.We":[54],"introduce":[55],"novel":[57],"metric":[58],"to":[59,75,123],"assess":[60],"monolinguality":[62],"features":[64,71,84,122],"obtained":[65],"from":[66],"SAEs,":[67],"discovering":[68],"that":[69,80],"some":[70,100],"are":[72],"strongly":[73],"related":[74],"specific":[76],"languages.Additionally,":[77],"we":[78,98,117],"show":[79],"ablating":[81,108,115],"only":[85],"significantly":[86],"reduces":[87],"abilities":[88],"one":[90],"language":[91,131],"LLMs,":[93],"leaving":[94],"others":[95],"almost":[96],"unaffected.Interestingly,":[97],"find":[99],"languages":[101],"multiple":[103],"synergistic":[104],"features,":[106],"them":[109],"together":[110],"yields":[111],"greater":[112],"improvement":[113],"than":[114],"individually.Moreover,":[116],"leverage":[118],"SAE-derived":[120],"language-specific":[121],"enhance":[124],"steering":[125],"vectors,":[126],"achieving":[127],"control":[128],"over":[129],"generated":[132],"LLMs.The":[134],"code":[135],"is":[136],"publicly":[137],"available":[138],"at":[139],"https://github.com/":[140],"Aatrox103/multilingual-llm-features.":[141]},"counts_by_year":[],"updated_date":"2026-07-14T23:27:15.235271","created_date":"2025-10-10T00:00:00"}