{"id":"https://openalex.org/W7166854504","doi":"https://doi.org/10.18653/v1/2026.acl-long.2214","title":"Beyond Variance: Knowledge-Aware LLM Compression via Fisher-Aligned Subspace Diagnostics","display_name":"Beyond Variance: Knowledge-Aware LLM Compression via Fisher-Aligned Subspace Diagnostics","publication_year":2026,"publication_date":"2026-01-01","ids":{"openalex":"https://openalex.org/W7166854504","doi":"https://doi.org/10.18653/v1/2026.acl-long.2214"},"language":null,"primary_location":{"id":"doi:10.18653/v1/2026.acl-long.2214","is_oa":true,"landing_page_url":"https://doi.org/10.18653/v1/2026.acl-long.2214","pdf_url":"https://aclanthology.org/2026.acl-long.2214.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 64th 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/2026.acl-long.2214.pdf","any_repository_has_fulltext":null},"authorships":[{"author_position":"first","author":{"id":"https://openalex.org/A5139844005","display_name":"Ibne Farabi Shihab","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Ibne Farabi Shihab","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5119832027","display_name":"Sanjeda Akter","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Sanjeda Akter","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"last","author":{"id":"https://openalex.org/A5139733942","display_name":"Anuj Sharma","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Anuj Sharma","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":0.84373338,"is_in_top_1_percent":false,"is_in_top_10_percent":false},"cited_by_percentile_year":null,"biblio":{"volume":null,"issue":null,"first_page":"47938","last_page":"47954"},"is_retracted":false,"is_paratext":false,"is_xpac":false,"primary_topic":{"id":"https://openalex.org/T12072","display_name":"Machine Learning and Algorithms","score":0.20180000364780426,"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/T12072","display_name":"Machine Learning and Algorithms","score":0.20180000364780426,"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/T10181","display_name":"Natural Language Processing Techniques","score":0.12479999661445618,"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":0.07339999824762344,"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/compression","display_name":"Compression (physics)","score":0.39570000767707825},{"id":"https://openalex.org/keywords/noise","display_name":"Noise (video)","score":0.30000001192092896},{"id":"https://openalex.org/keywords/data-compression","display_name":"Data compression","score":0.2696000039577484},{"id":"https://openalex.org/keywords/subspace-topology","display_name":"Subspace topology","score":0.2648000121116638}],"concepts":[{"id":"https://openalex.org/C41008148","wikidata":"https://www.wikidata.org/wiki/Q21198","display_name":"Computer science","level":0,"score":0.4837999939918518},{"id":"https://openalex.org/C180016635","wikidata":"https://www.wikidata.org/wiki/Q2712821","display_name":"Compression (physics)","level":2,"score":0.39570000767707825},{"id":"https://openalex.org/C154945302","wikidata":"https://www.wikidata.org/wiki/Q11660","display_name":"Artificial intelligence","level":1,"score":0.3603000044822693},{"id":"https://openalex.org/C99498987","wikidata":"https://www.wikidata.org/wiki/Q2210247","display_name":"Noise (video)","level":3,"score":0.30000001192092896},{"id":"https://openalex.org/C31972630","wikidata":"https://www.wikidata.org/wiki/Q844240","display_name":"Computer vision","level":1,"score":0.29190000891685486},{"id":"https://openalex.org/C11413529","wikidata":"https://www.wikidata.org/wiki/Q8366","display_name":"Algorithm","level":1,"score":0.2906999886035919},{"id":"https://openalex.org/C121332964","wikidata":"https://www.wikidata.org/wiki/Q413","display_name":"Physics","level":0,"score":0.2773999869823456},{"id":"https://openalex.org/C78548338","wikidata":"https://www.wikidata.org/wiki/Q2493","display_name":"Data compression","level":2,"score":0.2696000039577484},{"id":"https://openalex.org/C32834561","wikidata":"https://www.wikidata.org/wiki/Q660730","display_name":"Subspace topology","level":2,"score":0.2648000121116638},{"id":"https://openalex.org/C116834253","wikidata":"https://www.wikidata.org/wiki/Q2039217","display_name":"Identification (biology)","level":2,"score":0.25060001015663147}],"mesh":[],"locations_count":1,"locations":[{"id":"doi:10.18653/v1/2026.acl-long.2214","is_oa":true,"landing_page_url":"https://doi.org/10.18653/v1/2026.acl-long.2214","pdf_url":"https://aclanthology.org/2026.acl-long.2214.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 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)","raw_type":"proceedings-article"}],"best_oa_location":{"id":"doi:10.18653/v1/2026.acl-long.2214","is_oa":true,"landing_page_url":"https://doi.org/10.18653/v1/2026.acl-long.2214","pdf_url":"https://aclanthology.org/2026.acl-long.2214.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 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)","raw_type":"proceedings-article"},"sustainable_development_goals":[],"awards":[],"funders":[],"has_content":{"grobid_xml":true,"pdf":true},"content_urls":{"pdf":"https://content.openalex.org/works/W7166854504.pdf","grobid_xml":"https://content.openalex.org/works/W7166854504.grobid-xml"},"referenced_works_count":0,"referenced_works":[],"related_works":[],"abstract_inverted_index":{"Post-training":[0],"activation":[1],"compression":[2,42],"is":[3,99],"essential":[4],"for":[5,69],"deploying":[6],"Large":[7],"Language":[8],"Models":[9],"(LLMs)":[10],"on":[11,31,105,116],"resource-constrained":[12],"hardware.However,":[13],"standard":[14],"methods":[15,124],"like":[16],"Singular":[17],"Value":[18],"Decomposition":[19],"(SVD)":[20],"are":[21],"gradient-blind:":[22],"they":[23],"preserve":[24],"high-variance":[25],"dimensions":[26,67],"regardless":[27],"of":[28,56,139,153],"their":[29],"impact":[30],"factual":[32,70,97,137,164],"knowledge":[33,98],"preservation.We":[34],"introduce":[35],"Fisher-Aligned":[36],"Subspace":[37],"Compression":[38],"(FASC),":[39],"a":[40,53,87,131,140,150],"knowledge-aware":[41],"framework":[43],"that":[44,91,110,146],"selects":[45],"subspaces":[46],"by":[47],"directly":[48],"modeling":[49],"activation-gradient":[50,93],"coupling,":[51,94],"minimizing":[52],"second-order":[54],"surrogate":[55],"the":[57,61,81,136],"loss":[58],"function.FASC":[59],"leverages":[60],"Fisher":[62],"Information":[63],"Matrix":[64],"to":[65,122,134],"identify":[66],"critical":[68],"knowledge,":[71,155],"which":[72],"often":[73],"reside":[74],"in":[75],"low-variance":[76],"but":[77],"high-gradient-sensitivity":[78],"subspaces.We":[79],"propose":[80],"Dependence":[82],"Violation":[83],"Score":[84],"(\u03c1)":[85],"as":[86,149],"general-purpose":[88],"diagnostic":[89],"metric":[90],"quantifies":[92],"revealing":[95],"where":[96],"stored":[100,154],"within":[101],"transformer":[102],"architectures.Extensive":[103],"experiments":[104],"Mistral-7B":[106],"and":[107],"Llama-3-8B":[108],"demonstrate":[109],"FASC":[111],"preserves":[112],"6-8%":[113],"more":[114],"accuracy":[115],"knowledge-intensive":[117],"benchmarks":[118],"(MMLU,":[119],"LAMA)":[120],"compared":[121],"variance-based":[123],"at":[125],"50%":[126],"rank":[127],"reduction,":[128],"effectively":[129],"enabling":[130],"7B":[132],"model":[133],"match":[135],"recall":[138],"13B":[141],"uncompressed":[142],"model.Our":[143],"analysis":[144],"reveals":[145],"\u03c1":[147],"serves":[148],"fundamental":[151],"signal":[152],"with":[156],"high-\u03c1":[157],"layers":[158],"emerging":[159],"only":[160],"when":[161],"models":[162],"internalize":[163],"associations":[165],"during":[166],"training.":[167]},"counts_by_year":[],"updated_date":"2026-07-29T14:22:42.915294","created_date":"2026-07-02T00:00:00"}