{"id":"https://openalex.org/W7140146812","doi":"https://doi.org/10.18653/v1/2026.findings-eacl.96","title":"Feature Drift: How Fine-Tuning Repurposes Representations in LLMs","display_name":"Feature Drift: How Fine-Tuning Repurposes Representations in LLMs","publication_year":2026,"publication_date":"2026-01-01","ids":{"openalex":"https://openalex.org/W7140146812","doi":"https://doi.org/10.18653/v1/2026.findings-eacl.96"},"language":null,"primary_location":{"id":"doi:10.18653/v1/2026.findings-eacl.96","is_oa":true,"landing_page_url":"https://doi.org/10.18653/v1/2026.findings-eacl.96","pdf_url":"https://aclanthology.org/2026.findings-eacl.96.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: EACL 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-eacl.96.pdf","any_repository_has_fulltext":null},"authorships":[{"author_position":"first","author":{"id":"https://openalex.org/A5092698837","display_name":"Andrey V. Galichin","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Andrey V. Galichin","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5126264909","display_name":"Anton Korznikov","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Anton Korznikov","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5130336588","display_name":"Alexey Dontsov","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Alexey Dontsov","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5048793164","display_name":"Oleg Y. Rogov","orcid":"https://orcid.org/0000-0001-9672-2427"},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Oleg Rogov","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5130339990","display_name":"Elena Tutubalina","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Elena Tutubalina","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"last","author":{"id":"https://openalex.org/A5130411263","display_name":"Ivan Oseledets","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Ivan Oseledets","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":0.26220875,"is_in_top_1_percent":false,"is_in_top_10_percent":false},"cited_by_percentile_year":null,"biblio":{"volume":null,"issue":null,"first_page":"1878","last_page":"1887"},"is_retracted":false,"is_paratext":false,"is_xpac":false,"primary_topic":{"id":"https://openalex.org/T10181","display_name":"Natural Language Processing Techniques","score":0.10729999840259552,"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":0.10729999840259552,"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/T12761","display_name":"Data Stream Mining Techniques","score":0.09040000289678574,"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/T11975","display_name":"Evolutionary Algorithms and Applications","score":0.07779999822378159,"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/feature","display_name":"Feature (linguistics)","score":0.5249999761581421},{"id":"https://openalex.org/keywords/representation","display_name":"Representation (politics)","score":0.32899999618530273},{"id":"https://openalex.org/keywords/pattern-recognition","display_name":"Pattern recognition (psychology)","score":0.2957000136375427},{"id":"https://openalex.org/keywords/frame","display_name":"Frame (networking)","score":0.28839999437332153},{"id":"https://openalex.org/keywords/interpretation","display_name":"Interpretation (philosophy)","score":0.27300000190734863}],"concepts":[{"id":"https://openalex.org/C2776401178","wikidata":"https://www.wikidata.org/wiki/Q12050496","display_name":"Feature (linguistics)","level":2,"score":0.5249999761581421},{"id":"https://openalex.org/C154945302","wikidata":"https://www.wikidata.org/wiki/Q11660","display_name":"Artificial intelligence","level":1,"score":0.5181999802589417},{"id":"https://openalex.org/C41008148","wikidata":"https://www.wikidata.org/wiki/Q21198","display_name":"Computer science","level":0,"score":0.5084999799728394},{"id":"https://openalex.org/C2776359362","wikidata":"https://www.wikidata.org/wiki/Q2145286","display_name":"Representation (politics)","level":3,"score":0.32899999618530273},{"id":"https://openalex.org/C153180895","wikidata":"https://www.wikidata.org/wiki/Q7148389","display_name":"Pattern recognition (psychology)","level":2,"score":0.2957000136375427},{"id":"https://openalex.org/C126042441","wikidata":"https://www.wikidata.org/wiki/Q1324888","display_name":"Frame (networking)","level":2,"score":0.28839999437332153},{"id":"https://openalex.org/C204321447","wikidata":"https://www.wikidata.org/wiki/Q30642","display_name":"Natural language processing","level":1,"score":0.2734000086784363},{"id":"https://openalex.org/C527412718","wikidata":"https://www.wikidata.org/wiki/Q855395","display_name":"Interpretation (philosophy)","level":2,"score":0.27300000190734863},{"id":"https://openalex.org/C192209626","wikidata":"https://www.wikidata.org/wiki/Q190909","display_name":"Focus (optics)","level":2,"score":0.26759999990463257},{"id":"https://openalex.org/C9652623","wikidata":"https://www.wikidata.org/wiki/Q190109","display_name":"Field (mathematics)","level":2,"score":0.26750001311302185}],"mesh":[],"locations_count":1,"locations":[{"id":"doi:10.18653/v1/2026.findings-eacl.96","is_oa":true,"landing_page_url":"https://doi.org/10.18653/v1/2026.findings-eacl.96","pdf_url":"https://aclanthology.org/2026.findings-eacl.96.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: EACL 2026","raw_type":"proceedings-article"}],"best_oa_location":{"id":"doi:10.18653/v1/2026.findings-eacl.96","is_oa":true,"landing_page_url":"https://doi.org/10.18653/v1/2026.findings-eacl.96","pdf_url":"https://aclanthology.org/2026.findings-eacl.96.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: EACL 2026","raw_type":"proceedings-article"},"sustainable_development_goals":[{"id":"https://metadata.un.org/sdg/13","score":0.7391107678413391,"display_name":"Climate action"}],"awards":[],"funders":[{"id":"https://openalex.org/F4320324261","display_name":"National Research University Higher School of Economics","ror":"https://ror.org/055f7t516"}],"has_content":{"pdf":true,"grobid_xml":true},"content_urls":{"pdf":"https://content.openalex.org/works/W7140146812.pdf","grobid_xml":"https://content.openalex.org/works/W7140146812.grobid-xml"},"referenced_works_count":0,"referenced_works":[],"related_works":[],"abstract_inverted_index":{"Fine-tuning":[0],"LLMs":[1],"introduces":[2],"many":[3],"important":[4],"behaviors,":[5],"such":[6],"as":[7,65],"instruction-following":[8],"and":[9,101],"safety":[10],"alignment.This":[11],"makes":[12],"it":[13],"crucial":[14],"to":[15,51,56,127],"study":[16],"how":[17,148,166],"fine-tuning":[18,83,167],"changes":[19,168],"models'":[20],"internal":[21],"mechanisms.Sparse":[22],"Autoencoders":[23],"(SAEs)":[24],"offer":[25],"a":[26,115,161],"powerful":[27],"tool":[28],"for":[29,155],"interpreting":[30],"neural":[31],"networks":[32],"by":[33,94],"extracting":[34],"concepts":[35,86],"(features)":[36],"represented":[37],"in":[38,132],"their":[39],"activations.Previous":[40],"work":[41],"observed":[42],"that":[43,125,140,146],"SAEs":[44,97],"trained":[45],"on":[46,129],"base":[47,96,100,122],"models":[48],"transfer":[49],"effectively":[50],"instruction-tuned":[52],"(chat)":[53],"models,":[54],"attributed":[55],"activation":[57],"similarity.In":[58],"this":[59,93],"work,":[60],"we":[61,120],"propose":[62],"feature":[63,70,78],"drift":[64,113,126],"an":[66],"alternative":[67],"explanation:":[68],"the":[69,75],"space":[71],"remains":[72],"relevant,":[73],"but":[74],"distribution":[76],"of":[77,165],"activations":[79,103],"changes.In":[80],"other":[81],"words,":[82],"recombines":[84],"existing":[85,149],"rather":[87,152],"than":[88,153],"learning":[89],"new":[90,157],"ones.We":[91],"validate":[92],"showing":[95],"reconstruct":[98],"both":[99],"chat":[102,133],"comparably":[104],"despite":[105],"systematic":[106],"differences,":[107],"with":[108],"individual":[109],"features":[110,124,138,150],"exhibiting":[111],"clear":[112],"patterns.In":[114],"refusal":[116],"behavior":[117],"case":[118],"study,":[119],"identify":[121],"SAE":[123],"activate":[128],"harmful":[130],"instructions":[131],"models.Causal":[134],"interventions":[135],"using":[136],"these":[137],"confirm":[139],"they":[141],"mediate":[142],"refusal.Our":[143],"findings":[144],"suggest":[145],"monitoring":[147],"drift,":[151],"searching":[154],"entirely":[156],"features,":[158],"may":[159],"provide":[160],"more":[162],"complete":[163],"explanation":[164],"model":[169],"capabilities.":[170]},"counts_by_year":[],"updated_date":"2026-07-29T14:22:42.915294","created_date":"2026-03-24T00:00:00"}