{"id":"https://openalex.org/W7223710171","doi":"https://doi.org/10.48550/arxiv.2610.32866","title":"Interpretable sparse fine-tuning with TopKLoRA enables faithful circuit discovery","display_name":"Interpretable sparse fine-tuning with TopKLoRA enables faithful circuit discovery","publication_year":2026,"publication_date":"2026-10-23","ids":{"openalex":"https://openalex.org/W7223710171","doi":"https://doi.org/10.48550/arxiv.2610.32866"},"language":"en","primary_location":{"id":"pmh:oai:arXiv.org:2610.32866","doi":null,"is_oa":true,"landing_page_url":"https://arxiv.org/abs/2610.32866","pdf_url":"https://arxiv.org/pdf/2610.32866","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,"listed_in":[],"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","license_id":"https://openalex.org/licenses/cc-by","version":"submittedVersion","is_accepted":false,"is_published":false,"raw_source_name":null,"raw_type":"text"},"volume":null,"type":"preprint","is_traditional":true,"indexed_in":["arxiv","datacite"],"open_access":{"is_oa":true,"oa_status":"green","oa_url":"https://arxiv.org/pdf/2610.32866","any_repository_has_fulltext":true},"authorships":[{"author_position":"first","author":{"id":"https://openalex.org/A5159588008","display_name":"Marek Masiak","orcid":null,"observed_orcids":[]},"institutions":[{"id":"https://openalex.org/I40120149","display_name":"University of Oxford","ror":"https://ror.org/052gg0110","country_code":"GB","type":"education","lineage":["https://openalex.org/I40120149"]}],"countries":["GB"],"is_corresponding":false,"raw_author_name":"Masiak, Marek","raw_affiliation_strings":["Department of Engineering Science University of Oxford"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"Department of Engineering Science University of Oxford","institution_ids":["https://openalex.org/I40120149"]}]},{"author_position":"middle","author":{"id":"https://openalex.org/A5015878484","display_name":"Andrzej Szablewski","orcid":"https://orcid.org/0000-0002-0084-9966","observed_orcids":["https://orcid.org/0000-0002-0084-9966"]},"institutions":[{"id":"https://openalex.org/I241749","display_name":"University of Cambridge","ror":"https://ror.org/013meh722","country_code":"GB","type":"education","lineage":["https://openalex.org/I241749"]},{"id":"https://openalex.org/I40120149","display_name":"University of Oxford","ror":"https://ror.org/052gg0110","country_code":"GB","type":"education","lineage":["https://openalex.org/I40120149"]}],"countries":["GB"],"is_corresponding":false,"raw_author_name":"Szablewski, Andrzej","raw_affiliation_strings":["Department of Computer Science University of Cambridge","Department of Engineering Science University of Oxford"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"Department of Computer Science University of Cambridge","institution_ids":["https://openalex.org/I241749"]},{"raw_affiliation_string":"Department of Engineering Science University of Oxford","institution_ids":["https://openalex.org/I40120149"]}]},{"author_position":"middle","author":{"id":"https://openalex.org/A5004751764","display_name":"Nicola Cancedda","orcid":null,"observed_orcids":[]},"institutions":[{"id":"https://openalex.org/I40120149","display_name":"University of Oxford","ror":"https://ror.org/052gg0110","country_code":"GB","type":"education","lineage":["https://openalex.org/I40120149"]}],"countries":["GB"],"is_corresponding":false,"raw_author_name":"Cancedda, Nicola","raw_affiliation_strings":["Department of Engineering Science University of Oxford"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"Department of Engineering Science University of Oxford","institution_ids":["https://openalex.org/I40120149"]}]},{"author_position":"last","author":{"id":"https://openalex.org/A5092958466","display_name":"Constantin Venhoff","orcid":null,"observed_orcids":[]},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Venhoff, Constantin","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]}],"institutions":[],"countries_distinct_count":1,"institutions_distinct_count":2,"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":null,"last_page":null},"is_retracted":false,"is_paratext":false,"is_xpac":false,"primary_topic":{"id":"https://openalex.org/T11689","display_name":"Adversarial Robustness in Machine Learning","score":0.7753999829292297,"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/T11689","display_name":"Adversarial Robustness in Machine Learning","score":0.7753999829292297,"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/T12026","display_name":"Explainable Artificial Intelligence (XAI)","score":0.2084999978542328,"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":0.0010000000474974513,"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/backdoor-detection","display_name":"backdoor detection","score":0.6579999923706055},{"id":"https://openalex.org/keywords/circuit-discovery","display_name":"circuit discovery","score":0.6970000267028809},{"id":"https://openalex.org/keywords/parameter-efficient-fine-tuning","display_name":"parameter-efficient fine-tuning","score":0.796999990940094},{"id":"https://openalex.org/keywords/model-organisms","display_name":"model organisms","score":0.5299999713897705},{"id":"https://openalex.org/keywords/safety-training","display_name":"safety training","score":0.4650000035762787}],"concepts":[{"id":"https://openalex.org/C2781045450","wikidata":"https://www.wikidata.org/wiki/Q254569","display_name":"Backdoor","level":2,"score":0.9739000201225281},{"id":"https://openalex.org/C41008148","wikidata":"https://www.wikidata.org/wiki/Q21198","display_name":"Computer science","level":0,"score":0.7577999830245972},{"id":"https://openalex.org/C177284502","wikidata":"https://www.wikidata.org/wiki/Q1005390","display_name":"Adapter (computing)","level":2,"score":0.711899995803833},{"id":"https://openalex.org/C2780801425","wikidata":"https://www.wikidata.org/wiki/Q5164392","display_name":"Construct (python library)","level":2,"score":0.42899999022483826},{"id":"https://openalex.org/C154945302","wikidata":"https://www.wikidata.org/wiki/Q11660","display_name":"Artificial intelligence","level":1,"score":0.4269999861717224},{"id":"https://openalex.org/C134146338","wikidata":"https://www.wikidata.org/wiki/Q1815901","display_name":"Electronic circuit","level":2,"score":0.4106000065803528},{"id":"https://openalex.org/C124101348","wikidata":"https://www.wikidata.org/wiki/Q172491","display_name":"Data mining","level":1,"score":0.3946000039577484},{"id":"https://openalex.org/C207850805","wikidata":"https://www.wikidata.org/wiki/Q269608","display_name":"Reverse engineering","level":2,"score":0.3707999885082245},{"id":"https://openalex.org/C160920958","wikidata":"https://www.wikidata.org/wiki/Q7662746","display_name":"Synthetic data","level":2,"score":0.34049999713897705},{"id":"https://openalex.org/C119857082","wikidata":"https://www.wikidata.org/wiki/Q2539","display_name":"Machine learning","level":1,"score":0.3393999934196472},{"id":"https://openalex.org/C11413529","wikidata":"https://www.wikidata.org/wiki/Q8366","display_name":"Algorithm","level":1,"score":0.321399986743927},{"id":"https://openalex.org/C165696696","wikidata":"https://www.wikidata.org/wiki/Q11287","display_name":"Exploit","level":2,"score":0.31130000948905945},{"id":"https://openalex.org/C153180895","wikidata":"https://www.wikidata.org/wiki/Q7148389","display_name":"Pattern recognition (psychology)","level":2,"score":0.3005000054836273},{"id":"https://openalex.org/C32946077","wikidata":"https://www.wikidata.org/wiki/Q618079","display_name":"Network analysis","level":2,"score":0.2856999933719635},{"id":"https://openalex.org/C2779089604","wikidata":"https://www.wikidata.org/wiki/Q7169333","display_name":"Permission","level":2,"score":0.28130000829696655},{"id":"https://openalex.org/C80444323","wikidata":"https://www.wikidata.org/wiki/Q2878974","display_name":"Theoretical computer science","level":1,"score":0.2653000056743622}],"mesh":[],"locations_count":2,"locations":[{"id":"pmh:oai:arXiv.org:2610.32866","doi":null,"is_oa":true,"landing_page_url":"https://arxiv.org/abs/2610.32866","pdf_url":"https://arxiv.org/pdf/2610.32866","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,"listed_in":[],"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","license_id":"https://openalex.org/licenses/cc-by","version":"submittedVersion","is_accepted":false,"is_published":false,"raw_source_name":null,"raw_type":"text"},{"id":"doi:10.48550/arxiv.2610.32866","doi":"https://doi.org/10.48550/arxiv.2610.32866","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2610.32866","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,"listed_in":[],"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","license_id":"https://openalex.org/licenses/cc-by","version":null,"is_accepted":false,"is_published":false,"raw_source_name":null,"raw_type":"Preprint"}],"best_oa_location":{"id":"pmh:oai:arXiv.org:2610.32866","doi":null,"is_oa":true,"landing_page_url":"https://arxiv.org/abs/2610.32866","pdf_url":"https://arxiv.org/pdf/2610.32866","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,"listed_in":[],"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","license_id":"https://openalex.org/licenses/cc-by","version":"submittedVersion","is_accepted":false,"is_published":false,"raw_source_name":null,"raw_type":"text"},"sustainable_development_goals":[],"study_designs":[],"awards":[],"funders":[],"has_content":{"grobid_xml":true,"pdf":true},"content_urls":{"pdf":"https://content.openalex.org/works/W7223710171.pdf","grobid_xml":"https://content.openalex.org/works/W7223710171.grobid-xml"},"referenced_works_count":19,"referenced_works":["https://openalex.org/W2594633041","https://openalex.org/W3010694149","https://openalex.org/W3168867926","https://openalex.org/W4308023630","https://openalex.org/W4367701241","https://openalex.org/W4393336283","https://openalex.org/W4403322569","https://openalex.org/W4404783717","https://openalex.org/W4405621156","https://openalex.org/W4415068980","https://openalex.org/W4415796540","https://openalex.org/W4415798500","https://openalex.org/W4415798729","https://openalex.org/W4417091950","https://openalex.org/W4417143258","https://openalex.org/W7083687746","https://openalex.org/W7104916054","https://openalex.org/W7133199409","https://openalex.org/W7133218086"],"related_works":["https://openalex.org/W7117173611","https://openalex.org/W7170285508","https://openalex.org/W7162667472","https://openalex.org/W7140042152","https://openalex.org/W7116736472","https://openalex.org/W7171966741","https://openalex.org/W7140323808","https://openalex.org/W7123903022","https://openalex.org/W4415068600","https://openalex.org/W7083687746"],"abstract_inverted_index":{"Fine-tuning":[0],"on":[1,87,121],"datasets":[2],"of":[3,53,59,159,170,202,221],"unknown":[4],"origin":[5],"can":[6,26],"create":[7,112],"a":[8,105,122,133,163,181,212],"backdoor":[9,107],"in":[10,68,104],"the":[11,15,36,50,56,71,143,171,200],"trained":[12,85],"model":[13,113,185,225],"when":[14],"data":[16],"includes":[17],"poisoned":[18,123],"examples.":[19],"Difficult":[20],"to":[21,141,156,183,198],"localise":[22],"and":[23,119,131,139,193],"remove,":[24],"backdoors":[25],"persist":[27],"through":[28],"safety":[29],"training.":[30],"While":[31],"parameter-efficient":[32],"fine-tuning":[33,116,148],"methods":[34],"capture":[35],"full":[37,51],"parameter":[38],"changes":[39],"introduced":[40],"during":[41],"training,":[42],"they":[43],"are":[44,76,154],"not":[45],"easily":[46],"interpretable.":[47],"We":[48,98,207,227],"present":[49],"method":[52,149],"TopKLoRA,":[54],"extending":[55],"preliminary":[57],"proposal":[58],"Masiak":[60],"et":[61],"al.":[62],"(2025),":[63],"an":[64,157],"interpretable-by-design":[65,215],"LoRA":[66,102,165],"variant":[67],"which":[69,195],"only":[70],"$k$":[72],"largest":[73],"adapter":[74,103],"latents":[75,192],"active":[77],"at":[78],"each":[79,127],"token.":[80],"Unlike":[81],"SAEs,":[82],"TopKLoRA":[83,191],"is":[84,136],"purely":[86],"downstream":[88],"loss,":[89],"where":[90,110],"optimisation":[91],"pressure":[92],"incentivises":[93],"learning":[94],"functionally":[95],"causal":[96],"latents.":[97],"demonstrate":[99],"this":[100],"sparse":[101,151],"sleeper-agent":[106],"case":[108],"study":[109],"we":[111,129,179,196],"organisms":[114,186],"by":[115],"gemma-2-2b,":[117],"Qwen2.5-1.5B,":[118],"Qwen2.5-7B":[120],"instruction-following":[124,176],"dataset.":[125],"For":[126],"organism,":[128],"identify":[130],"ablate":[132],"circuit":[134,204],"that":[135,153,217],"both":[137],"necessary":[138],"sufficient":[140],"implement":[142],"backdoor.":[144],"Furthermore,":[145],"our":[146,203,229],"new":[147],"produces":[150],"circuits":[152,189],"up":[155],"order":[158],"magnitude":[160],"smaller":[161],"than":[162],"dense":[164],"baseline,":[166],"allowing":[167],"surgical":[168],"ablation":[169],"malicious":[172],"capability":[173],"without":[174,223],"sacrificing":[175],"performance.":[177,226],"Finally,":[178],"propose":[180],"technique":[182],"construct":[184],"with":[187],"localised":[188],"using":[190],"gradient-routing,":[194],"use":[197],"validate":[199],"findings":[201,210],"search":[205],"algorithm.":[206],"consider":[208],"these":[209],"as":[211],"step":[213],"toward":[214],"architectures":[216],"provide":[218],"faithful":[219],"units":[220],"analysis":[222],"compromising":[224],"share":[228],"code:":[230],"https://github.com/interpretable-finetuning/TopKLoRA.":[231]},"counts_by_year":[],"updated_date":"2026-11-10T11:39:43.021295","created_date":"2026-10-27T00:00:00"}