{"id":"https://openalex.org/W7167621636","doi":"https://doi.org/10.48550/arxiv.2607.03640","title":"Revealing Hidden Model Behaviors with Task-Specific Self-Reports","display_name":"Revealing Hidden Model Behaviors with Task-Specific Self-Reports","publication_year":2026,"publication_date":"2026-07-03","ids":{"openalex":"https://openalex.org/W7167621636","doi":"https://doi.org/10.48550/arxiv.2607.03640"},"language":null,"primary_location":{"id":"pmh:oai:arXiv.org:2607.03640","doi":null,"is_oa":true,"landing_page_url":"https://arxiv.org/abs/2607.03640","pdf_url":"https://arxiv.org/pdf/2607.03640","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":null,"license_id":null,"version":"submittedVersion","is_accepted":false,"is_published":false,"raw_source_name":null,"raw_type":"text"},"volume":null,"type":"preprint","indexed_in":["arxiv","datacite"],"open_access":{"is_oa":true,"oa_status":"green","oa_url":"https://arxiv.org/pdf/2607.03640","any_repository_has_fulltext":true},"authorships":[{"author_position":"first","author":{"id":"https://openalex.org/A5140182598","display_name":"Taras Kutsyk","orcid":null,"observed_orcids":[]},"institutions":[{"id":"https://openalex.org/I126596746","display_name":"Jagiellonian University","ror":"https://ror.org/03bqmcz70","country_code":"PL","type":"education","lineage":["https://openalex.org/I126596746"]}],"countries":["PL"],"is_corresponding":false,"raw_author_name":"Kutsyk, Taras","raw_affiliation_strings":["Jagiellonian University , Doctoral School of Exact and Natural Sciences","Jagiellonian University , Faculty of Mathematics and Computer Science"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"Jagiellonian University , Doctoral School of Exact and Natural Sciences","institution_ids":["https://openalex.org/I126596746"]},{"raw_affiliation_string":"Jagiellonian University , Faculty of Mathematics and Computer Science","institution_ids":["https://openalex.org/I126596746"]}]},{"author_position":"last","author":{"id":"https://openalex.org/A5058095594","display_name":"Bartosz Zieli\u0144ski","orcid":"https://orcid.org/0000-0002-3063-3621","observed_orcids":["https://orcid.org/0000-0002-3063-3621"]},"institutions":[{"id":"https://openalex.org/I126596746","display_name":"Jagiellonian University","ror":"https://ror.org/03bqmcz70","country_code":"PL","type":"education","lineage":["https://openalex.org/I126596746"]}],"countries":["PL"],"is_corresponding":false,"raw_author_name":"Zieli\u0144ski, Bartosz","raw_affiliation_strings":["Jagiellonian University , Faculty of Mathematics and Computer Science"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"Jagiellonian University , Faculty of Mathematics and Computer Science","institution_ids":["https://openalex.org/I126596746"]}]}],"institutions":[],"countries_distinct_count":1,"institutions_distinct_count":1,"corresponding_author_ids":[],"corresponding_institution_ids":["https://openalex.org/I126596746"],"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":0.6050000190734863,"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":0.6050000190734863,"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":0.3068999946117401,"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/T10883","display_name":"Ethics and Social Impacts of AI","score":0.026799999177455902,"subfield":{"id":"https://openalex.org/subfields/3311","display_name":"Safety Research"},"field":{"id":"https://openalex.org/fields/33","display_name":"Social Sciences"},"domain":{"id":"https://openalex.org/domains/2","display_name":"Social Sciences"}}],"keywords":[{"id":"https://openalex.org/keywords/self-report","display_name":"self-report","score":0.5989999771118164},{"id":"https://openalex.org/keywords/lora-adapters","display_name":"LoRA adapters","score":0.7239999771118164},{"id":"https://openalex.org/keywords/fine-tuned-language-models","display_name":"fine-tuned language models","score":0.8880000114440918},{"id":"https://openalex.org/keywords/model-introspection","display_name":"model introspection","score":0.7120000123977661},{"id":"https://openalex.org/keywords/model-auditing","display_name":"model auditing","score":0.4059999883174896},{"id":"https://openalex.org/keywords/hallucination-detection","display_name":"hallucination detection","score":0.6100000143051147},{"id":"https://openalex.org/keywords/misalignment","display_name":"misalignment","score":0.4690000116825104}],"concepts":[{"id":"https://openalex.org/C41008148","wikidata":"https://www.wikidata.org/wiki/Q21198","display_name":"Computer science","level":0,"score":0.6599000096321106},{"id":"https://openalex.org/C79581498","wikidata":"https://www.wikidata.org/wiki/Q1367530","display_name":"Suite","level":2,"score":0.5950000286102295},{"id":"https://openalex.org/C154945302","wikidata":"https://www.wikidata.org/wiki/Q11660","display_name":"Artificial intelligence","level":1,"score":0.5073999762535095},{"id":"https://openalex.org/C177284502","wikidata":"https://www.wikidata.org/wiki/Q1005390","display_name":"Adapter (computing)","level":2,"score":0.4832000136375427},{"id":"https://openalex.org/C2779955035","wikidata":"https://www.wikidata.org/wiki/Q4686785","display_name":"Advice (programming)","level":2,"score":0.3716000020503998},{"id":"https://openalex.org/C28490314","wikidata":"https://www.wikidata.org/wiki/Q189436","display_name":"Speech recognition","level":1,"score":0.34310001134872437},{"id":"https://openalex.org/C137293760","wikidata":"https://www.wikidata.org/wiki/Q3621696","display_name":"Language model","level":2,"score":0.33629998564720154},{"id":"https://openalex.org/C129671850","wikidata":"https://www.wikidata.org/wiki/Q210501","display_name":"Introspection","level":2,"score":0.33559998869895935},{"id":"https://openalex.org/C2779267917","wikidata":"https://www.wikidata.org/wiki/Q170028","display_name":"Deception","level":2,"score":0.31679999828338623},{"id":"https://openalex.org/C31972630","wikidata":"https://www.wikidata.org/wiki/Q844240","display_name":"Computer vision","level":1,"score":0.30169999599456787},{"id":"https://openalex.org/C199521495","wikidata":"https://www.wikidata.org/wiki/Q181487","display_name":"Audit","level":2,"score":0.2865999937057495},{"id":"https://openalex.org/C119857082","wikidata":"https://www.wikidata.org/wiki/Q2539","display_name":"Machine learning","level":1,"score":0.2863999903202057},{"id":"https://openalex.org/C2779843651","wikidata":"https://www.wikidata.org/wiki/Q7390335","display_name":"SIGNAL (programming language)","level":2,"score":0.2842999994754791},{"id":"https://openalex.org/C140547941","wikidata":"https://www.wikidata.org/wiki/Q7797194","display_name":"Threat model","level":2,"score":0.27410000562667847},{"id":"https://openalex.org/C67186912","wikidata":"https://www.wikidata.org/wiki/Q367664","display_name":"Data modeling","level":2,"score":0.2709999978542328},{"id":"https://openalex.org/C2779696439","wikidata":"https://www.wikidata.org/wiki/Q7512811","display_name":"Signature (topology)","level":2,"score":0.27079999446868896},{"id":"https://openalex.org/C77052588","wikidata":"https://www.wikidata.org/wiki/Q644307","display_name":"Constant false alarm rate","level":2,"score":0.2685000002384186},{"id":"https://openalex.org/C184337299","wikidata":"https://www.wikidata.org/wiki/Q1437428","display_name":"Semantics (computer science)","level":2,"score":0.2615000009536743},{"id":"https://openalex.org/C153180895","wikidata":"https://www.wikidata.org/wiki/Q7148389","display_name":"Pattern recognition (psychology)","level":2,"score":0.25929999351501465},{"id":"https://openalex.org/C26517878","wikidata":"https://www.wikidata.org/wiki/Q228039","display_name":"Key (lock)","level":2,"score":0.2563999891281128},{"id":"https://openalex.org/C23224414","wikidata":"https://www.wikidata.org/wiki/Q176769","display_name":"Hidden Markov model","level":2,"score":0.25119999051094055}],"mesh":[],"locations_count":2,"locations":[{"id":"pmh:oai:arXiv.org:2607.03640","doi":null,"is_oa":true,"landing_page_url":"https://arxiv.org/abs/2607.03640","pdf_url":"https://arxiv.org/pdf/2607.03640","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":null,"license_id":null,"version":"submittedVersion","is_accepted":false,"is_published":false,"raw_source_name":null,"raw_type":"text"},{"id":"doi:10.48550/arxiv.2607.03640","doi":"https://doi.org/10.48550/arxiv.2607.03640","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2607.03640","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":null,"license_id":null,"version":null,"is_accepted":false,"is_published":null,"raw_source_name":null,"raw_type":"Preprint"}],"best_oa_location":{"id":"pmh:oai:arXiv.org:2607.03640","doi":null,"is_oa":true,"landing_page_url":"https://arxiv.org/abs/2607.03640","pdf_url":"https://arxiv.org/pdf/2607.03640","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":null,"license_id":null,"version":"submittedVersion","is_accepted":false,"is_published":false,"raw_source_name":null,"raw_type":"text"},"sustainable_development_goals":[],"sustainable_development_goals_aurora":[{"score":0.5731571912765503,"display_name":"Peace, justice, and strong institutions","id":"https://metadata.un.org/sdg/16"}],"study_designs":[],"awards":[],"funders":[{"id":"https://openalex.org/F4320326815","display_name":"Infrastruktura PL-Grid","ror":"https://ror.org/01m3qaz74"},{"id":"https://openalex.org/F4320335596","display_name":"Wroclawskie Centrum Sieciowo-Superkomputerowe, Politechnika Wroclawska","ror":"https://ror.org/03fqwqm07"},{"id":"https://openalex.org/F4320337160","display_name":"Academic Computer Centre Cyfronet, AGH University of Science and Technology","ror":"https://ror.org/0575jwx13"}],"has_content":{"pdf":true,"grobid_xml":true},"content_urls":{"pdf":"https://content.openalex.org/works/W7167621636.pdf","grobid_xml":"https://content.openalex.org/works/W7167621636.grobid-xml"},"referenced_works_count":0,"referenced_works":[],"related_works":[],"abstract_inverted_index":{"Fine-tuning":[0],"can":[1],"give":[2,10,18],"a":[3,6,14,23,26,37,43,139,145],"language":[4],"model":[5,45,57,79,147],"hidden":[7,49,72],"behavior--it":[8],"may":[9],"false":[11],"answers":[12],"under":[13],"narrow":[15],"condition,":[16],"or":[17],"harmful":[19],"advice":[20],"only":[21,55],"when":[22,77],"prompt":[24],"touches":[25],"particular":[27],"topic.":[28],"We":[29],"introduce":[30],"the":[31,56,59,71,78,86,96,132],"Stabilized":[32],"Adapter":[33],"for":[34],"self-Report":[35],"(SAR),":[36],"lightweight":[38],"LoRA":[39],"adapter":[40],"that":[41,85],"makes":[42],"fine-tuned":[44,146],"describe":[46],"its":[47],"own":[48],"behavior":[50,73],"in":[51,74],"plain":[52],"language,":[53],"using":[54],"and":[58,129,148],"dataset":[60],"it":[61,111,113,152],"was":[62],"trained":[63],"on.":[64],"Across":[65],"seven":[66],"implanted":[67],"behaviors,":[68],"SAR":[69,119],"detects":[70,100],"every":[75,124],"one--even":[76],"has":[80],"generalized":[81],"into":[82],"broad":[83],"misalignment":[84],"training":[87],"data":[88],"alone":[89],"does":[90],"not":[91],"predict.":[92],"Introspection":[93],"Adapters":[94],"(IA),":[95],"closest":[97],"existing":[98],"baseline,":[99],"some":[101],"behaviors":[102],"from":[103],"our":[104],"suite":[105],"but":[106],"misses":[107],"others":[108],"entirely--and":[109],"where":[110,126],"misses,":[112],"hallucinates,":[114],"consistently":[115],"reporting":[116],"wrong":[117],"behaviors.":[118],"retains":[120],"positive":[121],"signal":[122],"on":[123],"setting":[125],"IA":[127],"fails":[128],"roughly":[130],"halves":[131],"rate":[133],"of":[134,156],"hallucinations.":[135],"This":[136],"gives":[137],"practitioners":[138],"more":[140],"reliable":[141],"tool":[142],"to":[143],"audit":[144],"answer":[149],"``what":[150],"did":[151],"actually":[153],"learn?''":[154],"type":[155],"questions.":[157]},"counts_by_year":[],"updated_date":"2026-10-15T10:00:46.086208","created_date":"2026-07-08T00:00:00"}