{"id":"https://openalex.org/W7214638293","doi":"https://doi.org/10.48550/arxiv.2609.31603","title":"User Model Extraction via Belief Self-Distillation","display_name":"User Model Extraction via Belief Self-Distillation","publication_year":2026,"publication_date":"2026-09-25","ids":{"openalex":"https://openalex.org/W7214638293","doi":"https://doi.org/10.48550/arxiv.2609.31603"},"language":"en","primary_location":{"id":"pmh:oai:arXiv.org:2609.31603","doi":null,"is_oa":true,"landing_page_url":"https://arxiv.org/abs/2609.31603","pdf_url":"https://arxiv.org/pdf/2609.31603","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","is_traditional":true,"indexed_in":["arxiv","datacite"],"open_access":{"is_oa":true,"oa_status":"green","oa_url":"https://arxiv.org/pdf/2609.31603","any_repository_has_fulltext":true},"authorships":[{"author_position":"first","author":{"id":"https://openalex.org/A5137363149","display_name":"Ali Holmov","orcid":null,"observed_orcids":[]},"institutions":[{"id":"https://openalex.org/I62916508","display_name":"Technical University of Munich","ror":"https://ror.org/02kkvpp62","country_code":"DE","type":"education","lineage":["https://openalex.org/I62916508"]}],"countries":["DE"],"is_corresponding":false,"raw_author_name":"Holmov, Ali","raw_affiliation_strings":["Technical University of Munich"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"Technical University of Munich","institution_ids":["https://openalex.org/I62916508"]}]},{"author_position":"middle","author":{"id":"https://openalex.org/A5153212307","display_name":"Yiran Huang","orcid":null,"observed_orcids":[]},"institutions":[{"id":"https://openalex.org/I62916508","display_name":"Technical University of Munich","ror":"https://ror.org/02kkvpp62","country_code":"DE","type":"education","lineage":["https://openalex.org/I62916508"]}],"countries":["DE"],"is_corresponding":false,"raw_author_name":"Huang, Yiran","raw_affiliation_strings":["Technical University of Munich"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"Technical University of Munich","institution_ids":["https://openalex.org/I62916508"]}]},{"author_position":"middle","author":{"id":"https://openalex.org/A5070294236","display_name":"Kirill Bykov","orcid":"https://orcid.org/0000-0002-3358-2858","observed_orcids":["https://orcid.org/0000-0002-3358-2858"]},"institutions":[{"id":"https://openalex.org/I3018134672","display_name":"Helmholtz Munich","ror":"https://ror.org/00cfam450","country_code":"DE","type":"facility","lineage":["https://openalex.org/I1305996414","https://openalex.org/I3018134672"]},{"id":"https://openalex.org/I62916508","display_name":"Technical University of Munich","ror":"https://ror.org/02kkvpp62","country_code":"DE","type":"education","lineage":["https://openalex.org/I62916508"]}],"countries":["DE"],"is_corresponding":false,"raw_author_name":"Bykov, Kirill","raw_affiliation_strings":["Helmholtz Zentrum M\u00fcnchen https://github.com","Technical University of Munich"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"Helmholtz Zentrum M\u00fcnchen https://github.com","institution_ids":["https://openalex.org/I3018134672"]},{"raw_affiliation_string":"Technical University of Munich","institution_ids":["https://openalex.org/I62916508"]}]},{"author_position":"last","author":{"id":"https://openalex.org/A5040372929","display_name":"Zeynep Akata","orcid":"https://orcid.org/0000-0002-1432-7747","observed_orcids":["https://orcid.org/0000-0002-1432-7747"]},"institutions":[{"id":"https://openalex.org/I3018134672","display_name":"Helmholtz Munich","ror":"https://ror.org/00cfam450","country_code":"DE","type":"facility","lineage":["https://openalex.org/I1305996414","https://openalex.org/I3018134672"]},{"id":"https://openalex.org/I62916508","display_name":"Technical University of Munich","ror":"https://ror.org/02kkvpp62","country_code":"DE","type":"education","lineage":["https://openalex.org/I62916508"]}],"countries":["DE"],"is_corresponding":false,"raw_author_name":"Akata, Zeynep","raw_affiliation_strings":["Helmholtz Zentrum M\u00fcnchen https://github.com","Technical University of Munich"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"Helmholtz Zentrum M\u00fcnchen https://github.com","institution_ids":["https://openalex.org/I3018134672"]},{"raw_affiliation_string":"Technical University of Munich","institution_ids":["https://openalex.org/I62916508"]}]}],"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/T12026","display_name":"Explainable Artificial Intelligence (XAI)","score":0.9223999977111816,"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/T12026","display_name":"Explainable Artificial Intelligence (XAI)","score":0.9223999977111816,"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":0.027799999341368675,"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":0.02539999969303608,"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/large-language-models","display_name":"large language models","score":0.9259999990463257},{"id":"https://openalex.org/keywords/user-intention-recognition","display_name":"user intention recognition","score":0.5479999780654907},{"id":"https://openalex.org/keywords/refusal-behavior","display_name":"refusal behavior","score":0.7770000100135803},{"id":"https://openalex.org/keywords/ai-safety","display_name":"AI safety","score":0.7080000042915344},{"id":"https://openalex.org/keywords/self-distillation","display_name":"self-distillation","score":0.4339999854564667}],"concepts":[{"id":"https://openalex.org/C41008148","wikidata":"https://www.wikidata.org/wiki/Q21198","display_name":"Computer science","level":0,"score":0.6977999806404114},{"id":"https://openalex.org/C2776359362","wikidata":"https://www.wikidata.org/wiki/Q2145286","display_name":"Representation (politics)","level":3,"score":0.6438000202178955},{"id":"https://openalex.org/C11671645","wikidata":"https://www.wikidata.org/wiki/Q5054567","display_name":"Causal model","level":2,"score":0.544700026512146},{"id":"https://openalex.org/C154945302","wikidata":"https://www.wikidata.org/wiki/Q11660","display_name":"Artificial intelligence","level":1,"score":0.4595000147819519},{"id":"https://openalex.org/C107457646","wikidata":"https://www.wikidata.org/wiki/Q207434","display_name":"Human\u2013computer interaction","level":1,"score":0.42590001225471497},{"id":"https://openalex.org/C67712803","wikidata":"https://www.wikidata.org/wiki/Q7901853","display_name":"User modeling","level":3,"score":0.41519999504089355},{"id":"https://openalex.org/C64357122","wikidata":"https://www.wikidata.org/wiki/Q1149766","display_name":"Causality (physics)","level":2,"score":0.40950000286102295},{"id":"https://openalex.org/C48103436","wikidata":"https://www.wikidata.org/wiki/Q599031","display_name":"State (computer science)","level":2,"score":0.400299996137619},{"id":"https://openalex.org/C195324797","wikidata":"https://www.wikidata.org/wiki/Q33742","display_name":"Natural language","level":2,"score":0.39340001344680786},{"id":"https://openalex.org/C2776608160","wikidata":"https://www.wikidata.org/wiki/Q4785462","display_name":"Natural (archaeology)","level":2,"score":0.3693999946117401},{"id":"https://openalex.org/C204321447","wikidata":"https://www.wikidata.org/wiki/Q30642","display_name":"Natural language processing","level":1,"score":0.3671000003814697},{"id":"https://openalex.org/C2779439875","wikidata":"https://www.wikidata.org/wiki/Q1078276","display_name":"Natural language understanding","level":3,"score":0.35339999198913574},{"id":"https://openalex.org/C44291984","wikidata":"https://www.wikidata.org/wiki/Q1074173","display_name":"Question answering","level":2,"score":0.3375999927520752},{"id":"https://openalex.org/C168167062","wikidata":"https://www.wikidata.org/wiki/Q1117970","display_name":"Component (thermodynamics)","level":2,"score":0.3082999885082245},{"id":"https://openalex.org/C2985684807","wikidata":"https://www.wikidata.org/wiki/Q1513879","display_name":"Text generation","level":2,"score":0.30070000886917114},{"id":"https://openalex.org/C45493050","wikidata":"https://www.wikidata.org/wiki/Q7884934","display_name":"Unified Model","level":2,"score":0.29499998688697815},{"id":"https://openalex.org/C195818886","wikidata":"https://www.wikidata.org/wiki/Q5421724","display_name":"Expressive power","level":2,"score":0.29100000858306885},{"id":"https://openalex.org/C2776187449","wikidata":"https://www.wikidata.org/wiki/Q1513879","display_name":"Natural language generation","level":3,"score":0.2827000021934509},{"id":"https://openalex.org/C161301231","wikidata":"https://www.wikidata.org/wiki/Q3478658","display_name":"Knowledge representation and reasoning","level":2,"score":0.28200000524520874},{"id":"https://openalex.org/C180747234","wikidata":"https://www.wikidata.org/wiki/Q23373","display_name":"Cognitive psychology","level":1,"score":0.2741999924182892},{"id":"https://openalex.org/C59404180","wikidata":"https://www.wikidata.org/wiki/Q17013334","display_name":"Feature learning","level":2,"score":0.2644999921321869},{"id":"https://openalex.org/C137293760","wikidata":"https://www.wikidata.org/wiki/Q3621696","display_name":"Language model","level":2,"score":0.2628999948501587},{"id":"https://openalex.org/C115086926","wikidata":"https://www.wikidata.org/wiki/Q17004651","display_name":"Causal reasoning","level":3,"score":0.25999999046325684},{"id":"https://openalex.org/C195807954","wikidata":"https://www.wikidata.org/wiki/Q1662562","display_name":"Information extraction","level":2,"score":0.25940001010894775},{"id":"https://openalex.org/C175154964","wikidata":"https://www.wikidata.org/wiki/Q380077","display_name":"Task analysis","level":3,"score":0.2590000033378601},{"id":"https://openalex.org/C27415008","wikidata":"https://www.wikidata.org/wiki/Q7256382","display_name":"Psychological intervention","level":2,"score":0.2581999897956848},{"id":"https://openalex.org/C125411270","wikidata":"https://www.wikidata.org/wiki/Q18653","display_name":"Encoding (memory)","level":2,"score":0.25690001249313354}],"mesh":[],"locations_count":2,"locations":[{"id":"pmh:oai:arXiv.org:2609.31603","doi":null,"is_oa":true,"landing_page_url":"https://arxiv.org/abs/2609.31603","pdf_url":"https://arxiv.org/pdf/2609.31603","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.2609.31603","doi":"https://doi.org/10.48550/arxiv.2609.31603","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2609.31603","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:2609.31603","doi":null,"is_oa":true,"landing_page_url":"https://arxiv.org/abs/2609.31603","pdf_url":"https://arxiv.org/pdf/2609.31603","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":[],"study_designs":[],"awards":[],"funders":[],"has_content":{"pdf":true,"grobid_xml":true},"content_urls":{"pdf":"https://content.openalex.org/works/W7214638293.pdf","grobid_xml":"https://content.openalex.org/works/W7214638293.grobid-xml"},"referenced_works_count":10,"referenced_works":["https://openalex.org/W2971569798","https://openalex.org/W4399657295","https://openalex.org/W4402671231","https://openalex.org/W4405181600","https://openalex.org/W4414939355","https://openalex.org/W4417465430","https://openalex.org/W6910566006","https://openalex.org/W7160938135","https://openalex.org/W7166841348","https://openalex.org/W7202371739"],"related_works":["https://openalex.org/W7220869551","https://openalex.org/W7131113151","https://openalex.org/W7167580951","https://openalex.org/W7222742684","https://openalex.org/W7212036357","https://openalex.org/W7117982355","https://openalex.org/W7105801875","https://openalex.org/W7164695304","https://openalex.org/W7165478335","https://openalex.org/W4416035194"],"abstract_inverted_index":{"Large":[0],"language":[1],"models":[2,165,181],"(LLMs)":[3],"implicitly":[4],"infer":[5],"attributes":[6],"of":[7],"their":[8,12,157],"users":[9],"and":[10,22,37,51,103,168],"adapt":[11],"behavior":[13],"accordingly,":[14],"yet":[15],"these":[16,160],"beliefs":[17,66,102],"remain":[18],"difficult":[19],"to":[20],"inspect":[21],"causally":[23,169],"manipulate.":[24],"We":[25,140],"introduce":[26],"Belief":[27],"Self-Distillation":[28],"(BSD),":[29],"a":[30,42,85,143,152],"unified":[31],"read-write":[32],"framework":[33],"that":[34,46,115],"bridges":[35],"linear":[36],"causal":[38,88],"probing":[39],"by":[40],"learning":[41],"compact":[43],"user":[44,101,128,164],"representation":[45],"can":[47,90],"be":[48,91],"both":[49],"decoded":[50],"written":[52],"back":[53],"into":[54],"the":[55,121,125,137],"model.":[56],"The":[57],"frozen":[58],"LLM":[59],"acts":[60],"as":[61,166],"its":[62],"own":[63],"teacher,":[64],"distilling":[65],"from":[67],"natural":[68],"conversations":[69],"without":[70],"external":[71],"annotations.":[72],"Unlike":[73],"conventional":[74],"probing,":[75],"BSD":[76,98],"isolates":[77],"not":[78,118],"only":[79,119],"information":[80],"present":[81],"in":[82],"activations,":[83],"but":[84,123],"state":[86],"whose":[87],"role":[89],"directly":[92],"tested.":[93],"Across":[94],"multiple":[95],"model":[96],"families,":[97],"faithfully":[99],"recovers":[100],"enables":[104],"substantially":[105],"stronger":[106],"interventions":[107],"than":[108],"matched":[109],"hidden-state":[110],"steering.":[111],"Crucially,":[112],"we":[113],"find":[114],"refusal":[116,134],"depends":[117],"on":[120,124,151,185],"request,":[122],"model's":[126],"inferred":[127],"intent:":[129],"changing":[130],"this":[131],"belief":[132],"alters":[133],"while":[135],"holding":[136],"request":[138],"fixed.":[139],"further":[141],"uncover":[142],"striking":[144],"cross-model":[145],"regularity:":[146],"independently":[147],"trained":[148],"LLMs":[149],"converge":[150],"shared":[153],"geometry":[154],"for":[155,176],"representing":[156],"users.":[158],"Together,":[159],"results":[161],"reveal":[162],"implicit":[163],"readable":[167],"writable":[170],"internal":[171],"states":[172],"with":[173],"direct":[174],"implications":[175],"AI":[177],"safety,":[178],"shaping":[179],"how":[180],"condition":[182],"safety":[183],"decisions":[184],"whom":[186],"they":[187,189],"believe":[188],"are":[190],"interacting":[191],"with.":[192]},"counts_by_year":[],"updated_date":"2026-11-07T11:25:07.637180","created_date":"2026-09-29T00:00:00"}