{"id":"https://openalex.org/W7214718045","doi":"https://doi.org/10.48550/arxiv.2609.34970","title":"See it, Say it, Sorted: Mechanistic Diagnosis and Parameter-Space Mitigation of Emergent Misalignment in LLMs","display_name":"See it, Say it, Sorted: Mechanistic Diagnosis and Parameter-Space Mitigation of Emergent Misalignment in LLMs","publication_year":2026,"publication_date":"2026-09-28","ids":{"openalex":"https://openalex.org/W7214718045","doi":"https://doi.org/10.48550/arxiv.2609.34970"},"language":"en","primary_location":{"id":"pmh:oai:arXiv.org:2609.34970","doi":null,"is_oa":true,"landing_page_url":"https://arxiv.org/abs/2609.34970","pdf_url":"https://arxiv.org/pdf/2609.34970","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/2609.34970","any_repository_has_fulltext":true},"authorships":[{"author_position":"first","author":{"id":"https://openalex.org/A5108848915","display_name":"Wenjun Que","orcid":null,"observed_orcids":[]},"institutions":[{"id":"https://openalex.org/I66867065","display_name":"East China Normal University","ror":"https://ror.org/02n96ep67","country_code":"CN","type":"education","lineage":["https://openalex.org/I66867065"]}],"countries":["CN"],"is_corresponding":false,"raw_author_name":"Que, Weiqiao","raw_affiliation_strings":["School of Computer Science and Technology , East China Normal University , China"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"School of Computer Science and Technology , East China Normal University , China","institution_ids":["https://openalex.org/I66867065"]}]},{"author_position":"middle","author":{"id":"https://openalex.org/A5153475767","display_name":"Ruizhe Li","orcid":null,"observed_orcids":[]},"institutions":[{"id":"https://openalex.org/I79619799","display_name":"University of Birmingham","ror":"https://ror.org/03angcq70","country_code":"GB","type":"education","lineage":["https://openalex.org/I79619799"]}],"countries":["GB"],"is_corresponding":false,"raw_author_name":"Li, Ruizhe","raw_affiliation_strings":["School of Computer Science , University of Birmingham , UK"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"School of Computer Science , University of Birmingham , UK","institution_ids":["https://openalex.org/I79619799"]}]},{"author_position":"middle","author":{"id":"https://openalex.org/A5153546261","display_name":"Chengyu Wang","orcid":null,"observed_orcids":[]},"institutions":[{"id":"https://openalex.org/I45928872","display_name":"Alibaba Group (China)","ror":"https://ror.org/00k642b80","country_code":"CN","type":"company","lineage":["https://openalex.org/I45928872"]}],"countries":["CN"],"is_corresponding":false,"raw_author_name":"Wang, Chengyu","raw_affiliation_strings":["Alibaba Group , China"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"Alibaba Group , China","institution_ids":["https://openalex.org/I45928872"]}]},{"author_position":"middle","author":{"id":"https://openalex.org/A5017516574","display_name":"Dakan Wang","orcid":null,"observed_orcids":[]},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Wang, Dakan","raw_affiliation_strings":["Exacity Inc"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"Exacity Inc","institution_ids":[]}]},{"author_position":"middle","author":{"id":"https://openalex.org/A5076265623","display_name":"Emine Y\u0131lmaz","orcid":"https://orcid.org/0000-0003-4734-4532","observed_orcids":["https://orcid.org/0000-0003-4734-4532","https://orcid.org/0000-0002-3434-8932"]},"institutions":[{"id":"https://openalex.org/I45129253","display_name":"University College London","ror":"https://ror.org/02jx3x895","country_code":"GB","type":"education","lineage":["https://openalex.org/I124357947","https://openalex.org/I45129253"]}],"countries":["GB"],"is_corresponding":false,"raw_author_name":"Yilmaz, Emine","raw_affiliation_strings":["Centre for Artificial Intelligence , University College London , UK"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"Centre for Artificial Intelligence , University College London , UK","institution_ids":["https://openalex.org/I45129253"]}]},{"author_position":"last","author":{"id":"https://openalex.org/A5153494543","display_name":"Xiaofeng He","orcid":null,"observed_orcids":[]},"institutions":[{"id":"https://openalex.org/I66867065","display_name":"East China Normal University","ror":"https://ror.org/02n96ep67","country_code":"CN","type":"education","lineage":["https://openalex.org/I66867065"]}],"countries":["CN"],"is_corresponding":false,"raw_author_name":"He, Xiaofeng","raw_affiliation_strings":["School of Computer Science and Technology , East China Normal University , China"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"School of Computer Science and Technology , East China Normal University , China","institution_ids":["https://openalex.org/I66867065"]}]}],"institutions":[],"countries_distinct_count":2,"institutions_distinct_count":4,"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":0.9570000171661377,"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":0.9570000171661377,"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":0.017999999225139618,"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":0.006599999964237213,"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/emergent-misalignment","display_name":"emergent misalignment","score":0.996999979019165},{"id":"https://openalex.org/keywords/large-language-models","display_name":"large language models","score":0.890999972820282},{"id":"https://openalex.org/keywords/safety-alignment","display_name":"safety alignment","score":0.6420000195503235},{"id":"https://openalex.org/keywords/training-dynamics","display_name":"training dynamics","score":0.5929999947547913},{"id":"https://openalex.org/keywords/domain-adaptation","display_name":"domain adaptation","score":0.5149999856948853}],"concepts":[{"id":"https://openalex.org/C32834561","wikidata":"https://www.wikidata.org/wiki/Q660730","display_name":"Subspace topology","level":2,"score":0.6373999714851379},{"id":"https://openalex.org/C127705205","wikidata":"https://www.wikidata.org/wiki/Q5748245","display_name":"Heuristics","level":2,"score":0.5670999884605408},{"id":"https://openalex.org/C203616005","wikidata":"https://www.wikidata.org/wiki/Q620495","display_name":"Hessian matrix","level":2,"score":0.5069000124931335},{"id":"https://openalex.org/C41008148","wikidata":"https://www.wikidata.org/wiki/Q21198","display_name":"Computer science","level":0,"score":0.4553999900817871},{"id":"https://openalex.org/C2776196576","wikidata":"https://www.wikidata.org/wiki/Q196113","display_name":"Camouflage","level":2,"score":0.42410001158714294},{"id":"https://openalex.org/C36503486","wikidata":"https://www.wikidata.org/wiki/Q11235244","display_name":"Domain (mathematical analysis)","level":2,"score":0.3889999985694885},{"id":"https://openalex.org/C154945302","wikidata":"https://www.wikidata.org/wiki/Q11660","display_name":"Artificial intelligence","level":1,"score":0.38830000162124634},{"id":"https://openalex.org/C2777363581","wikidata":"https://www.wikidata.org/wiki/Q15098235","display_name":"Harm","level":2,"score":0.3801000118255615},{"id":"https://openalex.org/C12713177","wikidata":"https://www.wikidata.org/wiki/Q1900281","display_name":"Perspective (graphical)","level":2,"score":0.3675999939441681},{"id":"https://openalex.org/C139807058","wikidata":"https://www.wikidata.org/wiki/Q352374","display_name":"Adaptation (eye)","level":2,"score":0.36230000853538513},{"id":"https://openalex.org/C195065555","wikidata":"https://www.wikidata.org/wiki/Q214881","display_name":"Curvature","level":2,"score":0.3587000072002411},{"id":"https://openalex.org/C184047640","wikidata":"https://www.wikidata.org/wiki/Q182593","display_name":"Illusion","level":2,"score":0.35519999265670776},{"id":"https://openalex.org/C112313634","wikidata":"https://www.wikidata.org/wiki/Q7886648","display_name":"Complement (music)","level":5,"score":0.3409000039100647},{"id":"https://openalex.org/C13662910","wikidata":"https://www.wikidata.org/wiki/Q193139","display_name":"Trajectory","level":2,"score":0.31929999589920044},{"id":"https://openalex.org/C112930515","wikidata":"https://www.wikidata.org/wiki/Q4389547","display_name":"Risk analysis (engineering)","level":1,"score":0.3093999922275543},{"id":"https://openalex.org/C180747234","wikidata":"https://www.wikidata.org/wiki/Q23373","display_name":"Cognitive psychology","level":1,"score":0.3027999997138977},{"id":"https://openalex.org/C2775936607","wikidata":"https://www.wikidata.org/wiki/Q466845","display_name":"Tracking (education)","level":2,"score":0.301800012588501},{"id":"https://openalex.org/C149782125","wikidata":"https://www.wikidata.org/wiki/Q160039","display_name":"Econometrics","level":1,"score":0.2856999933719635},{"id":"https://openalex.org/C2779662365","wikidata":"https://www.wikidata.org/wiki/Q5416694","display_name":"Event (particle physics)","level":2,"score":0.28139999508857727},{"id":"https://openalex.org/C166052673","wikidata":"https://www.wikidata.org/wiki/Q83021","display_name":"Empirical evidence","level":2,"score":0.2777000069618225},{"id":"https://openalex.org/C162324750","wikidata":"https://www.wikidata.org/wiki/Q8134","display_name":"Economics","level":0,"score":0.27320000529289246},{"id":"https://openalex.org/C119857082","wikidata":"https://www.wikidata.org/wiki/Q2539","display_name":"Machine learning","level":1,"score":0.27239999175071716},{"id":"https://openalex.org/C38652104","wikidata":"https://www.wikidata.org/wiki/Q3510521","display_name":"Computer security","level":1,"score":0.2696000039577484},{"id":"https://openalex.org/C120936955","wikidata":"https://www.wikidata.org/wiki/Q2155640","display_name":"Empirical research","level":2,"score":0.26510000228881836},{"id":"https://openalex.org/C2777211547","wikidata":"https://www.wikidata.org/wiki/Q17141490","display_name":"Training (meteorology)","level":2,"score":0.26350000500679016},{"id":"https://openalex.org/C95713431","wikidata":"https://www.wikidata.org/wiki/Q631425","display_name":"Vulnerability (computing)","level":2,"score":0.2581000030040741}],"mesh":[],"locations_count":2,"locations":[{"id":"pmh:oai:arXiv.org:2609.34970","doi":null,"is_oa":true,"landing_page_url":"https://arxiv.org/abs/2609.34970","pdf_url":"https://arxiv.org/pdf/2609.34970","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.2609.34970","doi":"https://doi.org/10.48550/arxiv.2609.34970","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2609.34970","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:2609.34970","doi":null,"is_oa":true,"landing_page_url":"https://arxiv.org/abs/2609.34970","pdf_url":"https://arxiv.org/pdf/2609.34970","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":{"grobid_xml":true,"pdf":true},"content_urls":{"pdf":"https://content.openalex.org/works/W7214718045.pdf","grobid_xml":"https://content.openalex.org/works/W7214718045.grobid-xml"},"referenced_works_count":10,"referenced_works":["https://openalex.org/W4405655184","https://openalex.org/W4415068980","https://openalex.org/W4415108257","https://openalex.org/W7078232827","https://openalex.org/W7106695004","https://openalex.org/W7127283896","https://openalex.org/W7157609348","https://openalex.org/W7161110339","https://openalex.org/W7163992305","https://openalex.org/W7166807399"],"related_works":["https://openalex.org/W4415193668","https://openalex.org/W7166847658","https://openalex.org/W7204883081","https://openalex.org/W7106165751","https://openalex.org/W4415344850","https://openalex.org/W4392077870","https://openalex.org/W7161759866","https://openalex.org/W7214517898","https://openalex.org/W7160904105","https://openalex.org/W7165447259"],"abstract_inverted_index":{"Safety-aligned":[0],"LLMs":[1],"can":[2],"exhibit":[3],"emergent":[4],"misalignment":[5],"(EM):":[6],"narrow":[7],"domain":[8],"adaptation":[9],"unexpectedly":[10],"triggers":[11],"catastrophic":[12],"safety":[13],"failures":[14],"across":[15,104],"unrelated":[16],"domains.":[17],"Prior":[18],"static":[19],"analyses":[20],"leave":[21],"training":[22,44],"dynamics":[23],"unmapped,":[24],"while":[25],"existing":[26],"defenses":[27],"rely":[28],"on":[29,53],"heuristics":[30],"that":[31,47,82],"degrade":[32],"utility.":[33],"We":[34],"present":[35],"a":[36,77],"dynamic,":[37],"second-order":[38],"geometric":[39],"study":[40],"of":[41,90,108,133,143],"EM.":[42],"Tracking":[43],"trajectories":[45],"reveals":[46],"directional":[48],"Hessian":[49],"curvature":[50],"concentrates":[51],"sharply":[52],"semantic":[54],"pivot":[55],"tokens.":[56],"Grassmannian":[57],"projections":[58],"show":[59],"that,":[60],"in":[61,153],"most":[62],"settings,":[63],"harmful-safe":[64],"gap":[65],"widens":[66],"mainly":[67],"because":[68],"safe-gradient":[69],"overlap":[70],"declines.":[71],"Leveraging":[72],"these":[73,138],"insights,":[74],"we":[75],"introduce":[76],"parameter-level":[78],"Geometric":[79],"Mitigation":[80],"Framework":[81],"orthogonally":[83],"projects":[84],"empirical":[85],"harmful":[86,127],"gradient":[87],"subspace":[88,128,148],"out":[89],"parameter":[91],"updates.":[92],"On":[93],"Qwen2.5-14B-IT,":[94],"our":[95],"defense":[96],"suppresses":[97],"free-generation":[98],"EM":[99,118,135,157],"by":[100],"up":[101],"to":[102],"80.0%;":[103],"the":[105,130,141,146],"other":[106],"three":[107],"four":[109],"open-weight":[110],"instruction-based":[111],"model":[112],"families":[113],"(3B--20B),":[114],"where":[115,155],"single-layer":[116],"behavioral":[117,144,156],"is":[119,158],"already":[120],"near":[121,159],"zero,":[122],"teacher-forced":[123],"evaluation":[124],"shows":[125],"same":[126,147],"controls":[129],"conditional":[131],"support":[132],"frozen":[134],"responses.":[136],"Crucially,":[137],"diagnostics":[139],"unmask":[140],"illusion":[142],"safety:":[145],"remains":[149],"measurable":[150],"and":[151],"steerable":[152],"models":[154],"zero.":[160],"Code:":[161],"https://github.com/WeiqiaoQUE/mechanistic-emergent-misalignment.":[162]},"counts_by_year":[],"updated_date":"2026-11-07T11:25:07.637180","created_date":"2026-09-30T00:00:00"}