{"id":"https://openalex.org/W7204180367","doi":"https://doi.org/10.48550/arxiv.2608.23497","title":"Mitigating Reasoning-Induced Misalignment via Safety-Direction Penalty","display_name":"Mitigating Reasoning-Induced Misalignment via Safety-Direction Penalty","publication_year":2026,"publication_date":"2026-08-24","ids":{"openalex":"https://openalex.org/W7204180367","doi":"https://doi.org/10.48550/arxiv.2608.23497"},"language":"en","primary_location":{"id":"pmh:oai:arXiv.org:2608.23497","doi":null,"is_oa":true,"landing_page_url":"https://arxiv.org/abs/2608.23497","pdf_url":"https://arxiv.org/pdf/2608.23497","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","indexed_in":["arxiv","datacite"],"open_access":{"is_oa":true,"oa_status":"green","oa_url":"https://arxiv.org/pdf/2608.23497","any_repository_has_fulltext":true},"authorships":[{"author_position":"first","author":{"id":"https://openalex.org/A5021199407","display_name":"Yipeng Zhao","orcid":"https://orcid.org/0009-0005-4583-6818","observed_orcids":["https://orcid.org/0009-0005-4583-6818"]},"institutions":[{"id":"https://openalex.org/I185261750","display_name":"University of Toronto","ror":"https://ror.org/03dbr7087","country_code":"CA","type":"education","lineage":["https://openalex.org/I185261750"]}],"countries":["CA"],"is_corresponding":false,"raw_author_name":"Zhao, Yipeng","raw_affiliation_strings":["University of Toronto"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"University of Toronto","institution_ids":["https://openalex.org/I185261750"]}]},{"author_position":"middle","author":{"id":"https://openalex.org/A5148122936","display_name":"Qishun Yang","orcid":null,"observed_orcids":[]},"institutions":[{"id":"https://openalex.org/I71920554","display_name":"King Abdullah University of Science and Technology","ror":"https://ror.org/01q3tbs38","country_code":"SA","type":"education","lineage":["https://openalex.org/I71920554"]}],"countries":["SA"],"is_corresponding":false,"raw_author_name":"Yang, Qishun","raw_affiliation_strings":["King Abdullah University of Science and Technology"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"King Abdullah University of Science and Technology","institution_ids":["https://openalex.org/I71920554"]}]},{"author_position":"middle","author":{"id":"https://openalex.org/A5138208742","display_name":"Shenzhe Zhu","orcid":null,"observed_orcids":[]},"institutions":[{"id":"https://openalex.org/I185261750","display_name":"University of Toronto","ror":"https://ror.org/03dbr7087","country_code":"CA","type":"education","lineage":["https://openalex.org/I185261750"]},{"id":"https://openalex.org/I86519309","display_name":"The University of Texas at Austin","ror":"https://ror.org/00hj54h04","country_code":"US","type":"education","lineage":["https://openalex.org/I86519309"]}],"countries":["CA","US"],"is_corresponding":false,"raw_author_name":"Zhu, Shenzhe","raw_affiliation_strings":["University of Texas at Austin","University of Toronto"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"University of Texas at Austin","institution_ids":["https://openalex.org/I86519309"]},{"raw_affiliation_string":"University of Toronto","institution_ids":["https://openalex.org/I185261750"]}]},{"author_position":"middle","author":{"id":"https://openalex.org/A5128817865","display_name":"Shu Yang","orcid":null,"observed_orcids":[]},"institutions":[{"id":"https://openalex.org/I71920554","display_name":"King Abdullah University of Science and Technology","ror":"https://ror.org/01q3tbs38","country_code":"SA","type":"education","lineage":["https://openalex.org/I71920554"]}],"countries":["SA"],"is_corresponding":false,"raw_author_name":"Yang, Shu","raw_affiliation_strings":["King Abdullah University of Science and Technology"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"King Abdullah University of Science and Technology","institution_ids":["https://openalex.org/I71920554"]}]},{"author_position":"last","author":{"id":"https://openalex.org/A5100401482","display_name":"Di Wang","orcid":"https://orcid.org/0000-0003-4908-0243","observed_orcids":["https://orcid.org/0000-0003-4908-0243","https://orcid.org/0009-0001-5741-6694"]},"institutions":[{"id":"https://openalex.org/I71920554","display_name":"King Abdullah University of Science and Technology","ror":"https://ror.org/01q3tbs38","country_code":"SA","type":"education","lineage":["https://openalex.org/I71920554"]}],"countries":["SA"],"is_corresponding":false,"raw_author_name":"Wang, Di","raw_affiliation_strings":["King Abdullah University of Science and Technology"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"King Abdullah University of Science and Technology","institution_ids":["https://openalex.org/I71920554"]}]}],"institutions":[],"countries_distinct_count":3,"institutions_distinct_count":3,"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":0.8877000212669373,"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":0.8877000212669373,"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":0.05380000174045563,"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"}},{"id":"https://openalex.org/T12026","display_name":"Explainable Artificial Intelligence (XAI)","score":0.023099999874830246,"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-model-safety","display_name":"large language model safety","score":0.7730000019073486},{"id":"https://openalex.org/keywords/chain-of-thought-fine-tuning","display_name":"chain-of-thought fine-tuning","score":0.8399999737739563},{"id":"https://openalex.org/keywords/cross-architecture-generalization","display_name":"cross-architecture generalization","score":0.5490000247955322},{"id":"https://openalex.org/keywords/qwen2-5","display_name":"Qwen2.5","score":0.7480000257492065}],"concepts":[{"id":"https://openalex.org/C107551265","wikidata":"https://www.wikidata.org/wiki/Q1458245","display_name":"Displacement (psychology)","level":2,"score":0.5967000126838684},{"id":"https://openalex.org/C41008148","wikidata":"https://www.wikidata.org/wiki/Q21198","display_name":"Computer science","level":0,"score":0.5458999872207642},{"id":"https://openalex.org/C185798385","wikidata":"https://www.wikidata.org/wiki/Q1161707","display_name":"Benchmark (surveying)","level":2,"score":0.5440999865531921},{"id":"https://openalex.org/C2776359362","wikidata":"https://www.wikidata.org/wiki/Q2145286","display_name":"Representation (politics)","level":3,"score":0.5116000175476074},{"id":"https://openalex.org/C125411270","wikidata":"https://www.wikidata.org/wiki/Q18653","display_name":"Encoding (memory)","level":2,"score":0.46650001406669617},{"id":"https://openalex.org/C18762648","wikidata":"https://www.wikidata.org/wiki/Q42213","display_name":"Work (physics)","level":2,"score":0.44780001044273376},{"id":"https://openalex.org/C2778012447","wikidata":"https://www.wikidata.org/wiki/Q1034415","display_name":"Scope (computer science)","level":2,"score":0.41100001335144043},{"id":"https://openalex.org/C2778572836","wikidata":"https://www.wikidata.org/wiki/Q380933","display_name":"Space (punctuation)","level":2,"score":0.4000999927520752},{"id":"https://openalex.org/C154945302","wikidata":"https://www.wikidata.org/wiki/Q11660","display_name":"Artificial intelligence","level":1,"score":0.3828999996185303},{"id":"https://openalex.org/C177264268","wikidata":"https://www.wikidata.org/wiki/Q1514741","display_name":"Set (abstract data type)","level":2,"score":0.32919999957084656},{"id":"https://openalex.org/C26517878","wikidata":"https://www.wikidata.org/wiki/Q228039","display_name":"Key (lock)","level":2,"score":0.3188000023365021},{"id":"https://openalex.org/C132835097","wikidata":"https://www.wikidata.org/wiki/Q7663745","display_name":"System safety","level":2,"score":0.3089999854564667},{"id":"https://openalex.org/C159694833","wikidata":"https://www.wikidata.org/wiki/Q2321565","display_name":"Iterative method","level":2,"score":0.2989000082015991},{"id":"https://openalex.org/C126042441","wikidata":"https://www.wikidata.org/wiki/Q1324888","display_name":"Frame (networking)","level":2,"score":0.2777999937534332},{"id":"https://openalex.org/C2778355321","wikidata":"https://www.wikidata.org/wiki/Q17079427","display_name":"Identity (music)","level":2,"score":0.2549000084400177},{"id":"https://openalex.org/C31972630","wikidata":"https://www.wikidata.org/wiki/Q844240","display_name":"Computer vision","level":1,"score":0.2547000050544739},{"id":"https://openalex.org/C2779227376","wikidata":"https://www.wikidata.org/wiki/Q6505497","display_name":"Layer (electronics)","level":2,"score":0.2500999867916107}],"mesh":[],"locations_count":2,"locations":[{"id":"pmh:oai:arXiv.org:2608.23497","doi":null,"is_oa":true,"landing_page_url":"https://arxiv.org/abs/2608.23497","pdf_url":"https://arxiv.org/pdf/2608.23497","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.2608.23497","doi":"https://doi.org/10.48550/arxiv.2608.23497","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2608.23497","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:2608.23497","doi":null,"is_oa":true,"landing_page_url":"https://arxiv.org/abs/2608.23497","pdf_url":"https://arxiv.org/pdf/2608.23497","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":{"pdf":true,"grobid_xml":true},"content_urls":{"pdf":"https://content.openalex.org/works/W7204180367.pdf","grobid_xml":"https://content.openalex.org/works/W7204180367.grobid-xml"},"referenced_works_count":0,"referenced_works":[],"related_works":["https://openalex.org/W4414878341","https://openalex.org/W7079542053","https://openalex.org/W7214592101","https://openalex.org/W7155451762","https://openalex.org/W4415082370","https://openalex.org/W7127083425","https://openalex.org/W7083810573","https://openalex.org/W4415024673","https://openalex.org/W4416548702","https://openalex.org/W7129660752"],"abstract_inverted_index":{"Reasoning-Induced":[0],"Misalignment,":[1],"where":[2,140],"fine-tuning":[3,115],"on":[4],"reasoning":[5,94,104,118,197],"data":[6],"containing":[7],"no":[8],"harmful":[9,21],"content,":[10],"including":[11],"mathematics,":[12],"code,":[13],"and":[14,37,80,106,122,134,162,189],"problem-solving":[15],"with":[16,124],"chain-of-thought":[17],"traces":[18],"can":[19],"induce":[20],"behaviors":[22],"of":[23,32,60,78,151],"LLM,":[24],"posing":[25],"a":[26,69,75,89],"serious":[27],"challenge":[28],"to":[29,51],"the":[30,58,61,81,107,137,149,153,159,163,167,171,178,181],"safety":[31,91,109,120,129,160,193],"LLM":[33],"reasoning.":[34],"Cross-architecture,":[35],"cross-scale,":[36],"cross-dataset":[38],"checks":[39],"show":[40,127],"that":[41,116],"RIM":[42,50,79],"does":[43],"not":[44,56],"always":[45],"emerge.":[46],"Previous":[47],"work":[48],"attributed":[49],"neuron-level":[52],"entanglement,":[53],"but":[54],"did":[55],"identify":[57],"geometry":[59],"representation":[62],"space":[63],"underlying":[64],"this":[65,141],"entanglement":[66],"or":[67],"propose":[68],"training-time":[70],"fix.":[71],"We":[72],"provide":[73],"both:":[74],"representation-space":[76],"analysis":[77,97],"Safety-Direction":[82],"Penalty":[83],"(SDP),":[84],"which":[85],"penalizes":[86],"movement":[87],"along":[88,158],"learned":[90],"direction":[92],"during":[93],"fine-tuning.":[95],"The":[96],"extracts":[98],"two":[99],"activation-space":[100],"directions,":[101],"one":[102],"encoding":[103],"ability":[105],"other":[108],"behavior.":[110],"These":[111,146],"directions":[112],"are":[113],"coupled:":[114],"improves":[117],"shifts":[119,126,176],"representations,":[121],"prompts":[123],"larger":[125,128],"degradation.":[130],"CKA":[131],"distance":[132],"ratios":[133],"probes":[135],"locate":[136],"safety-decision":[138],"layers":[139],"shift":[142],"is":[143],"most":[144],"relevant.":[145],"findings":[147],"guide":[148,184],"design":[150],"SDP:":[152],"coupling":[154],"motivates":[155],"penalizing":[156],"displacement":[157],"direction,":[161],"layer":[164],"localization":[165],"sets":[166],"initial":[168,172],"scope.":[169],"When":[170],"scope":[173],"leaves":[174],"compensatory":[175],"beyond":[177],"penalized":[179],"layers,":[180],"same":[182],"diagnostics":[183],"iterative":[185],"expansion.":[186],"On":[187],"Qwen2.5-3B":[188],"7B,":[190],"SDP":[191],"restores":[192],"while":[194],"preserving":[195],"benchmark":[196],"performance.":[198]},"counts_by_year":[],"updated_date":"2026-10-28T10:33:24.094632","created_date":"2026-08-26T00:00:00"}