{"id":"https://openalex.org/W6929314784","doi":"https://doi.org/10.48550/arxiv.2507.22149","title":"When Truthful Representations Flip Under Deceptive Instructions?","display_name":"When Truthful Representations Flip Under Deceptive Instructions?","publication_year":2025,"publication_date":"2025-07-29","ids":{"openalex":"https://openalex.org/W6929314784","doi":"https://doi.org/10.48550/arxiv.2507.22149"},"language":"en","primary_location":{"id":"doi:10.48550/arxiv.2507.22149","doi":"https://doi.org/10.48550/arxiv.2507.22149","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2507.22149","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":false,"raw_source_name":null,"raw_type":"Preprint"},"volume":null,"type":"preprint","indexed_in":["datacite"],"open_access":{"is_oa":true,"oa_status":"green","oa_url":"https://doi.org/10.48550/arxiv.2507.22149","any_repository_has_fulltext":true},"authorships":[{"author_position":"first","author":{"id":null,"display_name":"Long, Xianxuan","orcid":null,"observed_orcids":[]},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Long, Xianxuan","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":null,"display_name":"Fu, Yao","orcid":null,"observed_orcids":[]},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Fu, Yao","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":null,"display_name":"Li, Runchao","orcid":null,"observed_orcids":[]},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Li, Runchao","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5120308840","display_name":"Mu Sheng","orcid":null,"observed_orcids":[]},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Sheng, Mu","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":null,"display_name":"Yu, Haotian","orcid":null,"observed_orcids":[]},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Yu, Haotian","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5101502420","display_name":"Xiaotian Han","orcid":"https://orcid.org/0000-0002-4018-4364","observed_orcids":["https://orcid.org/0000-0002-4018-4364"]},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Han, Xiaotian","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"last","author":{"id":null,"display_name":"Li, Pan","orcid":null,"observed_orcids":[]},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Li, Pan","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]}],"institutions":[],"countries_distinct_count":0,"institutions_distinct_count":0,"corresponding_author_ids":[],"corresponding_institution_ids":[],"apc_list":null,"apc_paid":null,"fwci":null,"has_fulltext":false,"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":true,"primary_topic":{"id":"https://openalex.org/T11689","display_name":"Adversarial Robustness in Machine Learning","score":0.8179000020027161,"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":0.8179000020027161,"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":0.11140000075101852,"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":0.03240000084042549,"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":0.8330000042915344},{"id":"https://openalex.org/keywords/internal-representations","display_name":"internal representations","score":0.8899999856948853},{"id":"https://openalex.org/keywords/linear-probing","display_name":"linear probing","score":0.6570000052452087},{"id":"https://openalex.org/keywords/sparse-autoencoders","display_name":"sparse autoencoders","score":0.9950000047683716},{"id":"https://openalex.org/keywords/fact-checking","display_name":"fact-checking","score":0.8899999856948853},{"id":"https://openalex.org/keywords/llama-3-1-8b-instruct","display_name":"Llama-3.1-8B-Instruct","score":0.7879999876022339},{"id":"https://openalex.org/keywords/deception-detection","display_name":"deception detection","score":0.7089999914169312},{"id":"https://openalex.org/keywords/layer-by-layer-analysis","display_name":"layer-wise analysis","score":0.421999990940094}],"concepts":[{"id":"https://openalex.org/C41008148","wikidata":"https://www.wikidata.org/wiki/Q21198","display_name":"Computer science","level":0,"score":0.6912999749183655},{"id":"https://openalex.org/C100776233","wikidata":"https://www.wikidata.org/wiki/Q2532492","display_name":"Bridge (graph theory)","level":2,"score":0.5005999803543091},{"id":"https://openalex.org/C2776359362","wikidata":"https://www.wikidata.org/wiki/Q2145286","display_name":"Representation (politics)","level":3,"score":0.4650999903678894},{"id":"https://openalex.org/C154945302","wikidata":"https://www.wikidata.org/wiki/Q11660","display_name":"Artificial intelligence","level":1,"score":0.45899999141693115},{"id":"https://openalex.org/C204321447","wikidata":"https://www.wikidata.org/wiki/Q30642","display_name":"Natural language processing","level":1,"score":0.4584999978542328},{"id":"https://openalex.org/C2778585151","wikidata":"https://www.wikidata.org/wiki/Q5282225","display_name":"Dishonesty","level":2,"score":0.43950000405311584},{"id":"https://openalex.org/C180747234","wikidata":"https://www.wikidata.org/wiki/Q23373","display_name":"Cognitive psychology","level":1,"score":0.4268999993801117},{"id":"https://openalex.org/C15744967","wikidata":"https://www.wikidata.org/wiki/Q9418","display_name":"Psychology","level":0,"score":0.4083000123500824},{"id":"https://openalex.org/C2776437466","wikidata":"https://www.wikidata.org/wiki/Q60920","display_name":"Lie detection","level":3,"score":0.391400009393692},{"id":"https://openalex.org/C107457646","wikidata":"https://www.wikidata.org/wiki/Q207434","display_name":"Human\u2013computer interaction","level":1,"score":0.353300005197525},{"id":"https://openalex.org/C188147891","wikidata":"https://www.wikidata.org/wiki/Q147638","display_name":"Cognitive science","level":1,"score":0.31619998812675476},{"id":"https://openalex.org/C26760741","wikidata":"https://www.wikidata.org/wiki/Q160402","display_name":"Perception","level":2,"score":0.2786000072956085},{"id":"https://openalex.org/C26517878","wikidata":"https://www.wikidata.org/wiki/Q228039","display_name":"Key (lock)","level":2,"score":0.25619998574256897},{"id":"https://openalex.org/C119421448","wikidata":"https://www.wikidata.org/wiki/Q2268776","display_name":"Lying","level":2,"score":0.25189998745918274},{"id":"https://openalex.org/C2779267917","wikidata":"https://www.wikidata.org/wiki/Q170028","display_name":"Deception","level":2,"score":0.2500999867916107}],"mesh":[],"locations_count":1,"locations":[{"id":"doi:10.48550/arxiv.2507.22149","doi":"https://doi.org/10.48550/arxiv.2507.22149","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2507.22149","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":"doi:10.48550/arxiv.2507.22149","doi":"https://doi.org/10.48550/arxiv.2507.22149","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2507.22149","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":false,"raw_source_name":null,"raw_type":"Preprint"},"sustainable_development_goals":[],"sustainable_development_goals_aurora":[{"score":0.6667523980140686,"display_name":"Peace, justice, and strong institutions","id":"https://metadata.un.org/sdg/16"}],"study_designs":[],"awards":[],"funders":[],"has_content":{"pdf":false,"grobid_xml":false},"content_urls":null,"referenced_works_count":0,"referenced_works":[],"related_works":[],"abstract_inverted_index":{"Large":[0],"language":[1],"models":[2],"(LLMs)":[3],"tend":[4],"to":[5,10,27,52,98,109,133,140],"follow":[6],"maliciously":[7],"crafted":[8],"instructions":[9,19,103],"generate":[11],"deceptive":[12,18,55,134],"responses,":[13],"posing":[14],"safety":[15],"challenges.":[16],"How":[17],"alter":[20],"the":[21,60,74,89,101],"internal":[22,61,90],"representations":[23,46,62,111],"of":[24,63,154,171],"LLM":[25,160],"compared":[26,108],"truthful":[28,51],"ones":[29],"remains":[30],"poorly":[31],"understood":[32],"beyond":[33],"output":[34,78],"analysis.":[35],"To":[36],"bridge":[37],"this":[38],"gap,":[39],"we":[40,72,93],"investigate":[41],"when":[42],"and":[43,65,119,136,151,162,168,178],"how":[44],"these":[45],"``flip'',":[47],"such":[48],"as":[49],"from":[50],"deceptive,":[53],"under":[54],"versus":[56],"truthful/neutral":[57],"instructions.":[58],"Analyzing":[59],"Llama-3.1-8B-Instruct":[64],"Gemma-2-9B-Instruct":[66],"on":[67,88,122],"a":[68],"factual":[69],"verification":[70],"task,":[71],"find":[73],"model's":[75],"instructed":[76,155,180],"True/False":[77],"is":[79],"predictable":[80],"via":[81],"linear":[82],"probes":[83],"across":[84],"all":[85],"conditions":[86],"based":[87],"representation.":[91],"Further,":[92],"use":[94,137],"Sparse":[95],"Autoencoders":[96],"(SAEs)":[97],"show":[99],"that":[100],"Deceptive":[102],"induce":[104],"significant":[105],"representational":[106,144],"shifts":[107],"Truthful/Neutral":[110],"(which":[112],"are":[113],"similar),":[114],"concentrated":[115],"in":[116,182],"early-to-mid":[117],"layers":[118],"detectable":[120],"even":[121],"complex":[123],"datasets.":[124],"We":[125],"also":[126],"identify":[127],"specific":[128],"SAE":[129],"features":[130],"highly":[131],"sensitive":[132],"instruction":[135],"targeted":[138],"visualizations":[139],"confirm":[141],"distinct":[142],"truthful/deceptive":[143],"subspaces.":[145],"%":[146],"Our":[147,164],"analysis":[148],"pinpoints":[149],"layer-wise":[150],"feature-level":[152],"correlates":[153],"dishonesty,":[156],"offering":[157,173],"insights":[158,175],"for":[159,176],"detection":[161],"control.":[163],"findings":[165],"expose":[166],"feature-":[167],"layer-level":[169],"signatures":[170],"deception,":[172],"new":[174],"detecting":[177],"mitigating":[179],"dishonesty":[181],"LLMs.":[183]},"counts_by_year":[],"updated_date":"2026-10-14T16:39:01.820549","created_date":"2025-10-10T00:00:00"}