{"id":"https://openalex.org/W4415257565","doi":"https://doi.org/10.48550/arxiv.2510.11905","title":"LLM Knowledge is Brittle: Truthfulness Representations Rely on Superficial Resemblance","display_name":"LLM Knowledge is Brittle: Truthfulness Representations Rely on Superficial Resemblance","publication_year":2025,"publication_date":"2025-10-13","ids":{"openalex":"https://openalex.org/W4415257565","doi":"https://doi.org/10.48550/arxiv.2510.11905"},"language":"en","primary_location":{"id":"pmh:oai:arXiv.org:2510.11905","doi":null,"is_oa":true,"landing_page_url":"http://arxiv.org/abs/2510.11905","pdf_url":"https://arxiv.org/pdf/2510.11905","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/2510.11905","any_repository_has_fulltext":true},"authorships":[{"author_position":"first","author":{"id":"https://openalex.org/A5002464117","display_name":"Patrick Haller","orcid":"https://orcid.org/0000-0002-8968-7587","observed_orcids":["https://orcid.org/0000-0002-8968-7587"]},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Haller, Patrick","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5030917007","display_name":"Mark Ibrahim","orcid":"https://orcid.org/0009-0005-2135-7476","observed_orcids":["https://orcid.org/0009-0005-2135-7476"]},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Ibrahim, Mark","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5084173427","display_name":"Polina Kirichenko","orcid":null,"observed_orcids":[]},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Kirichenko, Polina","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5071931582","display_name":"Levent Sagun","orcid":"https://orcid.org/0000-0001-5403-4124","observed_orcids":["https://orcid.org/0000-0001-5403-4124"]},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Sagun, Levent","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"last","author":{"id":"https://openalex.org/A5024504757","display_name":"Samuel J. Bell","orcid":"https://orcid.org/0000-0002-9437-5449","observed_orcids":["https://orcid.org/0000-0002-9437-5449"]},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Bell, Samuel J.","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":false,"primary_topic":{"id":"https://openalex.org/T10028","display_name":"Topic Modeling","score":0.8036999702453613,"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/T10028","display_name":"Topic Modeling","score":0.8036999702453613,"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":0.15080000460147858,"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":0.012799999676644802,"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":0.9300000071525574},{"id":"https://openalex.org/keywords/out-of-distribution-generalization","display_name":"out-of-distribution generalization","score":0.7739999890327454},{"id":"https://openalex.org/keywords/knowledge-probing","display_name":"knowledge probing","score":0.4779999852180481}],"concepts":[{"id":"https://openalex.org/C63479239","wikidata":"https://www.wikidata.org/wiki/Q7353546","display_name":"Robustness (evolution)","level":3,"score":0.6966999769210815},{"id":"https://openalex.org/C2777026412","wikidata":"https://www.wikidata.org/wiki/Q2684591","display_name":"Statement (logic)","level":2,"score":0.6660000085830688},{"id":"https://openalex.org/C41008148","wikidata":"https://www.wikidata.org/wiki/Q21198","display_name":"Computer science","level":0,"score":0.5649999976158142},{"id":"https://openalex.org/C154945302","wikidata":"https://www.wikidata.org/wiki/Q11660","display_name":"Artificial intelligence","level":1,"score":0.4271000027656555},{"id":"https://openalex.org/C204321447","wikidata":"https://www.wikidata.org/wiki/Q30642","display_name":"Natural language processing","level":1,"score":0.4018000066280365},{"id":"https://openalex.org/C2776359362","wikidata":"https://www.wikidata.org/wiki/Q2145286","display_name":"Representation (politics)","level":3,"score":0.39809998869895935},{"id":"https://openalex.org/C2777152325","wikidata":"https://www.wikidata.org/wiki/Q108163","display_name":"Proposition","level":2,"score":0.3718000054359436},{"id":"https://openalex.org/C2986065213","wikidata":"https://www.wikidata.org/wiki/Q743861","display_name":"Implicit knowledge","level":2,"score":0.33959999680519104},{"id":"https://openalex.org/C15744967","wikidata":"https://www.wikidata.org/wiki/Q9418","display_name":"Psychology","level":0,"score":0.3172000050544739},{"id":"https://openalex.org/C161301231","wikidata":"https://www.wikidata.org/wiki/Q3478658","display_name":"Knowledge representation and reasoning","level":2,"score":0.30720001459121704},{"id":"https://openalex.org/C185798385","wikidata":"https://www.wikidata.org/wiki/Q1161707","display_name":"Benchmark (surveying)","level":2,"score":0.301800012588501},{"id":"https://openalex.org/C2777267654","wikidata":"https://www.wikidata.org/wiki/Q3519023","display_name":"Test (biology)","level":2,"score":0.28929999470710754},{"id":"https://openalex.org/C2779525943","wikidata":"https://www.wikidata.org/wiki/Q1187300","display_name":"Grammaticality","level":3,"score":0.2858000099658966},{"id":"https://openalex.org/C202532154","wikidata":"https://www.wikidata.org/wiki/Q4374193","display_name":"Problem statement","level":2,"score":0.2824000120162964},{"id":"https://openalex.org/C180747234","wikidata":"https://www.wikidata.org/wiki/Q23373","display_name":"Cognitive psychology","level":1,"score":0.2800000011920929},{"id":"https://openalex.org/C66746571","wikidata":"https://www.wikidata.org/wiki/Q1134833","display_name":"ENCODE","level":3,"score":0.2614000141620636},{"id":"https://openalex.org/C111472728","wikidata":"https://www.wikidata.org/wiki/Q9471","display_name":"Epistemology","level":1,"score":0.2535000145435333}],"mesh":[],"locations_count":2,"locations":[{"id":"pmh:oai:arXiv.org:2510.11905","doi":null,"is_oa":true,"landing_page_url":"http://arxiv.org/abs/2510.11905","pdf_url":"https://arxiv.org/pdf/2510.11905","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.2510.11905","doi":"https://doi.org/10.48550/arxiv.2510.11905","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2510.11905","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"}],"best_oa_location":{"id":"pmh:oai:arXiv.org:2510.11905","doi":null,"is_oa":true,"landing_page_url":"http://arxiv.org/abs/2510.11905","pdf_url":"https://arxiv.org/pdf/2510.11905","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":false,"grobid_xml":false},"content_urls":null,"referenced_works_count":0,"referenced_works":[],"related_works":["https://openalex.org/W4416812383","https://openalex.org/W4387561538","https://openalex.org/W7053377175","https://openalex.org/W7221401661","https://openalex.org/W4411875187","https://openalex.org/W7134954275","https://openalex.org/W4388032052","https://openalex.org/W4389519449","https://openalex.org/W4399405791","https://openalex.org/W4403928636"],"abstract_inverted_index":{"For":[0],"Large":[1],"Language":[2],"Models":[3],"(LLMs)":[4],"to":[5,44,113,166],"be":[6,15,37,87],"reliable,":[7],"they":[8,182],"must":[9],"learn":[10,211],"robust":[11],"knowledge":[12,63,101,146,214,247],"that":[13,33,75,108,152,216],"can":[14,36,86,173],"generally":[16],"applied":[17],"in":[18],"diverse":[19],"settings":[20],"--":[21,81],"often":[22,174],"unlike":[23],"those":[24,167],"seen":[25,168],"during":[26,169],"training.":[27],"Yet,":[28],"extensive":[29],"research":[30,240],"has":[31],"shown":[32],"LLM":[34,76,139],"performance":[35],"brittle,":[38],"with":[39],"models":[40],"exhibiting":[41],"excessive":[42],"sensitivity":[43],"trivial":[45],"input":[46],"variations.":[47],"In":[48],"this":[49,54,67,188],"work,":[50],"we":[51,69,95,127],"explore":[52,66],"whether":[53],"brittleness":[55],"is":[56,190],"a":[57,202,225],"direct":[58],"result":[59],"of":[60,99,155,231,245],"unstable":[61],"internal":[62,153],"representations.":[64,248],"To":[65],"question,":[68],"build":[70],"on":[71,106,193,241],"previous":[72],"work":[73,223],"showing":[74],"representations":[77,154,215],"encode":[78],"statement":[79,156],"truthfulness":[80,157,232],"i.e.,":[82],"true,":[83],"factual":[84],"statements":[85,133,180],"easily":[88],"separated":[89],"from":[90],"false,":[91],"inaccurate":[92],"ones.":[93],"Specifically,":[94],"test":[96],"the":[97,160,185,194,229,243],"robustness":[98,244],"learned":[100,246],"by":[102],"evaluating":[103],"representation":[104],"separability":[105,130],"samples":[107],"have":[109],"undergone":[110],"superficial":[111],"transformations":[112],"drive":[114],"them":[115],"out-of-distribution":[116],"(OOD),":[117],"such":[118],"as":[119,132,159],"typos":[120],"or":[121],"reformulations.":[122],"By":[123],"applying":[124],"semantically-preserving":[125],"perturbations,":[126],"study":[128],"how":[129],"degrades":[131],"become":[134,163],"more":[135,235],"OOD,":[136],"across":[137],"four":[138],"families,":[140],"five":[141],"evaluation":[142],"datasets,":[143],"and":[144,178,234],"three":[145],"probing":[147],"methods.":[148],"Our":[149,222],"results":[150],"reveal":[151],"collapse":[158],"samples'":[161],"presentations":[162],"less":[164],"similar":[165],"pre-training.":[170],"While":[171],"LLMs":[172,209],"distinguish":[175],"between":[176],"true":[177],"false":[179],"when":[181],"closely":[183],"resemble":[184],"pre-training":[186],"data,":[187],"ability":[189],"highly":[191],"dependent":[192],"statement's":[195],"exact":[196],"surface":[197],"form.":[198],"These":[199],"findings":[200],"offer":[201],"possible":[203],"explanation":[204],"for":[205,218,228,238],"brittle":[206],"benchmark":[207],"performance:":[208],"may":[210],"shallow,":[212],"non-robust":[213],"allow":[217],"only":[219],"limited":[220],"generalizability.":[221],"presents":[224],"fundamental":[226],"challenge":[227],"utility":[230],"probes,":[233],"broadly,":[236],"calls":[237],"further":[239],"improving":[242]},"counts_by_year":[],"updated_date":"2026-08-26T07:47:46.906454","created_date":"2025-10-17T00:00:00"}