{"id":"https://openalex.org/W7214050872","doi":"https://doi.org/10.48550/arxiv.2609.22337","title":"When and Why Do Linear Bias Probes Fail? A Geometric and Statistical Theory of Bias Detectability in Large Language Model Representations","display_name":"When and Why Do Linear Bias Probes Fail? A Geometric and Statistical Theory of Bias Detectability in Large Language Model Representations","publication_year":2026,"publication_date":"2026-07-21","ids":{"openalex":"https://openalex.org/W7214050872","doi":"https://doi.org/10.48550/arxiv.2609.22337"},"language":"en","primary_location":{"id":"pmh:oai:arXiv.org:2609.22337","doi":null,"is_oa":true,"landing_page_url":"https://arxiv.org/abs/2609.22337","pdf_url":"https://arxiv.org/pdf/2609.22337","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/2609.22337","any_repository_has_fulltext":true},"authorships":[{"author_position":"first","author":{"id":"https://openalex.org/A5141104442","display_name":"Mo Hai","orcid":null,"observed_orcids":[]},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Hai, Mo","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"last","author":{"id":"https://openalex.org/A5130133365","display_name":"Haifeng Li","orcid":null,"observed_orcids":[]},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Li, Haifeng","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":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/T10883","display_name":"Ethics and Social Impacts of AI","score":0.9526000022888184,"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/T10883","display_name":"Ethics and Social Impacts of AI","score":0.9526000022888184,"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":0.023000000044703484,"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/T12262","display_name":"Hate Speech and Cyberbullying Detection","score":0.004100000020116568,"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/linear-probing","display_name":"linear probing","score":0.968999981880188},{"id":"https://openalex.org/keywords/large-language-models","display_name":"large language models","score":0.8960000276565552},{"id":"https://openalex.org/keywords/hidden-representations","display_name":"hidden representations","score":0.8389999866485596},{"id":"https://openalex.org/keywords/counterfactual-evaluation","display_name":"counterfactual evaluation","score":0.7139999866485596},{"id":"https://openalex.org/keywords/manifold-curvature","display_name":"manifold curvature","score":0.6769999861717224},{"id":"https://openalex.org/keywords/generalization-bounds","display_name":"generalization bounds","score":0.6299999952316284},{"id":"https://openalex.org/keywords/minimax-lower-bounds","display_name":"minimax lower bounds","score":0.8809999823570251},{"id":"https://openalex.org/keywords/power-analysis","display_name":"power analysis","score":0.6669999957084656},{"id":"https://openalex.org/keywords/sample-size-calculation","display_name":"sample size determination","score":0.5019999742507935}],"concepts":[{"id":"https://openalex.org/C33923547","wikidata":"https://www.wikidata.org/wiki/Q395","display_name":"Mathematics","level":0,"score":0.6306999921798706},{"id":"https://openalex.org/C177148314","wikidata":"https://www.wikidata.org/wiki/Q170084","display_name":"Generalization","level":2,"score":0.6258999705314636},{"id":"https://openalex.org/C149728462","wikidata":"https://www.wikidata.org/wiki/Q751319","display_name":"Minimax","level":2,"score":0.6129999756813049},{"id":"https://openalex.org/C195065555","wikidata":"https://www.wikidata.org/wiki/Q214881","display_name":"Curvature","level":2,"score":0.5690000057220459},{"id":"https://openalex.org/C165064840","wikidata":"https://www.wikidata.org/wiki/Q1321061","display_name":"Matching (statistics)","level":2,"score":0.5496000051498413},{"id":"https://openalex.org/C1921717","wikidata":"https://www.wikidata.org/wiki/Q1334846","display_name":"Mahalanobis distance","level":2,"score":0.47690001130104065},{"id":"https://openalex.org/C529865628","wikidata":"https://www.wikidata.org/wiki/Q1790740","display_name":"Manifold (fluid mechanics)","level":2,"score":0.42750000953674316},{"id":"https://openalex.org/C114289077","wikidata":"https://www.wikidata.org/wiki/Q3284399","display_name":"Statistical model","level":2,"score":0.3995000123977661},{"id":"https://openalex.org/C163258240","wikidata":"https://www.wikidata.org/wiki/Q25342","display_name":"Power (physics)","level":2,"score":0.38670000433921814},{"id":"https://openalex.org/C163175372","wikidata":"https://www.wikidata.org/wiki/Q3339222","display_name":"Linear model","level":2,"score":0.3806000053882599},{"id":"https://openalex.org/C11413529","wikidata":"https://www.wikidata.org/wiki/Q8366","display_name":"Algorithm","level":1,"score":0.37400001287460327},{"id":"https://openalex.org/C137293760","wikidata":"https://www.wikidata.org/wiki/Q3621696","display_name":"Language model","level":2,"score":0.3666999936103821},{"id":"https://openalex.org/C202632270","wikidata":"https://www.wikidata.org/wiki/Q7798106","display_name":"Threshold model","level":2,"score":0.35600000619888306},{"id":"https://openalex.org/C101454708","wikidata":"https://www.wikidata.org/wiki/Q17106019","display_name":"Standard Model (mathematical formulation)","level":3,"score":0.352400004863739},{"id":"https://openalex.org/C149629883","wikidata":"https://www.wikidata.org/wiki/Q660926","display_name":"Fraction (chemistry)","level":2,"score":0.3377000093460083},{"id":"https://openalex.org/C96608239","wikidata":"https://www.wikidata.org/wiki/Q1199823","display_name":"Statistical power","level":2,"score":0.3361999988555908},{"id":"https://openalex.org/C178635117","wikidata":"https://www.wikidata.org/wiki/Q747499","display_name":"RADIUS","level":2,"score":0.3361000120639801},{"id":"https://openalex.org/C28826006","wikidata":"https://www.wikidata.org/wiki/Q33521","display_name":"Applied mathematics","level":1,"score":0.334199994802475},{"id":"https://openalex.org/C105795698","wikidata":"https://www.wikidata.org/wiki/Q12483","display_name":"Statistics","level":1,"score":0.33410000801086426},{"id":"https://openalex.org/C73586568","wikidata":"https://www.wikidata.org/wiki/Q2600211","display_name":"Parameter space","level":2,"score":0.32989999651908875},{"id":"https://openalex.org/C77553402","wikidata":"https://www.wikidata.org/wiki/Q13222579","display_name":"Upper and lower bounds","level":2,"score":0.3260999917984009},{"id":"https://openalex.org/C2778572836","wikidata":"https://www.wikidata.org/wiki/Q380933","display_name":"Space (punctuation)","level":2,"score":0.3221000134944916},{"id":"https://openalex.org/C2780586882","wikidata":"https://www.wikidata.org/wiki/Q7520643","display_name":"Simple (philosophy)","level":2,"score":0.31859999895095825},{"id":"https://openalex.org/C48057537","wikidata":"https://www.wikidata.org/wiki/Q3551145","display_name":"Statistical theory","level":2,"score":0.3034999966621399},{"id":"https://openalex.org/C198531522","wikidata":"https://www.wikidata.org/wiki/Q485146","display_name":"Sample (material)","level":2,"score":0.30239999294281006},{"id":"https://openalex.org/C87040749","wikidata":"https://www.wikidata.org/wiki/Q428971","display_name":"Power law","level":2,"score":0.30079999566078186},{"id":"https://openalex.org/C62799726","wikidata":"https://www.wikidata.org/wiki/Q190056","display_name":"Hilbert space","level":2,"score":0.2962000072002411},{"id":"https://openalex.org/C117765406","wikidata":"https://www.wikidata.org/wiki/Q5362437","display_name":"Generalization error","level":3,"score":0.296099990606308},{"id":"https://openalex.org/C116834253","wikidata":"https://www.wikidata.org/wiki/Q2039217","display_name":"Identification (biology)","level":2,"score":0.290800005197525},{"id":"https://openalex.org/C87007009","wikidata":"https://www.wikidata.org/wiki/Q210832","display_name":"Statistical hypothesis testing","level":2,"score":0.2897999882698059},{"id":"https://openalex.org/C121864883","wikidata":"https://www.wikidata.org/wiki/Q677916","display_name":"Statistical physics","level":1,"score":0.2870999872684479},{"id":"https://openalex.org/C2776061190","wikidata":"https://www.wikidata.org/wiki/Q7451805","display_name":"Separation (statistics)","level":2,"score":0.2630000114440918},{"id":"https://openalex.org/C149782125","wikidata":"https://www.wikidata.org/wiki/Q160039","display_name":"Econometrics","level":1,"score":0.2551000118255615},{"id":"https://openalex.org/C2780990831","wikidata":"https://www.wikidata.org/wiki/Q319141","display_name":"Conjecture","level":2,"score":0.2506999969482422}],"mesh":[],"locations_count":2,"locations":[{"id":"pmh:oai:arXiv.org:2609.22337","doi":null,"is_oa":true,"landing_page_url":"https://arxiv.org/abs/2609.22337","pdf_url":"https://arxiv.org/pdf/2609.22337","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.2609.22337","doi":"https://doi.org/10.48550/arxiv.2609.22337","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2609.22337","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:2609.22337","doi":null,"is_oa":true,"landing_page_url":"https://arxiv.org/abs/2609.22337","pdf_url":"https://arxiv.org/pdf/2609.22337","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/W7214050872.pdf","grobid_xml":"https://content.openalex.org/works/W7214050872.grobid-xml"},"referenced_works_count":4,"referenced_works":["https://openalex.org/W4387355345","https://openalex.org/W4387561538","https://openalex.org/W4387929831","https://openalex.org/W4405655184"],"related_works":["https://openalex.org/W7154985695","https://openalex.org/W7154591723","https://openalex.org/W7165548713","https://openalex.org/W7083709762","https://openalex.org/W7215032912","https://openalex.org/W4403175214","https://openalex.org/W6885015010","https://openalex.org/W7162818373","https://openalex.org/W7137583598","https://openalex.org/W7154620814"],"abstract_inverted_index":{"Linear":[0],"probing":[1],"is":[2,153],"the":[3,11,97,114,172,213],"standard":[4],"instrument":[5],"for":[6,113,217],"detecting":[7],"social":[8],"biases":[9],"in":[10,28,120],"hidden":[12],"representations":[13,73],"of":[14,42,85],"large":[15],"language":[16],"models.":[17],"Yet":[18],"reported":[19],"probe":[20,53,147],"accuracies":[21],"come":[22],"almost":[23],"exclusively":[24],"from":[25,149,179],"\\emph{counterfactual}":[26],"evaluations":[27],"which":[29,142],"every":[30],"input":[31],"carries":[32,44],"an":[33,57,61,109],"explicit":[34],"demographic":[35,45],"marker.":[36],"Once":[37],"only":[38],"a":[39,51,66,83,91,102,123,130,138,180,201,205,218],"fraction":[40],"$\u03b1$":[41],"inputs":[43],"information,":[46],"performance":[47],"degrades":[48],"sharply,":[49],"and":[50,136,162,208],"weak":[52],"may":[54],"reflect":[55],"either":[56],"unbiased":[58],"model":[59],"or":[60],"underpowered":[62],"detector.":[63],"We":[64],"develop":[65],"theory":[67],"that":[68],"resolves":[69],"this":[70],"ambiguity.":[71],"Modeling":[72],"as":[74],"two":[75],"class-conditional":[76],"clusters":[77],"with":[78,101,158,188],"Mahalanobis":[79],"separation":[80,128],"$s$":[81],"on":[82,129,155,163],"manifold":[84],"curvature":[86,124],"$\\kap$,":[87],"we":[88],"prove:":[89],"(i)":[90],"finite-sample":[92],"generalization":[93],"bound":[94],"governed":[95],"by":[96],"manifold's":[98],"extrinsic":[99],"radius":[100],"matching":[103],"$\\smash{\\sqrt{\\dB/n}}$":[104],"minimax":[105],"lower":[106],"bound;":[107],"(ii)":[108],"exact":[110],"purity":[111,173,207],"law":[112,174],"maximum":[115],"linear-probe":[116],"AUC,":[117],"strictly":[118],"increasing":[119],"$\u03b1$;":[121],"(iii)":[122],"ceiling:":[125],"ambient":[126],"chordal":[127],"space":[131],"form":[132],"cannot":[133],"exceed":[134],"$2/\\sqrt{\\kap}$;":[135],"(iv)":[137],"detectability":[139],"threshold":[140],"below":[141],"no":[143,189],"audit":[144],"can":[145],"distinguish":[146],"output":[148],"chance.":[150],"Every":[151],"theorem":[152],"validated":[154],"synthetic":[156],"manifolds":[157],"known":[159],"ground":[160],"truth":[161],"six":[164],"open-weight":[165],"models":[166],"$\\times$":[167],"four":[168],"bias":[169,198],"dimensions,":[170],"where":[171],"predicts":[175],"entire":[176],"AUC--$\u03b1$":[177],"curves":[178],"single":[181],"cross-fitted":[182],"$\\hat":[183],"s$":[184],"measured":[185],"at":[186],"$\u03b1=1$,":[187],"parameters":[190],"fitted":[191],"to":[192],"those":[193],"curves.":[194],"The":[195],"framework":[196],"turns":[197],"auditing":[199],"into":[200],"power":[202],"analysis:":[203],"given":[204],"target":[206],"effect":[209],"size,":[210],"it":[211],"prescribes":[212],"sample":[214],"budget":[215],"$n(\u03b1)$":[216],"conclusive":[219],"audit.":[220]},"counts_by_year":[],"updated_date":"2026-11-07T11:25:07.637180","created_date":"2026-09-23T00:00:00"}