{"id":"https://openalex.org/W7220735973","doi":"https://doi.org/10.48550/arxiv.2610.04907","title":"Why Subliminal Learning Needs So Much Data: A Noisy Inverse View through Steering Vector Recovery","display_name":"Why Subliminal Learning Needs So Much Data: A Noisy Inverse View through Steering Vector Recovery","publication_year":2026,"publication_date":"2026-10-04","ids":{"openalex":"https://openalex.org/W7220735973","doi":"https://doi.org/10.48550/arxiv.2610.04907"},"language":"en","primary_location":{"id":"pmh:oai:arXiv.org:2610.04907","doi":null,"is_oa":true,"landing_page_url":"https://arxiv.org/abs/2610.04907","pdf_url":"https://arxiv.org/pdf/2610.04907","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/2610.04907","any_repository_has_fulltext":true},"authorships":[{"author_position":"first","author":{"id":"https://openalex.org/A5157768581","display_name":"Luoyu Chen","orcid":null,"observed_orcids":[]},"institutions":[{"id":"https://openalex.org/I114017466","display_name":"University of Technology Sydney","ror":"https://ror.org/03f0f6041","country_code":"AU","type":"education","lineage":["https://openalex.org/I114017466"]}],"countries":["AU"],"is_corresponding":false,"raw_author_name":"Chen, Luoyu","raw_affiliation_strings":["University of Technology Sydney , Sydney , NSW , Australia"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"University of Technology Sydney , Sydney , NSW , Australia","institution_ids":["https://openalex.org/I114017466"]}]},{"author_position":"middle","author":{"id":"https://openalex.org/A5157729488","display_name":"Xiaoyu Ding","orcid":null,"observed_orcids":[]},"institutions":[{"id":"https://openalex.org/I114017466","display_name":"University of Technology Sydney","ror":"https://ror.org/03f0f6041","country_code":"AU","type":"education","lineage":["https://openalex.org/I114017466"]}],"countries":["AU"],"is_corresponding":false,"raw_author_name":"Ding, Xiaoyu","raw_affiliation_strings":["University of Technology Sydney , Sydney , NSW , Australia"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"University of Technology Sydney , Sydney , NSW , Australia","institution_ids":["https://openalex.org/I114017466"]}]},{"author_position":"middle","author":{"id":"https://openalex.org/A5157676619","display_name":"Weiqi Wang","orcid":null,"observed_orcids":[]},"institutions":[{"id":"https://openalex.org/I87445476","display_name":"Xi'an Jiaotong University","ror":"https://ror.org/017zhmm22","country_code":"CN","type":"education","lineage":["https://openalex.org/I87445476"]}],"countries":["CN"],"is_corresponding":false,"raw_author_name":"Wang, Weiqi","raw_affiliation_strings":["Xi'an Jiaotong University , Xi'an , Shaanxi , China"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"Xi'an Jiaotong University , Xi'an , Shaanxi , China","institution_ids":["https://openalex.org/I87445476"]}]},{"author_position":"middle","author":{"id":"https://openalex.org/A5157758665","display_name":"Chenhan Zhang","orcid":null,"observed_orcids":[]},"institutions":[{"id":"https://openalex.org/I114017466","display_name":"University of Technology Sydney","ror":"https://ror.org/03f0f6041","country_code":"AU","type":"education","lineage":["https://openalex.org/I114017466"]}],"countries":["AU"],"is_corresponding":false,"raw_author_name":"Zhang, Chenhan","raw_affiliation_strings":["University of Technology Sydney , Sydney , NSW , Australia"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"University of Technology Sydney , Sydney , NSW , Australia","institution_ids":["https://openalex.org/I114017466"]}]},{"author_position":"middle","author":{"id":"https://openalex.org/A5059567156","display_name":"Zhiyi Tian","orcid":"https://orcid.org/0000-0001-8905-0941","observed_orcids":["https://orcid.org/0000-0001-8905-0941"]},"institutions":[{"id":"https://openalex.org/I76569877","display_name":"Southeast University","ror":"https://ror.org/04ct4d772","country_code":"CN","type":"education","lineage":["https://openalex.org/I76569877"]}],"countries":["CN"],"is_corresponding":false,"raw_author_name":"Tian, Zhiyi","raw_affiliation_strings":["Southeast University , Nanjing , Jiangsu , China"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"Southeast University , Nanjing , Jiangsu , China","institution_ids":["https://openalex.org/I76569877"]}]},{"author_position":"middle","author":{"id":"https://openalex.org/A5137054424","display_name":"Jianhuan Huang","orcid":null,"observed_orcids":[]},"institutions":[{"id":"https://openalex.org/I114017466","display_name":"University of Technology Sydney","ror":"https://ror.org/03f0f6041","country_code":"AU","type":"education","lineage":["https://openalex.org/I114017466"]}],"countries":["AU"],"is_corresponding":false,"raw_author_name":"Huang, Jianhuan","raw_affiliation_strings":["University of Technology Sydney , Sydney , NSW , Australia"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"University of Technology Sydney , Sydney , NSW , Australia","institution_ids":["https://openalex.org/I114017466"]}]},{"author_position":"last","author":{"id":"https://openalex.org/A5157664721","display_name":"Shui Yu","orcid":null,"observed_orcids":[]},"institutions":[{"id":"https://openalex.org/I114017466","display_name":"University of Technology Sydney","ror":"https://ror.org/03f0f6041","country_code":"AU","type":"education","lineage":["https://openalex.org/I114017466"]}],"countries":["AU"],"is_corresponding":false,"raw_author_name":"Yu, Shui","raw_affiliation_strings":["University of Technology Sydney , Sydney , NSW , Australia"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"University of Technology Sydney , Sydney , NSW , Australia","institution_ids":["https://openalex.org/I114017466"]}]}],"institutions":[],"countries_distinct_count":2,"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/T11307","display_name":"Domain Adaptation and Few-Shot Learning","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/T11307","display_name":"Domain Adaptation and Few-Shot Learning","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/T10028","display_name":"Topic Modeling","score":0.01590000092983246,"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":0.0038999998942017555,"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/subliminal-learning","display_name":"subliminal learning","score":0.9670000076293945},{"id":"https://openalex.org/keywords/fisher-information-matrix","display_name":"Fisher information matrix","score":0.5320000052452087},{"id":"https://openalex.org/keywords/soft-supervision","display_name":"soft supervision","score":0.40299999713897705},{"id":"https://openalex.org/keywords/noisy-labels","display_name":"label noise","score":0.5820000171661377},{"id":"https://openalex.org/keywords/parameter-recovery","display_name":"parameter recovery","score":0.46799999475479126},{"id":"https://openalex.org/keywords/data-efficiency","display_name":"data efficiency","score":0.44999998807907104},{"id":"https://openalex.org/keywords/large-language-models","display_name":"large language models","score":0.47699999809265137}],"concepts":[{"id":"https://openalex.org/C41008148","wikidata":"https://www.wikidata.org/wiki/Q21198","display_name":"Computer science","level":0,"score":0.579200029373169},{"id":"https://openalex.org/C11413529","wikidata":"https://www.wikidata.org/wiki/Q8366","display_name":"Algorithm","level":1,"score":0.5078999996185303},{"id":"https://openalex.org/C151496658","wikidata":"https://www.wikidata.org/wiki/Q310661","display_name":"Subliminal stimuli","level":2,"score":0.5004000067710876},{"id":"https://openalex.org/C126780896","wikidata":"https://www.wikidata.org/wiki/Q899871","display_name":"Distortion (music)","level":4,"score":0.44190001487731934},{"id":"https://openalex.org/C79337645","wikidata":"https://www.wikidata.org/wiki/Q779824","display_name":"Outlier","level":2,"score":0.4189000129699707},{"id":"https://openalex.org/C153258448","wikidata":"https://www.wikidata.org/wiki/Q1199743","display_name":"Gradient descent","level":3,"score":0.4146000146865845},{"id":"https://openalex.org/C150921843","wikidata":"https://www.wikidata.org/wiki/Q1170431","display_name":"Resampling","level":2,"score":0.3971000015735626},{"id":"https://openalex.org/C207467116","wikidata":"https://www.wikidata.org/wiki/Q4385666","display_name":"Inverse","level":2,"score":0.39489999413490295},{"id":"https://openalex.org/C1893757","wikidata":"https://www.wikidata.org/wiki/Q3653001","display_name":"Inversion (geology)","level":3,"score":0.3865000009536743},{"id":"https://openalex.org/C99498987","wikidata":"https://www.wikidata.org/wiki/Q2210247","display_name":"Noise (video)","level":3,"score":0.36230000853538513},{"id":"https://openalex.org/C154945302","wikidata":"https://www.wikidata.org/wiki/Q11660","display_name":"Artificial intelligence","level":1,"score":0.3467000126838684},{"id":"https://openalex.org/C135252773","wikidata":"https://www.wikidata.org/wiki/Q1567213","display_name":"Inverse problem","level":2,"score":0.3400000035762787},{"id":"https://openalex.org/C33923547","wikidata":"https://www.wikidata.org/wiki/Q395","display_name":"Mathematics","level":0,"score":0.32580000162124634},{"id":"https://openalex.org/C140528856","wikidata":"https://www.wikidata.org/wiki/Q191884","display_name":"Inverse function","level":3,"score":0.31949999928474426},{"id":"https://openalex.org/C28719098","wikidata":"https://www.wikidata.org/wiki/Q44946","display_name":"Point (geometry)","level":2,"score":0.3172999918460846},{"id":"https://openalex.org/C64543145","wikidata":"https://www.wikidata.org/wiki/Q162942","display_name":"Intersection (aeronautics)","level":2,"score":0.31690001487731934},{"id":"https://openalex.org/C107673813","wikidata":"https://www.wikidata.org/wiki/Q812534","display_name":"Bayesian probability","level":2,"score":0.3127000033855438},{"id":"https://openalex.org/C49937458","wikidata":"https://www.wikidata.org/wiki/Q2599292","display_name":"Probabilistic logic","level":2,"score":0.28949999809265137},{"id":"https://openalex.org/C80444323","wikidata":"https://www.wikidata.org/wiki/Q2878974","display_name":"Theoretical computer science","level":1,"score":0.27709999680519104},{"id":"https://openalex.org/C191795146","wikidata":"https://www.wikidata.org/wiki/Q3878446","display_name":"Norm (philosophy)","level":2,"score":0.2736000120639801},{"id":"https://openalex.org/C140779682","wikidata":"https://www.wikidata.org/wiki/Q210868","display_name":"Sampling (signal processing)","level":3,"score":0.27219998836517334},{"id":"https://openalex.org/C114466953","wikidata":"https://www.wikidata.org/wiki/Q6034165","display_name":"Initialization","level":2,"score":0.2720000147819519},{"id":"https://openalex.org/C29406490","wikidata":"https://www.wikidata.org/wiki/Q1420659","display_name":"Fisher information","level":2,"score":0.2660999894142151},{"id":"https://openalex.org/C2777303404","wikidata":"https://www.wikidata.org/wiki/Q759757","display_name":"Convergence (economics)","level":2,"score":0.2614000141620636},{"id":"https://openalex.org/C36299963","wikidata":"https://www.wikidata.org/wiki/Q1369844","display_name":"Observability","level":2,"score":0.2565999925136566},{"id":"https://openalex.org/C2780909371","wikidata":"https://www.wikidata.org/wiki/Q4801092","display_name":"Artificial noise","level":4,"score":0.25600001215934753},{"id":"https://openalex.org/C45374587","wikidata":"https://www.wikidata.org/wiki/Q12525525","display_name":"Computation","level":2,"score":0.25220000743865967},{"id":"https://openalex.org/C32022120","wikidata":"https://www.wikidata.org/wiki/Q797225","display_name":"Interference (communication)","level":3,"score":0.25189998745918274}],"mesh":[],"locations_count":2,"locations":[{"id":"pmh:oai:arXiv.org:2610.04907","doi":null,"is_oa":true,"landing_page_url":"https://arxiv.org/abs/2610.04907","pdf_url":"https://arxiv.org/pdf/2610.04907","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.2610.04907","doi":"https://doi.org/10.48550/arxiv.2610.04907","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2610.04907","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-sa","license_id":"https://openalex.org/licenses/cc-by-sa","version":null,"is_accepted":false,"is_published":false,"raw_source_name":null,"raw_type":"Preprint"}],"best_oa_location":{"id":"pmh:oai:arXiv.org:2610.04907","doi":null,"is_oa":true,"landing_page_url":"https://arxiv.org/abs/2610.04907","pdf_url":"https://arxiv.org/pdf/2610.04907","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":[{"score":0.4156000018119812,"id":"https://metadata.un.org/sdg/4","display_name":"Quality education"}],"study_designs":[],"awards":[],"funders":[],"has_content":{"grobid_xml":true,"pdf":true},"content_urls":{"pdf":"https://content.openalex.org/works/W7220735973.pdf","grobid_xml":"https://content.openalex.org/works/W7220735973.grobid-xml"},"referenced_works_count":14,"referenced_works":["https://openalex.org/W1570089119","https://openalex.org/W1594234351","https://openalex.org/W1821462560","https://openalex.org/W1970789124","https://openalex.org/W2005836623","https://openalex.org/W2034978228","https://openalex.org/W2091178361","https://openalex.org/W2123845705","https://openalex.org/W2165715950","https://openalex.org/W2763894180","https://openalex.org/W7083819187","https://openalex.org/W7154461809","https://openalex.org/W7158313565","https://openalex.org/W7204533374"],"related_works":["https://openalex.org/W7163153322","https://openalex.org/W7158313565","https://openalex.org/W7201864956","https://openalex.org/W7221510522","https://openalex.org/W7206186785","https://openalex.org/W7157815529","https://openalex.org/W7083819187","https://openalex.org/W7213413124","https://openalex.org/W7221159880","https://openalex.org/W4417434172"],"abstract_inverted_index":{"Subliminal":[0],"learning":[1],"lets":[2],"a":[3,6,42,99,124,154,233],"student":[4],"inherit":[5],"teacher's":[7],"behavioral":[8],"trait":[9,40,135,256],"from":[10,98,213,232],"semantically":[11],"unrelated":[12],"data,":[13],"yet":[14],"published":[15],"demonstrations":[16],"typically":[17],"require":[18],"tens":[19,112],"of":[20,22,113,158,193,207],"thousands":[21],"carrier":[23,58,131,248],"examples.":[24],"We":[25,115],"ask":[26],"where":[27],"this":[28,117,163,245],"data":[29,224],"requirement":[30],"comes":[31],"from.":[32],"Our":[33],"testbed":[34],"is":[35,41,183,267],"subliminal":[36],"steering:":[37],"the":[38,73,130,134,174,191,204,208,227,255],"teacher":[39],"known":[43],"residual-stream":[44],"vector":[45],"$\u0394_T$,":[46],"so":[47,141],"transfer":[48],"can":[49],"be":[50],"measured":[51],"directly":[52],"as":[53,123,153,238,240],"parameter":[54],"recovery.":[55],"On":[56],"identical":[57],"prefixes,":[59],"we":[60],"compare":[61],"token-level":[62],"(hard)":[63],"NLL":[64],"supervision":[65,93,105],"with":[66,84,111,190,223],"full-distribution":[67],"(soft)":[68],"KL":[69],"supervision.":[70],"At":[71],"initialization":[72],"two":[74],"objectives":[75],"give":[76],"nearly":[77],"collinear":[78],"gradients,":[79],"and":[80,199,226],"both":[81],"align":[82],"poorly":[83],"$\u0394_T$.":[85,148],"Under":[86],"iterative":[87],"optimization,":[88],"however,":[89],"they":[90],"diverge:":[91],"soft":[92,161],"recovers":[94],"$\u0394_T$":[95],"almost":[96],"exactly":[97],"few":[100],"hundred":[101],"carriers,":[102],"while":[103],"hard":[104,169],"stays":[106],"well":[107,239],"below":[108],"it":[109,171],"even":[110],"thousands.":[114],"explain":[116],"gap":[118],"by":[119,263],"casting":[120],"steering-vector":[121],"distillation":[122],"noisy":[125],"linear":[126],"inverse":[127,157,164],"problem.":[128],"Locally,":[129],"task":[132],"maps":[133],"through":[136],"its":[137],"Fisher":[138,205,264],"matrix":[139],"$F$,":[140],"gradients":[142],"point":[143],"toward":[144],"$F\u0394_T$":[145],"rather":[146],"than":[147,257],"Gradient":[149],"descent":[150],"then":[151],"acts":[152],"progressively":[155],"less-damped":[156],"$F$.":[159],"With":[160,168],"targets,":[162],"restores":[165],"low-curvature":[166],"directions.":[167,180],"labels,":[170],"also":[172],"amplifies":[173],"sampling":[175],"noise":[176,261],"in":[177],"those":[178],"same":[179],"The":[181],"result":[182],"an":[184,218],"optimal":[185,219],"inversion":[186],"depth":[187,220],"that":[188,221,229],"grows":[189],"number":[192],"independent":[194],"carriers.":[195],"Experiments":[196],"on":[197],"Qwen2.5-7B":[198],"Gemma-2-9B":[200],"confirm":[201],"four":[202],"predictions:":[203],"distortion":[206],"initial":[209],"gradient,":[210],"recovery":[211],"ordered":[212],"steep":[214],"to":[215,253,258],"flat":[216],"directions,":[217],"shifts":[222],"scale,":[225],"finding":[228],"resampling":[230],"completions":[231],"fixed":[234],"prompt":[235],"pool":[236],"works":[237],"adding":[241],"new":[242],"prompts.":[243],"In":[244],"setting,":[246],"large":[247],"datasets":[249],"are":[250],"needed":[251],"less":[252],"reveal":[254],"suppress":[259],"label":[260],"amplified":[262],"inversion.":[265],"Code":[266],"available":[268],"at":[269],"\\url{https://github.com/luoyuchenmlcv/subliminal-data}.":[270]},"counts_by_year":[],"updated_date":"2026-11-07T11:25:07.637180","created_date":"2026-10-07T00:00:00"}