{"id":"https://openalex.org/W7208761799","doi":"https://doi.org/10.48550/arxiv.2609.02959","title":"The Direction of Ignorance: LLMs Encode and Temper a Unigram Prior","display_name":"The Direction of Ignorance: LLMs Encode and Temper a Unigram Prior","publication_year":2026,"publication_date":"2026-09-02","ids":{"openalex":"https://openalex.org/W7208761799","doi":"https://doi.org/10.48550/arxiv.2609.02959"},"language":"en","primary_location":{"id":"pmh:oai:arXiv.org:2609.02959","doi":null,"is_oa":true,"landing_page_url":"https://arxiv.org/abs/2609.02959","pdf_url":"https://arxiv.org/pdf/2609.02959","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/2609.02959","any_repository_has_fulltext":true},"authorships":[{"author_position":"first","author":{"id":"https://openalex.org/A5150033302","display_name":"Toni J. B. Liu","orcid":null,"observed_orcids":[]},"institutions":[{"id":"https://openalex.org/I205783295","display_name":"Cornell University","ror":"https://ror.org/05bnh6r87","country_code":"US","type":"education","lineage":["https://openalex.org/I205783295"]}],"countries":["US"],"is_corresponding":false,"raw_author_name":"Liu, Toni J. B.","raw_affiliation_strings":["\u2663 Cornell University , USA ,"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"\u2663 Cornell University , USA ,","institution_ids":["https://openalex.org/I205783295"]}]},{"author_position":"middle","author":{"id":"https://openalex.org/A5133229058","display_name":"Jiajun Bao","orcid":"https://orcid.org/0009-0001-0344-8527","observed_orcids":["https://orcid.org/0009-0001-0344-8527"]},"institutions":[{"id":"https://openalex.org/I205783295","display_name":"Cornell University","ror":"https://ror.org/05bnh6r87","country_code":"US","type":"education","lineage":["https://openalex.org/I205783295"]}],"countries":["US"],"is_corresponding":false,"raw_author_name":"Bao, Jiajun","raw_affiliation_strings":["\u2663 Cornell University , USA ,"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"\u2663 Cornell University , USA ,","institution_ids":["https://openalex.org/I205783295"]}]},{"author_position":"middle","author":{"id":"https://openalex.org/A5160216783","display_name":"Gurbir Arora","orcid":null,"observed_orcids":[]},"institutions":[{"id":"https://openalex.org/I47508984","display_name":"Imperial College London","ror":"https://ror.org/041kmwe10","country_code":"GB","type":"education","lineage":["https://openalex.org/I47508984"]}],"countries":["GB"],"is_corresponding":false,"raw_author_name":"Arora, Gurbir","raw_affiliation_strings":["\u2661 Imperial College London , UK , \u2662 Goodfire AI , USA"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"\u2661 Imperial College London , UK , \u2662 Goodfire AI , USA","institution_ids":["https://openalex.org/I47508984"]}]},{"author_position":"middle","author":{"id":"https://openalex.org/A5092231224","display_name":"Mitchell Ostrow","orcid":"https://orcid.org/0000-0002-9732-548X","observed_orcids":["https://orcid.org/0000-0002-9732-548X"]},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Ostrow, Mitchell","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5160159617","display_name":"Yizhou Liu","orcid":null,"observed_orcids":[]},"institutions":[{"id":"https://openalex.org/I63966007","display_name":"Massachusetts Institute of Technology","ror":"https://ror.org/042nb2s44","country_code":"US","type":"education","lineage":["https://openalex.org/I63966007"]}],"countries":["US"],"is_corresponding":false,"raw_author_name":"Liu, Yizhou","raw_affiliation_strings":["\u2660 MIT , USA ,"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"\u2660 MIT , USA ,","institution_ids":["https://openalex.org/I63966007"]}]},{"author_position":"middle","author":{"id":"https://openalex.org/A5082174299","display_name":"Rapha\u00ebl Sarfati","orcid":"https://orcid.org/0000-0003-4944-0632","observed_orcids":["https://orcid.org/0000-0003-4944-0632"]},"institutions":[{"id":"https://openalex.org/I63966007","display_name":"Massachusetts Institute of Technology","ror":"https://ror.org/042nb2s44","country_code":"US","type":"education","lineage":["https://openalex.org/I63966007"]}],"countries":["US"],"is_corresponding":false,"raw_author_name":"Sarfati, Rapha\u00ebl","raw_affiliation_strings":["\u2660 MIT , USA ,"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"\u2660 MIT , USA ,","institution_ids":["https://openalex.org/I63966007"]}]},{"author_position":"middle","author":{"id":"https://openalex.org/A5012268400","display_name":"Nicolas Boull\u00e9","orcid":"https://orcid.org/0000-0002-1425-8307","observed_orcids":["https://orcid.org/0000-0002-1425-8307"]},"institutions":[{"id":"https://openalex.org/I205783295","display_name":"Cornell University","ror":"https://ror.org/05bnh6r87","country_code":"US","type":"education","lineage":["https://openalex.org/I205783295"]}],"countries":["US"],"is_corresponding":false,"raw_author_name":"Boull\u00e9, Nicolas","raw_affiliation_strings":["\u2663 Cornell University , USA ,"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"\u2663 Cornell University , USA ,","institution_ids":["https://openalex.org/I205783295"]}]},{"author_position":"last","author":{"id":"https://openalex.org/A5082345762","display_name":"Christopher Earls","orcid":"https://orcid.org/0000-0001-8944-5572","observed_orcids":["https://orcid.org/0000-0001-8944-5572"]},"institutions":[{"id":"https://openalex.org/I63966007","display_name":"Massachusetts Institute of Technology","ror":"https://ror.org/042nb2s44","country_code":"US","type":"education","lineage":["https://openalex.org/I63966007"]}],"countries":["US"],"is_corresponding":false,"raw_author_name":"Earls, Christopher J.","raw_affiliation_strings":["\u2660 MIT , USA ,"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"\u2660 MIT , USA ,","institution_ids":["https://openalex.org/I63966007"]}]}],"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/T10028","display_name":"Topic Modeling","score":0.9140999913215637,"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":0.9140999913215637,"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/T10181","display_name":"Natural Language Processing Techniques","score":0.055799998342990875,"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":0.014800000004470348,"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/prior-distributions","display_name":"prior distributions","score":0.6579999923706055},{"id":"https://openalex.org/keywords/model-size-scaling","display_name":"model size scaling","score":0.6320000290870667},{"id":"https://openalex.org/keywords/reasoning-models","display_name":"reasoning models","score":0.6710000038146973},{"id":"https://openalex.org/keywords/gsm8k","display_name":"GSM8K","score":0.8059999942779541}],"concepts":[{"id":"https://openalex.org/C177769412","wikidata":"https://www.wikidata.org/wiki/Q278090","display_name":"Prior probability","level":3,"score":0.800000011920929},{"id":"https://openalex.org/C107673813","wikidata":"https://www.wikidata.org/wiki/Q812534","display_name":"Bayesian probability","level":2,"score":0.6754000186920166},{"id":"https://openalex.org/C2779343474","wikidata":"https://www.wikidata.org/wiki/Q3109175","display_name":"Context (archaeology)","level":2,"score":0.59579998254776},{"id":"https://openalex.org/C527412718","wikidata":"https://www.wikidata.org/wiki/Q855395","display_name":"Interpretation (philosophy)","level":2,"score":0.571399986743927},{"id":"https://openalex.org/C33923547","wikidata":"https://www.wikidata.org/wiki/Q395","display_name":"Mathematics","level":0,"score":0.5257999897003174},{"id":"https://openalex.org/C2778732403","wikidata":"https://www.wikidata.org/wiki/Q815577","display_name":"Ignorance","level":2,"score":0.45969998836517334},{"id":"https://openalex.org/C154945302","wikidata":"https://www.wikidata.org/wiki/Q11660","display_name":"Artificial intelligence","level":1,"score":0.45419999957084656},{"id":"https://openalex.org/C61797465","wikidata":"https://www.wikidata.org/wiki/Q1188986","display_name":"Term (time)","level":2,"score":0.4424999952316284},{"id":"https://openalex.org/C160234255","wikidata":"https://www.wikidata.org/wiki/Q812535","display_name":"Bayesian inference","level":3,"score":0.4408000111579895},{"id":"https://openalex.org/C11413529","wikidata":"https://www.wikidata.org/wiki/Q8366","display_name":"Algorithm","level":1,"score":0.40119999647140503},{"id":"https://openalex.org/C110121322","wikidata":"https://www.wikidata.org/wiki/Q865811","display_name":"Distribution (mathematics)","level":2,"score":0.39489999413490295},{"id":"https://openalex.org/C41008148","wikidata":"https://www.wikidata.org/wiki/Q21198","display_name":"Computer science","level":0,"score":0.36809998750686646},{"id":"https://openalex.org/C57493831","wikidata":"https://www.wikidata.org/wiki/Q3134666","display_name":"Projection (relational algebra)","level":2,"score":0.36739999055862427},{"id":"https://openalex.org/C48103436","wikidata":"https://www.wikidata.org/wiki/Q599031","display_name":"State (computer science)","level":2,"score":0.3580000102519989},{"id":"https://openalex.org/C106487976","wikidata":"https://www.wikidata.org/wiki/Q685816","display_name":"Matrix (chemical analysis)","level":2,"score":0.35679998993873596},{"id":"https://openalex.org/C2776502983","wikidata":"https://www.wikidata.org/wiki/Q690182","display_name":"Contrast (vision)","level":2,"score":0.35030001401901245},{"id":"https://openalex.org/C101112237","wikidata":"https://www.wikidata.org/wiki/Q4874481","display_name":"Bayesian statistics","level":4,"score":0.3249000012874603},{"id":"https://openalex.org/C149441793","wikidata":"https://www.wikidata.org/wiki/Q200726","display_name":"Probability distribution","level":2,"score":0.3174000084400177},{"id":"https://openalex.org/C105795698","wikidata":"https://www.wikidata.org/wiki/Q12483","display_name":"Statistics","level":1,"score":0.3093999922275543},{"id":"https://openalex.org/C153180895","wikidata":"https://www.wikidata.org/wiki/Q7148389","display_name":"Pattern recognition (psychology)","level":2,"score":0.3068000078201294},{"id":"https://openalex.org/C207201462","wikidata":"https://www.wikidata.org/wiki/Q182505","display_name":"Bayes' theorem","level":3,"score":0.30640000104904175},{"id":"https://openalex.org/C149782125","wikidata":"https://www.wikidata.org/wiki/Q160039","display_name":"Econometrics","level":1,"score":0.30230000615119934},{"id":"https://openalex.org/C51632099","wikidata":"https://www.wikidata.org/wiki/Q3985153","display_name":"Training set","level":2,"score":0.29739999771118164},{"id":"https://openalex.org/C175694140","wikidata":"https://www.wikidata.org/wiki/Q980329","display_name":"Orthographic projection","level":2,"score":0.28690001368522644},{"id":"https://openalex.org/C2780388253","wikidata":"https://www.wikidata.org/wiki/Q5421508","display_name":"Exponent","level":2,"score":0.2759999930858612},{"id":"https://openalex.org/C137293760","wikidata":"https://www.wikidata.org/wiki/Q3621696","display_name":"Language model","level":2,"score":0.27219998836517334},{"id":"https://openalex.org/C114289077","wikidata":"https://www.wikidata.org/wiki/Q3284399","display_name":"Statistical model","level":2,"score":0.2651999890804291}],"mesh":[],"locations_count":2,"locations":[{"id":"pmh:oai:arXiv.org:2609.02959","doi":null,"is_oa":true,"landing_page_url":"https://arxiv.org/abs/2609.02959","pdf_url":"https://arxiv.org/pdf/2609.02959","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.2609.02959","doi":"https://doi.org/10.48550/arxiv.2609.02959","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2609.02959","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:2609.02959","doi":null,"is_oa":true,"landing_page_url":"https://arxiv.org/abs/2609.02959","pdf_url":"https://arxiv.org/pdf/2609.02959","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/W7208761799.pdf","grobid_xml":"https://content.openalex.org/works/W7208761799.grobid-xml"},"referenced_works_count":29,"referenced_works":["https://openalex.org/W1816313093","https://openalex.org/W2020946357","https://openalex.org/W2116064496","https://openalex.org/W2154099718","https://openalex.org/W2962940259","https://openalex.org/W2963206148","https://openalex.org/W2964191999","https://openalex.org/W3157005959","https://openalex.org/W3176618728","https://openalex.org/W3196813608","https://openalex.org/W3199011812","https://openalex.org/W3202180397","https://openalex.org/W4213445359","https://openalex.org/W4327526719","https://openalex.org/W4385570061","https://openalex.org/W4385572928","https://openalex.org/W4385573636","https://openalex.org/W4389519139","https://openalex.org/W4389520513","https://openalex.org/W4391159099","https://openalex.org/W4391670818","https://openalex.org/W4399521940","https://openalex.org/W4401042301","https://openalex.org/W4403122979","https://openalex.org/W4404132457","https://openalex.org/W4404782813","https://openalex.org/W4404792970","https://openalex.org/W4411119128","https://openalex.org/W4417089514"],"related_works":["https://openalex.org/W7117968941","https://openalex.org/W4412945633","https://openalex.org/W7000020581","https://openalex.org/W6922518453","https://openalex.org/W7218798243","https://openalex.org/W4399399219","https://openalex.org/W7213438527","https://openalex.org/W4417089514","https://openalex.org/W4394653302","https://openalex.org/W7117161873"],"abstract_inverted_index":{"What":[0],"does":[1],"a":[2,17,78,111,115,124,164],"language":[3],"model":[4,39,55,136],"predict":[5],"when":[6,43],"it":[7,169,181],"has":[8],"few":[9],"clues?The":[10],"answer":[11],"lies":[12],"in":[13],"its":[14],"unembedding":[15,22],"geometry:":[16],"single":[18],"direction":[19,49,76],"of":[20,28,50,110],"the":[21,25,29,35,38,48,70,88,94,98,107,120],"matrix":[23],"encodes":[24],"unigram":[26,116],"distribution":[27],"training":[30],"corpus,":[31],"which":[32,84],"serves":[33,162],"as":[34,87,106,163],"Bayesian":[36,113],"prior":[37,80,117,145],"falls":[40],"back":[41],"on":[42,177],"uncertain.":[44],"This":[45,127],"structure,":[46],"termed":[47],"ignorance,":[51],"appears":[52],"across":[53,135],"all":[54],"families":[56],"examined":[57],"(Llama,":[58],"Qwen,":[59],"Gemma,":[60],"and":[61,123,138,147,174],"Pythia),":[62],"ranging":[63],"from":[64],"0.4B":[65],"to":[66,119],"405B":[67],"parameters.":[68],"Projecting":[69],"final":[71],"prediction":[72,99],"state":[73,100],"onto":[74],"this":[75],"yields":[77],"per-token":[79],"loading":[81],"factor":[82],"$\u03bb$,":[83],"declines":[85],"steadily":[86],"context":[89],"becomes":[90],"more":[91],"informative.":[92],"Formally,":[93],"same":[95],"projection":[96],"decomposes":[97],"into":[101],"two":[102,108],"orthogonal":[103],"vectors":[104],"interpretable":[105],"components":[109],"tempered":[112],"update:":[114],"raised":[118],"exponent":[121],"$\u03bb$":[122,132,161],"context-driven":[125,172],"likelihood.":[126],"geometric":[128],"probabilistic":[129],"connection":[130],"makes":[131],"meaningfully":[133],"comparable":[134],"sizes":[137],"families.":[139],"Larger":[140],"models":[141,149],"generally":[142],"exhibit":[143,150],"lower":[144,151],"reliance,":[146],"reasoning":[148],"reliance":[152],"than":[153],"their":[154],"base":[155],"counterparts.":[156],"Finally,":[157],"we":[158],"show":[159],"that":[160],"causal,":[165],"inference-time":[166],"control:":[167],"reducing":[168],"can":[170],"promote":[171],"continuations":[173],"improve":[175],"Llama-3.2-3B-Instruct":[176],"GSM8K,":[178],"whereas":[179],"amplifying":[180],"or":[182],"applying":[183],"matched":[184],"perturbations":[185],"along":[186],"random":[187],"directions":[188],"degrades":[189],"performance.":[190]},"counts_by_year":[],"updated_date":"2026-11-08T12:12:28.789561","created_date":"2026-09-05T00:00:00"}