{"id":"https://openalex.org/W7223829162","doi":"https://doi.org/10.5281/zenodo.23679670","title":"The Attention Dispersion Law and Dual-Path Decoupling in Autoregressive Transformers","display_name":"The Attention Dispersion Law and Dual-Path Decoupling in Autoregressive Transformers","publication_year":2026,"publication_date":"2026-10-27","ids":{"openalex":"https://openalex.org/W7223829162","doi":"https://doi.org/10.5281/zenodo.23679670"},"language":"en","primary_location":{"id":"doi:10.5281/zenodo.23679670","doi":"https://doi.org/10.5281/zenodo.23679670","is_oa":true,"landing_page_url":"https://doi.org/10.5281/zenodo.23679670","pdf_url":null,"source":{"id":"https://openalex.org/S4306400562","display_name":"Zenodo (CERN European Organization for Nuclear Research)","issn_l":null,"issn":null,"is_oa":true,"is_in_doaj":false,"is_core":false,"listed_in":[],"host_organization":"https://openalex.org/I67311998","host_organization_name":"European Organization for Nuclear Research","host_organization_lineage":["https://openalex.org/I67311998"],"host_organization_lineage_names":[],"type":"repository"},"license":"cc-by","license_id":"https://openalex.org/licenses/cc-by","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/10.5281/zenodo.23679670","any_repository_has_fulltext":true},"authorships":[{"author_position":"first","author":{"id":"https://openalex.org/A5159841681","display_name":"Vladimir V. Borovikov","orcid":null,"observed_orcids":[]},"institutions":[],"countries":[],"is_corresponding":true,"raw_author_name":"Borovikov, Vladimir V.","raw_affiliation_strings":["Independent Researcher"],"raw_orcid":"https://orcid.org/0009-0005-7773-3581","affiliations":[{"raw_affiliation_string":"Independent Researcher","institution_ids":[]}]}],"institutions":[],"countries_distinct_count":0,"institutions_distinct_count":0,"corresponding_author_ids":["https://openalex.org/A5159841681"],"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":0.6244999766349792,"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":0.6244999766349792,"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":0.24799999594688416,"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":0.054499998688697815,"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/attention-dispersion","display_name":"attention dispersion","score":0.5019999742507935},{"id":"https://openalex.org/keywords/autoregressive-transformer","display_name":"autoregressive transformers","score":0.7689999938011169},{"id":"https://openalex.org/keywords/large-language-models","display_name":"large language models","score":0.7900000214576721},{"id":"https://openalex.org/keywords/bpe-tokenization","display_name":"BPE tokenization","score":0.675000011920929},{"id":"https://openalex.org/keywords/mixture-of-experts","display_name":"mixture of experts","score":0.5139999985694885}],"concepts":[{"id":"https://openalex.org/C41008148","wikidata":"https://www.wikidata.org/wiki/Q21198","display_name":"Computer science","level":0,"score":0.5760999917984009},{"id":"https://openalex.org/C11413529","wikidata":"https://www.wikidata.org/wiki/Q8366","display_name":"Algorithm","level":1,"score":0.46720001101493835},{"id":"https://openalex.org/C159877910","wikidata":"https://www.wikidata.org/wiki/Q2202883","display_name":"Autoregressive model","level":2,"score":0.4578000009059906},{"id":"https://openalex.org/C137293760","wikidata":"https://www.wikidata.org/wiki/Q3621696","display_name":"Language model","level":2,"score":0.4375},{"id":"https://openalex.org/C2780531524","wikidata":"https://www.wikidata.org/wiki/Q551540","display_name":"Ghosting","level":2,"score":0.430400013923645},{"id":"https://openalex.org/C61224824","wikidata":"https://www.wikidata.org/wiki/Q2260434","display_name":"Mixture model","level":2,"score":0.39320001006126404},{"id":"https://openalex.org/C66322947","wikidata":"https://www.wikidata.org/wiki/Q11658","display_name":"Transformer","level":3,"score":0.3919999897480011},{"id":"https://openalex.org/C2779343474","wikidata":"https://www.wikidata.org/wiki/Q3109175","display_name":"Context (archaeology)","level":2,"score":0.37209999561309814},{"id":"https://openalex.org/C80444323","wikidata":"https://www.wikidata.org/wiki/Q2878974","display_name":"Theoretical computer science","level":1,"score":0.3646000027656555},{"id":"https://openalex.org/C28490314","wikidata":"https://www.wikidata.org/wiki/Q189436","display_name":"Speech recognition","level":1,"score":0.35269999504089355},{"id":"https://openalex.org/C24424167","wikidata":"https://www.wikidata.org/wiki/Q180907","display_name":"Bijection","level":2,"score":0.34459999203681946},{"id":"https://openalex.org/C154945302","wikidata":"https://www.wikidata.org/wiki/Q11660","display_name":"Artificial intelligence","level":1,"score":0.3431999981403351},{"id":"https://openalex.org/C68757728","wikidata":"https://www.wikidata.org/wiki/Q911082","display_name":"DEVS","level":3,"score":0.33399999141693115},{"id":"https://openalex.org/C2777532361","wikidata":"https://www.wikidata.org/wiki/Q687185","display_name":"Lexicalization","level":2,"score":0.31380000710487366},{"id":"https://openalex.org/C2781402376","wikidata":"https://www.wikidata.org/wiki/Q17126172","display_name":"Photon counting","level":3,"score":0.30720001459121704},{"id":"https://openalex.org/C150799807","wikidata":"https://www.wikidata.org/wiki/Q374081","display_name":"Subtractive color","level":2,"score":0.30059999227523804},{"id":"https://openalex.org/C519991488","wikidata":"https://www.wikidata.org/wiki/Q28865","display_name":"Python (programming language)","level":2,"score":0.29670000076293945},{"id":"https://openalex.org/C9376300","wikidata":"https://www.wikidata.org/wiki/Q168817","display_name":"Algebraic number","level":2,"score":0.29420000314712524},{"id":"https://openalex.org/C161407221","wikidata":"https://www.wikidata.org/wiki/Q4382939","display_name":"Cognitive model","level":3,"score":0.28630000352859497},{"id":"https://openalex.org/C33923547","wikidata":"https://www.wikidata.org/wiki/Q395","display_name":"Mathematics","level":0,"score":0.2838999927043915},{"id":"https://openalex.org/C55689738","wikidata":"https://www.wikidata.org/wiki/Q15963867","display_name":"Discrete time and continuous time","level":2,"score":0.2822999954223633},{"id":"https://openalex.org/C183322885","wikidata":"https://www.wikidata.org/wiki/Q17007702","display_name":"Context model","level":3,"score":0.2809999883174896},{"id":"https://openalex.org/C45768438","wikidata":"https://www.wikidata.org/wiki/Q337432","display_name":"Numerical cognition","level":3,"score":0.28040000796318054},{"id":"https://openalex.org/C199539241","wikidata":"https://www.wikidata.org/wiki/Q7748","display_name":"Law","level":1,"score":0.2777999937534332},{"id":"https://openalex.org/C97137487","wikidata":"https://www.wikidata.org/wiki/Q729138","display_name":"Integer (computer science)","level":2,"score":0.27230000495910645},{"id":"https://openalex.org/C64869954","wikidata":"https://www.wikidata.org/wiki/Q1859747","display_name":"False positive paradox","level":2,"score":0.27059999108314514},{"id":"https://openalex.org/C94375191","wikidata":"https://www.wikidata.org/wiki/Q11205","display_name":"Arithmetic","level":1,"score":0.2703999876976013},{"id":"https://openalex.org/C141513077","wikidata":"https://www.wikidata.org/wiki/Q378542","display_name":"Independent and identically distributed random variables","level":3,"score":0.2696000039577484},{"id":"https://openalex.org/C127162648","wikidata":"https://www.wikidata.org/wiki/Q16858953","display_name":"Channel (broadcasting)","level":2,"score":0.26649999618530273},{"id":"https://openalex.org/C89128539","wikidata":"https://www.wikidata.org/wiki/Q1949963","display_name":"Statistic","level":2,"score":0.25600001215934753},{"id":"https://openalex.org/C87007009","wikidata":"https://www.wikidata.org/wiki/Q210832","display_name":"Statistical hypothesis testing","level":2,"score":0.25119999051094055}],"mesh":[],"locations_count":4,"locations":[{"id":"doi:10.5281/zenodo.23679670","doi":"https://doi.org/10.5281/zenodo.23679670","is_oa":true,"landing_page_url":"https://doi.org/10.5281/zenodo.23679670","pdf_url":null,"source":{"id":"https://openalex.org/S4306400562","display_name":"Zenodo (CERN European Organization for Nuclear Research)","issn_l":null,"issn":null,"is_oa":true,"is_in_doaj":false,"is_core":false,"listed_in":[],"host_organization":"https://openalex.org/I67311998","host_organization_name":"European Organization for Nuclear Research","host_organization_lineage":["https://openalex.org/I67311998"],"host_organization_lineage_names":[],"type":"repository"},"license":"cc-by","license_id":"https://openalex.org/licenses/cc-by","version":null,"is_accepted":false,"is_published":false,"raw_source_name":null,"raw_type":"Preprint"},{"id":"doi:10.5281/zenodo.23679671","doi":"https://doi.org/10.5281/zenodo.23679671","is_oa":true,"landing_page_url":"https://doi.org/10.5281/zenodo.23679671","pdf_url":null,"source":{"id":"https://openalex.org/S4306400562","display_name":"Zenodo (CERN European Organization for Nuclear Research)","issn_l":null,"issn":null,"is_oa":true,"is_in_doaj":false,"is_core":false,"listed_in":[],"host_organization":"https://openalex.org/I67311998","host_organization_name":"European Organization for Nuclear Research","host_organization_lineage":["https://openalex.org/I67311998"],"host_organization_lineage_names":[],"type":"repository"},"license":"cc-by","license_id":"https://openalex.org/licenses/cc-by","version":null,"is_accepted":false,"is_published":false,"raw_source_name":null,"raw_type":"Preprint"},{"id":"doi:10.5281/zenodo.23813779","doi":"https://doi.org/10.5281/zenodo.23813779","is_oa":true,"landing_page_url":"https://doi.org/10.5281/zenodo.23813779","pdf_url":null,"source":{"id":"https://openalex.org/S4306400562","display_name":"Zenodo (CERN European Organization for Nuclear Research)","issn_l":null,"issn":null,"is_oa":true,"is_in_doaj":false,"is_core":false,"listed_in":[],"host_organization":"https://openalex.org/I67311998","host_organization_name":"European Organization for Nuclear Research","host_organization_lineage":["https://openalex.org/I67311998"],"host_organization_lineage_names":[],"type":"repository"},"license":"cc-by","license_id":"https://openalex.org/licenses/cc-by","version":null,"is_accepted":false,"is_published":false,"raw_source_name":null,"raw_type":"Preprint"},{"id":"doi:10.5281/zenodo.23813780","doi":"https://doi.org/10.5281/zenodo.23813780","is_oa":true,"landing_page_url":"https://doi.org/10.5281/zenodo.23813780","pdf_url":null,"source":{"id":"https://openalex.org/S4306400562","display_name":"Zenodo (CERN European Organization for Nuclear Research)","issn_l":null,"issn":null,"is_oa":true,"is_in_doaj":false,"is_core":false,"listed_in":[],"host_organization":"https://openalex.org/I67311998","host_organization_name":"European Organization for Nuclear Research","host_organization_lineage":["https://openalex.org/I67311998"],"host_organization_lineage_names":[],"type":"repository"},"license":"cc-by","license_id":"https://openalex.org/licenses/cc-by","version":null,"is_accepted":false,"is_published":false,"raw_source_name":null,"raw_type":"Preprint"}],"best_oa_location":{"id":"doi:10.5281/zenodo.23679670","doi":"https://doi.org/10.5281/zenodo.23679670","is_oa":true,"landing_page_url":"https://doi.org/10.5281/zenodo.23679670","pdf_url":null,"source":{"id":"https://openalex.org/S4306400562","display_name":"Zenodo (CERN European Organization for Nuclear Research)","issn_l":null,"issn":null,"is_oa":true,"is_in_doaj":false,"is_core":false,"listed_in":[],"host_organization":"https://openalex.org/I67311998","host_organization_name":"European Organization for Nuclear Research","host_organization_lineage":["https://openalex.org/I67311998"],"host_organization_lineage_names":[],"type":"repository"},"license":"cc-by","license_id":"https://openalex.org/licenses/cc-by","version":null,"is_accepted":false,"is_published":false,"raw_source_name":null,"raw_type":"Preprint"},"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/W7160428581","https://openalex.org/W7125430285","https://openalex.org/W7133246625","https://openalex.org/W7148512853","https://openalex.org/W7160910432","https://openalex.org/W7164069945","https://openalex.org/W7222787774","https://openalex.org/W4415796170","https://openalex.org/W7131250391","https://openalex.org/W7140305852"],"abstract_inverted_index":{"Large":[0],"language":[1,187],"models":[2],"(LLMs)":[3],"excel":[4],"at":[5],"smooth":[6],"semantic":[7],"continuation,":[8],"yet":[9,251],"their":[10],"behavior":[11],"remains":[12],"vulnerable":[13],"when":[14],"maintaining":[15],"discrete":[16,81,221,311],"invariants\u2014exact":[17],"integer":[18],"counting":[19,61,82,222],"constraints":[20],"over":[21],"long":[22],"sequences.":[23],"Across":[24],"a":[25,91,164,173,244],"pre-registered":[26],"benchmark":[27],"of":[28,79,93,101,156,334],"1,416":[29],"canonical":[30],"probes":[31,62,236],"(1,320":[32],"in":[33,41,63,70,246,267,341],"the":[34,76,80,102,157,208,217,220,238,257,263,288,338],"core":[35],"Argon":[36],"v8-v16":[37],"series,":[38],"plus":[39],"96":[40,135],"replication":[42],"v6)":[43],"evaluated":[44],"under":[45],"zero-temperature":[46],"on":[47],"modern":[48],"Mixture-of-Experts":[49],"transformer":[50],"runtimes":[51],"(with":[52],"strict":[53],"No-Tools":[54],"isolation":[55],"across":[56],"all":[57],"960":[58],"purely":[59],"cognitive":[60],"v9":[64],"and":[65,72,162,213,326],"v11-v16,":[66],"alongside":[67],"tool-manipulation":[68],"protocols":[69],"v8":[71],"v10),":[73],"we":[74,97,130,183],"investigate":[75],"mechanistic":[77],"locus":[78],"error":[83],"(the":[84],"N":[85],"->":[86],"N+1":[87],"fencepost":[88],"overshoot).":[89],"Through":[90],"sequence":[92],"anatomical":[94],"interventional":[95],"experiments,":[96],"demonstrate":[98],"that:1.":[99],"Exoneration":[100],"Decoder:":[103],"By":[104,160],"applying":[105],"an":[106,225],"algebraic":[107],"amplifier":[108],"(f(N)":[109],"=":[110,121,128,202],"10N":[111],"-":[112],"6,":[113],"v15)":[114],"to":[115,191,206,242],"separate":[116],"attention":[117,226,297],"scan":[118,149],"drift":[119,150],"(Delta":[120,127],"+/-10)":[122],"from":[123,256,277],"output":[124],"unembedding":[125],"glitches":[126],"+1),":[129],"find":[131],"strictly":[132],"0":[133],"/":[134,344],"(0.0%,":[136],"two-sided":[137],"95%":[138],"Clopper-Pearson":[139],"CI":[140],"[0.0%,":[141],"3.8%])":[142],"decoder":[143],"off-by-one":[144],"errors,":[145],"while":[146],"input":[147],"perceptual":[148],"accounts":[151],"for":[152],"34.4%":[153],"(33/96).2.":[154],"Disqualification":[155],"Tokenizer":[158],"Hypothesis:":[159],"byte-":[161],"character-matching":[163],"repetitive":[165],"Korean":[166],"subword":[167],"marker":[168],"(12":[169,179],"bytes":[170,180],"UTF-8)":[171],"against":[172],"clean":[174],"Unicode":[175],"geometric":[176],"box-drawing":[177],"block":[178],"UTF-8,":[181],"v16),":[182],"show":[184],"that":[185,219,272],"replacing":[186],"with":[188,216,322],"geometry":[189],"fails":[190,205],"eliminate":[192],"phantom":[193],"errors":[194],"(35.4%":[195],"vs":[196],"41.7%,":[197],"Fisher":[198],"exact":[199],"test":[200],"p":[201],"0.675).":[203],"This":[204,270],"support":[207],"strong":[209],"BPE":[210],"tokenization":[211],"hypothesis":[212,218],"is":[214,337],"consistent":[215],"limit":[223],"reflects":[224],"phase":[227,305],"boundary":[228],"constraint.3.":[229],"Dual-Path":[230],"Attention":[231,289],"Decoupling:":[232],"In":[233],"adversarial":[234],"gaslighting":[235],"(v13),":[237],"model":[239],"follows":[240],"instructions":[241],"emit":[243],"falsehood":[245],"its":[247],"visible":[248],"content":[249],"(93.8%),":[250],"spontaneous":[252],"\"truth":[253],"leaks\"":[254],"directly":[255],"latent":[258],"reasoning":[259],"channel":[260],"(thinking)":[261],"into":[262,287],"terminal":[264,273],"token":[265],"occur":[266],"2/48":[268],"probes.":[269],"demonstrates":[271],"logits":[274],"can":[275],"decouple":[276],"local":[278,296],"autoregressive":[279],"n-gram":[280],"context.":[281],"We":[282],"formalize":[283],"these":[284],"empirical":[285],"findings":[286],"Dispersion":[290],"Law:":[291],"as":[292],"monotonic":[293],"sequences":[294],"exceed":[295],"span":[298],"capacity":[299],"(N":[300],">=":[301],"8..16),":[302],"positional":[303],"embedding":[304],"interference":[306],"smears":[307],"receptive":[308],"boundaries,":[309],"inducing":[310],"overshoot.":[312],"Consequently,":[313],"deterministic":[314,327],"agent":[315],"architectures":[316],"should":[317],"complement":[318],"in-context":[319],"data":[320],"passing":[321,325],"external":[323],"reference":[324],"runtime":[328],"parity":[329],"gates.":[330],"The":[331],"Russian":[332],"version":[333],"this":[335,342],"preprint":[336],"second":[339],"file":[340],"record:main_PACKAGE_v2_ru.pdf.":[343],"\u0420\u0443\u0441\u0441\u043a\u0430\u044f":[345],"\u0432\u0435\u0440\u0441\u0438\u044f":[346],"\u2014":[347],"\u0432\u0442\u043e\u0440\u043e\u0439":[348],"\u0444\u0430\u0439\u043b":[349],"\u0437\u0430\u043f\u0438\u0441\u0438:main_PACKAGE_v2_ru.pdf.":[350]},"counts_by_year":[],"updated_date":"2026-11-02T09:45:03.842371","created_date":"2026-10-28T00:00:00"}