{"id":"https://openalex.org/W7128604689","doi":"https://doi.org/10.1109/focs63196.2025.00136","title":"Theoretical limitations of multi-layer Transformer","display_name":"Theoretical limitations of multi-layer Transformer","publication_year":2025,"publication_date":"2025-12-14","ids":{"openalex":"https://openalex.org/W7128604689","doi":"https://doi.org/10.1109/focs63196.2025.00136"},"language":null,"primary_location":{"id":"doi:10.1109/focs63196.2025.00136","is_oa":false,"landing_page_url":"https://doi.org/10.1109/focs63196.2025.00136","pdf_url":null,"source":null,"license":null,"license_id":null,"version":"publishedVersion","is_accepted":true,"is_published":true,"raw_source_name":"2025 IEEE 66th Annual Symposium on Foundations of Computer Science (FOCS)","raw_type":"proceedings-article"},"type":"article","indexed_in":["crossref"],"open_access":{"is_oa":false,"oa_status":"closed","oa_url":null,"any_repository_has_fulltext":false},"authorships":[{"author_position":"first","author":{"id":"https://openalex.org/A5123643155","display_name":"Lijie Chen","orcid":null},"institutions":[{"id":"https://openalex.org/I134446601","display_name":"Berkeley College","ror":"https://ror.org/02xewxa75","country_code":"US","type":"education","lineage":["https://openalex.org/I134446601"]},{"id":"https://openalex.org/I95457486","display_name":"University of California, Berkeley","ror":"https://ror.org/01an7q238","country_code":"US","type":"education","lineage":["https://openalex.org/I95457486"]}],"countries":["US"],"is_corresponding":true,"raw_author_name":"Lijie Chen","raw_affiliation_strings":["UC Berkeley,Berkeley,United States"],"affiliations":[{"raw_affiliation_string":"UC Berkeley,Berkeley,United States","institution_ids":["https://openalex.org/I134446601","https://openalex.org/I95457486"]}]},{"author_position":"middle","author":{"id":"https://openalex.org/A5125676982","display_name":"Binghui Peng","orcid":null},"institutions":[{"id":"https://openalex.org/I1291425158","display_name":"Google (United States)","ror":"https://ror.org/00njsd438","country_code":"US","type":"company","lineage":["https://openalex.org/I1291425158","https://openalex.org/I4210128969"]}],"countries":["US"],"is_corresponding":false,"raw_author_name":"Binghui Peng","raw_affiliation_strings":["Google research,New York,United States"],"affiliations":[{"raw_affiliation_string":"Google research,New York,United States","institution_ids":["https://openalex.org/I1291425158"]}]},{"author_position":"last","author":{"id":"https://openalex.org/A5101333353","display_name":"Hongxun Wu","orcid":null},"institutions":[{"id":"https://openalex.org/I134446601","display_name":"Berkeley College","ror":"https://ror.org/02xewxa75","country_code":"US","type":"education","lineage":["https://openalex.org/I134446601"]},{"id":"https://openalex.org/I95457486","display_name":"University of California, Berkeley","ror":"https://ror.org/01an7q238","country_code":"US","type":"education","lineage":["https://openalex.org/I95457486"]}],"countries":["US"],"is_corresponding":false,"raw_author_name":"Hongxun Wu","raw_affiliation_strings":["UC Berkeley,Berkeley,United States"],"affiliations":[{"raw_affiliation_string":"UC Berkeley,Berkeley,United States","institution_ids":["https://openalex.org/I134446601","https://openalex.org/I95457486"]}]}],"institutions":[],"countries_distinct_count":1,"institutions_distinct_count":3,"corresponding_author_ids":["https://openalex.org/A5123643155"],"corresponding_institution_ids":["https://openalex.org/I134446601","https://openalex.org/I95457486"],"apc_list":null,"apc_paid":null,"fwci":0.0,"has_fulltext":false,"cited_by_count":0,"citation_normalized_percentile":{"value":0.793558,"is_in_top_1_percent":false,"is_in_top_10_percent":false},"cited_by_percentile_year":null,"biblio":{"volume":null,"issue":null,"first_page":"2631","last_page":"2653"},"is_retracted":false,"is_paratext":false,"is_xpac":false,"primary_topic":{"id":"https://openalex.org/T10720","display_name":"Complexity and Algorithms in Graphs","score":0.3880000114440918,"subfield":{"id":"https://openalex.org/subfields/1703","display_name":"Computational Theory and Mathematics"},"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/T10720","display_name":"Complexity and Algorithms in Graphs","score":0.3880000114440918,"subfield":{"id":"https://openalex.org/subfields/1703","display_name":"Computational Theory and Mathematics"},"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/T12072","display_name":"Machine Learning and Algorithms","score":0.30660000443458557,"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/T11612","display_name":"Stochastic Gradient Optimization Techniques","score":0.053300000727176666,"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/upper-and-lower-bounds","display_name":"Upper and lower bounds","score":0.6383000016212463},{"id":"https://openalex.org/keywords/transformer","display_name":"Transformer","score":0.6133999824523926},{"id":"https://openalex.org/keywords/encoder","display_name":"Encoder","score":0.5800999999046326},{"id":"https://openalex.org/keywords/exponential-growth","display_name":"Exponential growth","score":0.5604000091552734},{"id":"https://openalex.org/keywords/sequence","display_name":"Sequence (biology)","score":0.46369999647140503},{"id":"https://openalex.org/keywords/task","display_name":"Task (project management)","score":0.424699991941452},{"id":"https://openalex.org/keywords/constant","display_name":"Constant (computer programming)","score":0.4187000095844269},{"id":"https://openalex.org/keywords/polynomial","display_name":"Polynomial","score":0.4056999981403351}],"concepts":[{"id":"https://openalex.org/C77553402","wikidata":"https://www.wikidata.org/wiki/Q13222579","display_name":"Upper and lower bounds","level":2,"score":0.6383000016212463},{"id":"https://openalex.org/C66322947","wikidata":"https://www.wikidata.org/wiki/Q11658","display_name":"Transformer","level":3,"score":0.6133999824523926},{"id":"https://openalex.org/C118505674","wikidata":"https://www.wikidata.org/wiki/Q42586063","display_name":"Encoder","level":2,"score":0.5800999999046326},{"id":"https://openalex.org/C41008148","wikidata":"https://www.wikidata.org/wiki/Q21198","display_name":"Computer science","level":0,"score":0.5640000104904175},{"id":"https://openalex.org/C75235859","wikidata":"https://www.wikidata.org/wiki/Q582659","display_name":"Exponential growth","level":2,"score":0.5604000091552734},{"id":"https://openalex.org/C2778112365","wikidata":"https://www.wikidata.org/wiki/Q3511065","display_name":"Sequence (biology)","level":2,"score":0.46369999647140503},{"id":"https://openalex.org/C11413529","wikidata":"https://www.wikidata.org/wiki/Q8366","display_name":"Algorithm","level":1,"score":0.45239999890327454},{"id":"https://openalex.org/C2780451532","wikidata":"https://www.wikidata.org/wiki/Q759676","display_name":"Task (project management)","level":2,"score":0.424699991941452},{"id":"https://openalex.org/C2777027219","wikidata":"https://www.wikidata.org/wiki/Q1284190","display_name":"Constant (computer programming)","level":2,"score":0.4187000095844269},{"id":"https://openalex.org/C90119067","wikidata":"https://www.wikidata.org/wiki/Q43260","display_name":"Polynomial","level":2,"score":0.4056999981403351},{"id":"https://openalex.org/C80444323","wikidata":"https://www.wikidata.org/wiki/Q2878974","display_name":"Theoretical computer science","level":1,"score":0.4050000011920929},{"id":"https://openalex.org/C33923547","wikidata":"https://www.wikidata.org/wiki/Q395","display_name":"Mathematics","level":0,"score":0.39890000224113464},{"id":"https://openalex.org/C159877910","wikidata":"https://www.wikidata.org/wiki/Q2202883","display_name":"Autoregressive model","level":2,"score":0.3686000108718872},{"id":"https://openalex.org/C26517878","wikidata":"https://www.wikidata.org/wiki/Q228039","display_name":"Key (lock)","level":2,"score":0.35249999165534973},{"id":"https://openalex.org/C33676613","wikidata":"https://www.wikidata.org/wiki/Q13415176","display_name":"Dimension (graph theory)","level":2,"score":0.35010001063346863},{"id":"https://openalex.org/C179145077","wikidata":"https://www.wikidata.org/wiki/Q5154130","display_name":"Communication complexity","level":2,"score":0.3492000102996826},{"id":"https://openalex.org/C40231798","wikidata":"https://www.wikidata.org/wiki/Q1333743","display_name":"Composition (language)","level":2,"score":0.3190999925136566},{"id":"https://openalex.org/C311688","wikidata":"https://www.wikidata.org/wiki/Q2393193","display_name":"Time complexity","level":2,"score":0.29499998688697815},{"id":"https://openalex.org/C179799912","wikidata":"https://www.wikidata.org/wiki/Q205084","display_name":"Computational complexity theory","level":2,"score":0.29429998993873596},{"id":"https://openalex.org/C137293760","wikidata":"https://www.wikidata.org/wiki/Q3621696","display_name":"Language model","level":2,"score":0.2809999883174896},{"id":"https://openalex.org/C57273362","wikidata":"https://www.wikidata.org/wiki/Q576722","display_name":"Decoding methods","level":2,"score":0.2689000070095062},{"id":"https://openalex.org/C158156997","wikidata":"https://www.wikidata.org/wiki/Q1416645","display_name":"Models of communication","level":2,"score":0.25870001316070557},{"id":"https://openalex.org/C118615104","wikidata":"https://www.wikidata.org/wiki/Q121416","display_name":"Discrete mathematics","level":1,"score":0.2563000023365021}],"mesh":[],"locations_count":1,"locations":[{"id":"doi:10.1109/focs63196.2025.00136","is_oa":false,"landing_page_url":"https://doi.org/10.1109/focs63196.2025.00136","pdf_url":null,"source":null,"license":null,"license_id":null,"version":"publishedVersion","is_accepted":true,"is_published":true,"raw_source_name":"2025 IEEE 66th Annual Symposium on Foundations of Computer Science (FOCS)","raw_type":"proceedings-article"}],"best_oa_location":null,"sustainable_development_goals":[],"awards":[],"funders":[],"has_content":{"grobid_xml":false,"pdf":false},"content_urls":null,"referenced_works_count":0,"referenced_works":[],"related_works":[],"abstract_inverted_index":{"Transformers,":[0],"especially":[1],"the":[2,6,31,37,62,108,117,170,183,187,192,199,265],"decoder-only":[3,69,80,204,218],"variants,":[4],"are":[5],"backbone":[7],"of":[8,21,39,92,98,160,179,202,221,246,268],"most":[9],"modern":[10],"large":[11],"language":[12],"models.":[13],"Yet,":[14],"we":[15,60,75,190,233],"have":[16],"a":[17,83,102,140,157,163,203,236,242],"very":[18],"limited":[19],"understanding":[20],"their":[22,222],"limitations":[23,53],"(i.e.,":[24],"what":[25],"tasks":[26],"they":[27],"cannot":[28],"solve)":[29],"besides":[30],"simplest":[32],"1-layer":[33],"case.":[34],"Due":[35],"to":[36,51,88,128,174,263],"difficulty":[38],"analyzing":[40],"multi-layer":[41,55,113],"models,":[42],"all":[43,247],"previous":[44],"work":[45],"relies":[46],"on":[47],"unproven":[48],"complexity":[49],"conjectures":[50],"show":[52],"for":[54,112,124,143],"Transformers.":[56],"In":[57,206],"this":[58,211,230],"work,":[59],"prove":[61,76,226],"first":[63,109],"unconditional":[64,133],"lower":[65,208,215,227],"bound":[66],"against":[67,217],"multilayer":[68],"transformers.":[70],"For":[71],"any":[72,78],"constant":[73],"L,":[74],"that":[77,116,145,165,197,240],"L-layer":[79,125],"transformer":[81],"needs":[82],"polynomial":[84],"model":[85,171,196,213,256],"dimension":[86],"$\\left(n^{\\Omega(1)}\\right)$":[87],"perform":[89],"sequential":[90],"composition":[91,119],"L":[93],"functions":[94],"over":[95],"an":[96,132,150,176],"input":[97],"n":[99],"tokens.":[100],"As":[101],"consequence,":[103],"our":[104,253],"results":[105],"give:":[106],"(1)":[107],"depthwidth":[110],"trade-off":[111],"transformers,":[114,219],"exhibiting":[115,139,162],"L-step":[118],"task":[120,142,164],"is":[121,172],"exponentially":[122,151,167],"harder":[123],"models":[126],"compared":[127],"$(L+1)$-layer":[129],"ones;":[130],"(2)":[131],"separation":[134],"between":[135],"encoder":[136],"and":[137,153,257],"decoder,":[138],"hard":[141],"decoders":[144],"can":[146],"be":[147,261],"solved":[148],"by":[149],"shallower":[152],"smaller":[154],"encoder;":[155],"(3)":[156],"provable":[158],"advantage":[159],"chain-of-thought,":[161],"becomes":[166],"easier":[168],"when":[169],"allowed":[173],"produce":[175],"intermediate":[177],"sequence":[178],"tokens":[180],"before":[181],"outputting":[182],"final":[184],"answer.":[185],"On":[186],"technical":[188],"side,":[189],"propose":[191],"multi-party":[193],"autoregressive":[194],"communication":[195,212,231,255],"abstracts":[198],"key":[200],"aspects":[201],"Transformer.":[205],"particular,":[207],"bounds":[209,216,228],"within":[210],"imply":[214],"independent":[220],"implementation":[223],"details.":[224],"To":[225],"in":[229],"model,":[232],"also":[234],"introduce":[235],"new":[237,254],"proof":[238,258],"technique":[239],"finds":[241],"certain":[243],"indistinguishable":[244],"decomposition":[245],"possible":[248],"inputs":[249],"iteratively.":[250],"We":[251],"believe":[252],"techniques":[259],"will":[260],"helpful":[262],"understand":[264],"computational":[266],"power":[267],"transformers":[269],"further.":[270]},"counts_by_year":[],"updated_date":"2026-02-13T13:36:01.753593","created_date":"2026-02-12T00:00:00"}