{"id":"https://openalex.org/W7208747583","doi":"https://doi.org/10.48550/arxiv.2609.03501","title":"Towards a Statistical Understanding of Mixture-of-Experts","display_name":"Towards a Statistical Understanding of Mixture-of-Experts","publication_year":2026,"publication_date":"2026-09-03","ids":{"openalex":"https://openalex.org/W7208747583","doi":"https://doi.org/10.48550/arxiv.2609.03501"},"language":"en","primary_location":{"id":"pmh:oai:arXiv.org:2609.03501","doi":null,"is_oa":true,"landing_page_url":"https://arxiv.org/abs/2609.03501","pdf_url":"https://arxiv.org/pdf/2609.03501","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/2609.03501","any_repository_has_fulltext":true},"authorships":[{"author_position":"first","author":{"id":"https://openalex.org/A5149986000","display_name":"Siyuan He","orcid":null,"observed_orcids":[]},"institutions":[{"id":"https://openalex.org/I99065089","display_name":"Tsinghua University","ror":"https://ror.org/03cve4549","country_code":"CN","type":"education","lineage":["https://openalex.org/I99065089"]}],"countries":["CN"],"is_corresponding":false,"raw_author_name":"He, Siyuan","raw_affiliation_strings":["Qiuzhen College , Tsinghua University , Beijing , China"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"Qiuzhen College , Tsinghua University , Beijing , China","institution_ids":["https://openalex.org/I99065089"]}]},{"author_position":"middle","author":{"id":"https://openalex.org/A5150042537","display_name":"Bokai Yang","orcid":null,"observed_orcids":[]},"institutions":[{"id":"https://openalex.org/I99065089","display_name":"Tsinghua University","ror":"https://ror.org/03cve4549","country_code":"CN","type":"education","lineage":["https://openalex.org/I99065089"]}],"countries":["CN"],"is_corresponding":false,"raw_author_name":"Yang, Bokai","raw_affiliation_strings":["Qiuzhen College , Tsinghua University , Beijing , China"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"Qiuzhen College , Tsinghua University , Beijing , China","institution_ids":["https://openalex.org/I99065089"]}]},{"author_position":"middle","author":{"id":"https://openalex.org/A5150021890","display_name":"Jie Hu","orcid":null,"observed_orcids":[]},"institutions":[{"id":"https://openalex.org/I99065089","display_name":"Tsinghua University","ror":"https://ror.org/03cve4549","country_code":"CN","type":"education","lineage":["https://openalex.org/I99065089"]}],"countries":["CN"],"is_corresponding":false,"raw_author_name":"Hu, Jie","raw_affiliation_strings":["Yau Mathematical Sciences Center , Tsinghua University , Beijing , China"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"Yau Mathematical Sciences Center , Tsinghua University , Beijing , China","institution_ids":["https://openalex.org/I99065089"]}]},{"author_position":"middle","author":{"id":"https://openalex.org/A5136147011","display_name":"Ziwen Gao","orcid":null,"observed_orcids":[]},"institutions":[{"id":"https://openalex.org/I66867065","display_name":"East China Normal University","ror":"https://ror.org/02n96ep67","country_code":"CN","type":"education","lineage":["https://openalex.org/I66867065"]}],"countries":["CN"],"is_corresponding":false,"raw_author_name":"Gao, Ziwen","raw_affiliation_strings":["KLATASDS-MOE , School of Statistics , East China Normal University , Shanghai , China"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"KLATASDS-MOE , School of Statistics , East China Normal University , Shanghai , China","institution_ids":["https://openalex.org/I66867065"]}]},{"author_position":"last","author":{"id":"https://openalex.org/A5150041454","display_name":"Yuhong Yang","orcid":null,"observed_orcids":[]},"institutions":[{"id":"https://openalex.org/I4403928416","display_name":"Beijing Institute of Mathematical Sciences and Applications","ror":"https://ror.org/05t6hvr95","country_code":"CN","type":"education","lineage":["https://openalex.org/I4403928416","https://openalex.org/I99065089"]},{"id":"https://openalex.org/I99065089","display_name":"Tsinghua University","ror":"https://ror.org/03cve4549","country_code":"CN","type":"education","lineage":["https://openalex.org/I99065089"]}],"countries":["CN"],"is_corresponding":false,"raw_author_name":"Yang, Yuhong","raw_affiliation_strings":["Beijing Institute of Mathematical Sciences and Applications (BIMSA) , Beijing , China","Yau Mathematical Sciences Center , Tsinghua University , Beijing , China"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"Beijing Institute of Mathematical Sciences and Applications (BIMSA) , Beijing , China","institution_ids":["https://openalex.org/I4403928416"]},{"raw_affiliation_string":"Yau Mathematical Sciences Center , Tsinghua University , Beijing , China","institution_ids":["https://openalex.org/I99065089"]}]}],"institutions":[],"countries_distinct_count":1,"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":0.1914999932050705,"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":0.1914999932050705,"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/T10320","display_name":"Neural Networks and Applications","score":0.19120000302791595,"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":0.17069999873638153,"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/mixture-of-experts","display_name":"mixture of experts","score":0.9879999756813049},{"id":"https://openalex.org/keywords/sparse-activation","display_name":"sparse activation","score":0.8799999952316284},{"id":"https://openalex.org/keywords/top-k-routing","display_name":"top-k routing","score":0.7250000238418579},{"id":"https://openalex.org/keywords/gating","display_name":"gating","score":0.5640000104904175},{"id":"https://openalex.org/keywords/expert-specialization","display_name":"expert specialization","score":0.7379999756813049}],"concepts":[{"id":"https://openalex.org/C41008148","wikidata":"https://www.wikidata.org/wiki/Q21198","display_name":"Computer science","level":0,"score":0.623199999332428},{"id":"https://openalex.org/C55166926","wikidata":"https://www.wikidata.org/wiki/Q2892946","display_name":"Oracle","level":2,"score":0.5679000020027161},{"id":"https://openalex.org/C117251300","wikidata":"https://www.wikidata.org/wiki/Q1849855","display_name":"Parametric statistics","level":2,"score":0.5480999946594238},{"id":"https://openalex.org/C74172769","wikidata":"https://www.wikidata.org/wiki/Q1446839","display_name":"Routing (electronic design automation)","level":2,"score":0.507099986076355},{"id":"https://openalex.org/C2982736386","wikidata":"https://www.wikidata.org/wiki/Q2539","display_name":"Statistical learning","level":2,"score":0.5040000081062317},{"id":"https://openalex.org/C119857082","wikidata":"https://www.wikidata.org/wiki/Q2539","display_name":"Machine learning","level":1,"score":0.49309998750686646},{"id":"https://openalex.org/C192209626","wikidata":"https://www.wikidata.org/wiki/Q190909","display_name":"Focus (optics)","level":2,"score":0.46650001406669617},{"id":"https://openalex.org/C154945302","wikidata":"https://www.wikidata.org/wiki/Q11660","display_name":"Artificial intelligence","level":1,"score":0.43959999084472656},{"id":"https://openalex.org/C114289077","wikidata":"https://www.wikidata.org/wiki/Q3284399","display_name":"Statistical model","level":2,"score":0.41359999775886536},{"id":"https://openalex.org/C155512373","wikidata":"https://www.wikidata.org/wiki/Q287450","display_name":"Residual","level":2,"score":0.4016999900341034},{"id":"https://openalex.org/C124101348","wikidata":"https://www.wikidata.org/wiki/Q172491","display_name":"Data mining","level":1,"score":0.3287999927997589},{"id":"https://openalex.org/C22367795","wikidata":"https://www.wikidata.org/wiki/Q7625208","display_name":"Structured prediction","level":2,"score":0.32019999623298645},{"id":"https://openalex.org/C2522767166","wikidata":"https://www.wikidata.org/wiki/Q2374463","display_name":"Data science","level":1,"score":0.29580000042915344},{"id":"https://openalex.org/C80444323","wikidata":"https://www.wikidata.org/wiki/Q2878974","display_name":"Theoretical computer science","level":1,"score":0.29330000281333923},{"id":"https://openalex.org/C58328972","wikidata":"https://www.wikidata.org/wiki/Q184609","display_name":"Expert system","level":2,"score":0.28760001063346863},{"id":"https://openalex.org/C155846161","wikidata":"https://www.wikidata.org/wiki/Q1143367","display_name":"Graphical model","level":2,"score":0.26930001378059387},{"id":"https://openalex.org/C2778334786","wikidata":"https://www.wikidata.org/wiki/Q1586270","display_name":"Variation (astronomy)","level":2,"score":0.2565999925136566},{"id":"https://openalex.org/C194544171","wikidata":"https://www.wikidata.org/wiki/Q21105679","display_name":"Gating","level":2,"score":0.2563000023365021},{"id":"https://openalex.org/C45340560","wikidata":"https://www.wikidata.org/wiki/Q215382","display_name":"Disjoint sets","level":2,"score":0.25209999084472656}],"mesh":[],"locations_count":2,"locations":[{"id":"pmh:oai:arXiv.org:2609.03501","doi":null,"is_oa":true,"landing_page_url":"https://arxiv.org/abs/2609.03501","pdf_url":"https://arxiv.org/pdf/2609.03501","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.2609.03501","doi":"https://doi.org/10.48550/arxiv.2609.03501","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2609.03501","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:2609.03501","doi":null,"is_oa":true,"landing_page_url":"https://arxiv.org/abs/2609.03501","pdf_url":"https://arxiv.org/pdf/2609.03501","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":{"pdf":true,"grobid_xml":true},"content_urls":{"pdf":"https://content.openalex.org/works/W7208747583.pdf","grobid_xml":"https://content.openalex.org/works/W7208747583.grobid-xml"},"referenced_works_count":7,"referenced_works":["https://openalex.org/W4390723197","https://openalex.org/W4393969316","https://openalex.org/W4400434490","https://openalex.org/W4417093066","https://openalex.org/W7105141428","https://openalex.org/W7160847478","https://openalex.org/W7161147422"],"related_works":["https://openalex.org/W7221828767","https://openalex.org/W4416532856","https://openalex.org/W7139967476","https://openalex.org/W7134266197","https://openalex.org/W4417093066","https://openalex.org/W4417287601","https://openalex.org/W7222598568","https://openalex.org/W7160635260","https://openalex.org/W7164004074","https://openalex.org/W7083826191"],"abstract_inverted_index":{"Mixture-of-experts":[0],"(MoE)":[1],"architectures":[2,152],"increase":[3],"model":[4],"capacity":[5],"by":[6,194],"combining":[7],"a":[8,20,74,174],"collection":[9],"of":[10,23,39,76,118,132,140],"expert":[11,142,183,187],"predictors":[12],"through":[13,129,181],"input-dependent":[14,182],"routing,":[15,44],"while":[16,121],"often":[17],"activating":[18],"only":[19,51],"small":[21],"subset":[22],"experts":[24,164],"for":[25,93,178],"each":[26],"input.":[27],"Despite":[28],"their":[29,40],"growing":[30],"importance":[31],"in":[32,151,185],"modern":[33],"large-scale":[34],"models,":[35],"the":[36,85,116,130],"statistical":[37,176],"roles":[38],"design":[41],"choices,":[42],"especially":[43],"sparse":[45,97,111],"activation,":[46],"and":[47,79,96,105,108,144,189],"shared":[48,147],"experts,":[49,101,148],"remain":[50],"partially":[52],"understood,":[53],"as":[54,73,149,154],"existing":[55],"theory":[56],"has":[57],"largely":[58],"focused":[59],"on":[60,166],"parametric":[61],"or":[62],"correctly":[63],"specified":[64],"MoE":[65,72,180],"models.":[66],"In":[67],"this":[68,82],"paper,":[69],"we":[70],"view":[71],"form":[75],"localized":[77,119],"aggregation":[78,120],"show":[80,145],"how":[81,110,146],"localization":[83],"reshapes":[84],"approximation-estimation-computation":[86],"tradeoff.":[87],"We":[88,125],"derive":[89],"oracle":[90],"risk":[91],"bounds":[92],"learning":[94],"dense":[95],"routing":[98,113,136],"with":[99],"evolving":[100],"separating":[102],"approximation,":[103],"expert-learning,":[104],"router-estimation":[106],"errors,":[107],"characterize":[109],"Top-K":[112],"can":[114,156],"retain":[115],"benefits":[117],"controlling":[122],"per-input":[123],"computation.":[124],"also":[126],"interpret":[127],"gating":[128],"geometry":[131],"input":[133],"space,":[134],"relating":[135],"performance":[137],"to":[138],"regions":[139],"local":[141,168,195],"advantage,":[143],"adopted":[150],"such":[153],"DeepSeekMoE,":[155],"extract":[157],"common":[158],"predictive":[159,196],"structure":[160],"so":[161],"that":[162],"routed":[163],"focus":[165],"residual":[167],"variation.":[169],"Together,":[170],"these":[171],"results":[172],"provide":[173],"unified":[175],"framework":[177],"understanding":[179],"aggregation,":[184],"which":[186],"specialization":[188],"computational":[190],"tradeoffs":[191],"are":[192],"governed":[193],"structure.":[197]},"counts_by_year":[],"updated_date":"2026-11-07T11:25:07.637180","created_date":"2026-09-05T00:00:00"}