{"id":"https://openalex.org/W7135232160","doi":"https://doi.org/10.48550/arxiv.2603.11535","title":"Expert Threshold Routing for Autoregressive Language Modeling with Dynamic Computation Allocation and Load Balancing","display_name":"Expert Threshold Routing for Autoregressive Language Modeling with Dynamic Computation Allocation and Load Balancing","publication_year":2026,"publication_date":"2026-03-12","ids":{"openalex":"https://openalex.org/W7135232160","doi":"https://doi.org/10.48550/arxiv.2603.11535"},"language":"en","primary_location":{"id":"pmh:oai:arXiv.org:2603.11535","doi":null,"is_oa":true,"landing_page_url":"http://arxiv.org/abs/2603.11535","pdf_url":"https://arxiv.org/pdf/2603.11535","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","license_id":"https://openalex.org/licenses/cc-by","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/2603.11535","any_repository_has_fulltext":true},"authorships":[{"author_position":"first","author":{"id":"https://openalex.org/A5129028445","display_name":"Hanchi Sun","orcid":"https://orcid.org/0009-0000-3267-1093","observed_orcids":["https://orcid.org/0009-0000-3267-1093"]},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Sun, Hanchi","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5128988119","display_name":"Yixin Liu","orcid":null,"observed_orcids":[]},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Liu, Yixin","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5129072670","display_name":"Yonghui Wu","orcid":null,"observed_orcids":[]},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Wu, Yonghui","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"last","author":{"id":"https://openalex.org/A5133187192","display_name":"Lichao Sun","orcid":null,"observed_orcids":[]},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Sun, Lichao","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]}],"institutions":[],"countries_distinct_count":0,"institutions_distinct_count":0,"corresponding_author_ids":[],"corresponding_institution_ids":[],"apc_list":null,"apc_paid":null,"fwci":null,"has_fulltext":true,"cited_by_count":6,"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":0.47749999165534973,"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":0.47749999165534973,"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":0.2321999967098236,"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/T10036","display_name":"Advanced Neural Network Applications","score":0.1599999964237213,"subfield":{"id":"https://openalex.org/subfields/1707","display_name":"Computer Vision and Pattern Recognition"},"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":0.8679999709129333},{"id":"https://openalex.org/keywords/load-balancing","display_name":"load balancing","score":0.9739999771118164}],"concepts":[{"id":"https://openalex.org/C48145219","wikidata":"https://www.wikidata.org/wiki/Q1335365","display_name":"Security token","level":2,"score":0.8075000047683716},{"id":"https://openalex.org/C41008148","wikidata":"https://www.wikidata.org/wiki/Q21198","display_name":"Computer science","level":0,"score":0.7572000026702881},{"id":"https://openalex.org/C45374587","wikidata":"https://www.wikidata.org/wiki/Q12525525","display_name":"Computation","level":2,"score":0.7038999795913696},{"id":"https://openalex.org/C159877910","wikidata":"https://www.wikidata.org/wiki/Q2202883","display_name":"Autoregressive model","level":2,"score":0.6227999925613403},{"id":"https://openalex.org/C138959212","wikidata":"https://www.wikidata.org/wiki/Q1806783","display_name":"Load balancing (electrical power)","level":3,"score":0.4156000018119812},{"id":"https://openalex.org/C58328972","wikidata":"https://www.wikidata.org/wiki/Q184609","display_name":"Expert system","level":2,"score":0.4068000018596649},{"id":"https://openalex.org/C126255220","wikidata":"https://www.wikidata.org/wiki/Q141495","display_name":"Mathematical optimization","level":1,"score":0.38339999318122864},{"id":"https://openalex.org/C99844830","wikidata":"https://www.wikidata.org/wiki/Q102441924","display_name":"Scaling","level":2,"score":0.3765999972820282},{"id":"https://openalex.org/C151376022","wikidata":"https://www.wikidata.org/wiki/Q168698","display_name":"Exponential function","level":2,"score":0.3686000108718872},{"id":"https://openalex.org/C74172769","wikidata":"https://www.wikidata.org/wiki/Q1446839","display_name":"Routing (electronic design automation)","level":2,"score":0.36579999327659607},{"id":"https://openalex.org/C125411270","wikidata":"https://www.wikidata.org/wiki/Q18653","display_name":"Encoding (memory)","level":2,"score":0.36169999837875366},{"id":"https://openalex.org/C188198153","wikidata":"https://www.wikidata.org/wiki/Q1613840","display_name":"Limiting","level":2,"score":0.35040000081062317},{"id":"https://openalex.org/C137293760","wikidata":"https://www.wikidata.org/wiki/Q3621696","display_name":"Language model","level":2,"score":0.3465999960899353},{"id":"https://openalex.org/C24856439","wikidata":"https://www.wikidata.org/wiki/Q352483","display_name":"Adaptive routing","level":5,"score":0.3393000066280365},{"id":"https://openalex.org/C2779370713","wikidata":"https://www.wikidata.org/wiki/Q357554","display_name":"Load management","level":2,"score":0.31459999084472656},{"id":"https://openalex.org/C77618280","wikidata":"https://www.wikidata.org/wiki/Q1155772","display_name":"Scheme (mathematics)","level":2,"score":0.31290000677108765},{"id":"https://openalex.org/C11413529","wikidata":"https://www.wikidata.org/wiki/Q8366","display_name":"Algorithm","level":1,"score":0.2849999964237213},{"id":"https://openalex.org/C37404715","wikidata":"https://www.wikidata.org/wiki/Q380679","display_name":"Dynamic programming","level":2,"score":0.2761000096797943},{"id":"https://openalex.org/C49937458","wikidata":"https://www.wikidata.org/wiki/Q2599292","display_name":"Probabilistic logic","level":2,"score":0.2572999894618988},{"id":"https://openalex.org/C168031717","wikidata":"https://www.wikidata.org/wiki/Q1530280","display_name":"Balance (ability)","level":2,"score":0.25110000371932983}],"mesh":[],"locations_count":2,"locations":[{"id":"pmh:oai:arXiv.org:2603.11535","doi":null,"is_oa":true,"landing_page_url":"http://arxiv.org/abs/2603.11535","pdf_url":"https://arxiv.org/pdf/2603.11535","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","license_id":"https://openalex.org/licenses/cc-by","version":"submittedVersion","is_accepted":false,"is_published":false,"raw_source_name":null,"raw_type":"text"},{"id":"doi:10.48550/arxiv.2603.11535","doi":"https://doi.org/10.48550/arxiv.2603.11535","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2603.11535","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","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":"pmh:oai:arXiv.org:2603.11535","doi":null,"is_oa":true,"landing_page_url":"http://arxiv.org/abs/2603.11535","pdf_url":"https://arxiv.org/pdf/2603.11535","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","license_id":"https://openalex.org/licenses/cc-by","version":"submittedVersion","is_accepted":false,"is_published":false,"raw_source_name":null,"raw_type":"text"},"sustainable_development_goals":[],"study_designs":[],"awards":[],"funders":[{"id":"https://openalex.org/F4320309480","display_name":"Nvidia","ror":"https://ror.org/03jdj4y14"}],"has_content":{"grobid_xml":true,"pdf":true},"content_urls":{"pdf":"https://content.openalex.org/works/W7135232160.pdf","grobid_xml":"https://content.openalex.org/works/W7135232160.grobid-xml"},"referenced_works_count":24,"referenced_works":["https://openalex.org/W2581624817","https://openalex.org/W3177813494","https://openalex.org/W3210904070","https://openalex.org/W4311252752","https://openalex.org/W4388979610","https://openalex.org/W4389524599","https://openalex.org/W4391376534","https://openalex.org/W4392822410","https://openalex.org/W4393967974","https://openalex.org/W4398796940","https://openalex.org/W4401306836","https://openalex.org/W4401306979","https://openalex.org/W4402670023","https://openalex.org/W4402671950","https://openalex.org/W4403322029","https://openalex.org/W4404781605","https://openalex.org/W4405903187","https://openalex.org/W4412886668","https://openalex.org/W4414939355","https://openalex.org/W4415798713","https://openalex.org/W4416035996","https://openalex.org/W4416691570","https://openalex.org/W6929596217","https://openalex.org/W7125503976"],"related_works":["https://openalex.org/W4399353875","https://openalex.org/W7221679618","https://openalex.org/W7155001241","https://openalex.org/W7160847478","https://openalex.org/W7148955909","https://openalex.org/W4404781605","https://openalex.org/W7166831075","https://openalex.org/W4412888297","https://openalex.org/W7151295702","https://openalex.org/W7148613496"],"abstract_inverted_index":{"Token-choice":[0],"Mixture-of-Experts":[1],"(TC-MoE)":[2],"routes":[3],"each":[4,31,51],"token":[5,44,52],"to":[6,20,56,100,114],"a":[7],"fixed":[8],"number":[9],"of":[10],"experts,":[11],"limiting":[12],"dynamic":[13,67],"computation":[14,68],"allocation":[15,69],"and":[16,49],"requiring":[17],"auxiliary":[18,75],"losses":[19],"maintain":[21],"load":[22,72],"balance.":[23],"We":[24],"propose":[25],"Expert":[26],"Threshold":[27],"(ET)":[28],"routing,":[29],"where":[30],"expert":[32,58],"maintains":[33],"an":[34,57],"exponential":[35],"moving":[36],"average":[37],"(EMA)":[38],"threshold":[39],"estimated":[40],"from":[41],"the":[42,63,87,116],"global":[43],"distribution.":[45],"At":[46],"both":[47],"training":[48],"inference,":[50],"is":[53],"independently":[54],"routed":[55],"if":[59],"its":[60],"score":[61],"exceeds":[62],"expert's":[64],"threshold,":[65],"enabling":[66],"while":[70],"achieving":[71],"balance":[73],"without":[74],"losses.":[76],"This":[77],"fully":[78],"causal":[79],"mechanism":[80],"eliminates":[81],"dependence":[82],"on":[83,103],"other":[84],"tokens":[85],"in":[86],"batch,":[88],"making":[89],"it":[90],"well-suited":[91],"for":[92],"autoregressive":[93],"language":[94],"modeling.":[95],"In":[96],"pretraining":[97],"experiments":[98],"scaling":[99],"2.4B":[101],"parameters":[102],"FineWeb-Edu,":[104],"ET":[105],"achieves":[106],"0.067":[107],"lower":[108],"cross-entropy":[109],"loss":[110],"than":[111],"TC-MoE,":[112],"equivalent":[113],"reaching":[115],"same":[117],"performance":[118],"with":[119],"1.6$\\times$":[120],"fewer":[121],"tokens.":[122]},"counts_by_year":[{"year":2026,"cited_by_count":6}],"updated_date":"2026-11-07T11:25:07.637180","created_date":"2026-03-14T00:00:00"}