{"id":"https://openalex.org/W4415248342","doi":"https://doi.org/10.18653/v1/2026.findings-acl.2140","title":"Faster MoE LLM Inference for Extremely Large Models","display_name":"Faster MoE LLM Inference for Extremely Large Models","publication_year":2026,"publication_date":"2026-01-01","ids":{"openalex":"https://openalex.org/W4415248342","doi":"https://doi.org/10.18653/v1/2026.findings-acl.2140"},"language":"en","primary_location":{"id":"doi:10.18653/v1/2026.findings-acl.2140","doi":"https://doi.org/10.18653/v1/2026.findings-acl.2140","is_oa":true,"landing_page_url":"https://doi.org/10.18653/v1/2026.findings-acl.2140","pdf_url":"https://aclanthology.org/2026.findings-acl.2140.pdf","source":{"id":"https://openalex.org/S4363605144","display_name":"Findings of the Association for Computational Linguistics: ACL (Findings of ACL)","issn_l":null,"issn":null,"is_oa":false,"is_in_doaj":false,"is_core":false,"listed_in":[],"host_organization":null,"host_organization_name":null,"host_organization_lineage":[],"host_organization_lineage_names":[],"type":"conference"},"license":"cc-by","license_id":"https://openalex.org/licenses/cc-by","version":"publishedVersion","is_accepted":true,"is_published":true,"raw_source_name":"Findings of the Association for Computational Linguistics: ACL 2026","raw_type":"proceedings-article"},"volume":null,"type":"conference-paper","indexed_in":["arxiv","crossref","datacite"],"open_access":{"is_oa":true,"oa_status":"hybrid","oa_url":"https://aclanthology.org/2026.findings-acl.2140.pdf","any_repository_has_fulltext":true},"authorships":[{"author_position":"first","author":{"id":"https://openalex.org/A5121529722","display_name":"Haoqi Yang","orcid":null,"observed_orcids":[]},"institutions":[{"id":"https://openalex.org/I37461747","display_name":"Wuhan University","ror":"https://ror.org/033vjfk17","country_code":"CN","type":"education","lineage":["https://openalex.org/I37461747"]}],"countries":["CN"],"is_corresponding":false,"raw_author_name":"Haoqi Yang","raw_affiliation_strings":["School of Computer Science , Wuhan University , Wuhan , China"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"School of Computer Science , Wuhan University , Wuhan , China","institution_ids":["https://openalex.org/I37461747"]}]},{"author_position":"middle","author":{"id":"https://openalex.org/A5119181873","display_name":"Shi Luohe","orcid":null,"observed_orcids":[]},"institutions":[{"id":"https://openalex.org/I37461747","display_name":"Wuhan University","ror":"https://ror.org/033vjfk17","country_code":"CN","type":"education","lineage":["https://openalex.org/I37461747"]}],"countries":["CN"],"is_corresponding":false,"raw_author_name":"Luohe Shi","raw_affiliation_strings":["School of Computer Science , Wuhan University , Wuhan , China"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"School of Computer Science , Wuhan University , Wuhan , China","institution_ids":["https://openalex.org/I37461747"]}]},{"author_position":"middle","author":{"id":"https://openalex.org/A5100640251","display_name":"Qiwei Li","orcid":"https://orcid.org/0000-0002-1020-3050","observed_orcids":["https://orcid.org/0000-0002-1020-3050","https://orcid.org/0009-0001-1527-4931"]},"institutions":[{"id":"https://openalex.org/I37461747","display_name":"Wuhan University","ror":"https://ror.org/033vjfk17","country_code":"CN","type":"education","lineage":["https://openalex.org/I37461747"]}],"countries":["CN"],"is_corresponding":false,"raw_author_name":"Qiwei Li","raw_affiliation_strings":["School of Computer Science , Wuhan University , Wuhan , China"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"School of Computer Science , Wuhan University , Wuhan , China","institution_ids":["https://openalex.org/I37461747"]}]},{"author_position":"middle","author":{"id":"https://openalex.org/A5002584135","display_name":"Zuchao Li","orcid":"https://orcid.org/0000-0003-0436-8446","observed_orcids":["https://orcid.org/0000-0003-0436-8446"]},"institutions":[{"id":"https://openalex.org/I37461747","display_name":"Wuhan University","ror":"https://ror.org/033vjfk17","country_code":"CN","type":"education","lineage":["https://openalex.org/I37461747"]}],"countries":["CN"],"is_corresponding":false,"raw_author_name":"Zuchao Li","raw_affiliation_strings":["School of Artificial Intelligence , Wuhan University , Wuhan , China"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"School of Artificial Intelligence , Wuhan University , Wuhan , China","institution_ids":["https://openalex.org/I37461747"]}]},{"author_position":"middle","author":{"id":"https://openalex.org/A5100338689","display_name":"Ping Wang","orcid":"https://orcid.org/0000-0002-8854-2079","observed_orcids":["https://orcid.org/0000-0002-8854-2079"]},"institutions":[{"id":"https://openalex.org/I37461747","display_name":"Wuhan University","ror":"https://ror.org/033vjfk17","country_code":"CN","type":"education","lineage":["https://openalex.org/I37461747"]}],"countries":["CN"],"is_corresponding":false,"raw_author_name":"Ping Wang","raw_affiliation_strings":["School of Information Management , Wuhan University , Wuhan , China"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"School of Information Management , Wuhan University , Wuhan , China","institution_ids":["https://openalex.org/I37461747"]}]},{"author_position":"middle","author":{"id":"https://openalex.org/A5143296774","display_name":"Hao Huang","orcid":null,"observed_orcids":[]},"institutions":[{"id":"https://openalex.org/I37461747","display_name":"Wuhan University","ror":"https://ror.org/033vjfk17","country_code":"CN","type":"education","lineage":["https://openalex.org/I37461747"]}],"countries":["CN"],"is_corresponding":false,"raw_author_name":"Hao Huang","raw_affiliation_strings":["School of Computer Science , Wuhan University , Wuhan , China"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"School of Computer Science , Wuhan University , Wuhan , China","institution_ids":["https://openalex.org/I37461747"]}]},{"author_position":"last","author":{"id":"https://openalex.org/A5003608372","display_name":"Hongyun Zhao","orcid":"https://orcid.org/0000-0002-1251-3888","observed_orcids":["https://orcid.org/0000-0002-1251-3888","https://orcid.org/0009-0006-0560-8375","https://orcid.org/0000-0003-2938-5613","https://orcid.org/0000-0002-3539-3532","https://orcid.org/0000-0001-6039-9736","https://orcid.org/0000-0002-8879-9595","https://orcid.org/0000-0001-7903-581X","https://orcid.org/0000-0003-4756-771X","https://orcid.org/0000-0001-6618-3039","https://orcid.org/0000-0001-8554-9411","https://orcid.org/0000-0002-4575-811X","https://orcid.org/0000-0002-6298-1217","https://orcid.org/0000-0002-7890-0055","https://orcid.org/0000-0003-4746-389X","https://orcid.org/0009-0005-8958-7008","https://orcid.org/0009-0005-9482-8953","https://orcid.org/0009-0006-0216-3307","https://orcid.org/0009-0008-5370-813X","https://orcid.org/0000-0001-5407-2569","https://orcid.org/0000-0002-4438-7791","https://orcid.org/0000-0002-7765-2985","https://orcid.org/0000-0003-2176-9030","https://orcid.org/0000-0003-2384-1606","https://orcid.org/0009-0000-6368-5359","https://orcid.org/0009-0003-4754-1701","https://orcid.org/0009-0003-6018-5115","https://orcid.org/0009-0003-9594-2358","https://orcid.org/0009-0003-9797-6746","https://orcid.org/0009-0004-0308-6214","https://orcid.org/0009-0006-0226-1804","https://orcid.org/0009-0006-6579-8131","https://orcid.org/0009-0006-8593-6079","https://orcid.org/0009-0007-6680-596X","https://orcid.org/0009-0008-5350-8429","https://orcid.org/0009-0008-7243-9799","https://orcid.org/0009-0008-9829-3900"]},"institutions":[{"id":"https://openalex.org/I183067930","display_name":"Shanghai Jiao Tong University","ror":"https://ror.org/0220qvk04","country_code":"CN","type":"education","lineage":["https://openalex.org/I183067930"]}],"countries":["CN"],"is_corresponding":false,"raw_author_name":"Hai Zhao","raw_affiliation_strings":["School of Computer Science , Shanghai Jiao Tong University , Shanghai , China"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"School of Computer Science , Shanghai Jiao Tong University , Shanghai , China","institution_ids":["https://openalex.org/I183067930"]}]}],"institutions":[],"countries_distinct_count":1,"institutions_distinct_count":2,"corresponding_author_ids":[],"corresponding_institution_ids":[],"apc_list":null,"apc_paid":null,"fwci":0.0,"has_fulltext":true,"cited_by_count":0,"citation_normalized_percentile":{"value":0.00701414,"is_in_top_1_percent":false,"is_in_top_10_percent":false},"cited_by_percentile_year":null,"biblio":{"volume":null,"issue":null,"first_page":"43133","last_page":"43151"},"is_retracted":false,"is_paratext":false,"is_xpac":false,"primary_topic":{"id":"https://openalex.org/T10036","display_name":"Advanced Neural Network Applications","score":0.875,"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"}},"topics":[{"id":"https://openalex.org/T10036","display_name":"Advanced Neural Network Applications","score":0.875,"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"}},{"id":"https://openalex.org/T10028","display_name":"Topic Modeling","score":0.04270000010728836,"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/T10054","display_name":"Parallel Computing and Optimization Techniques","score":0.023499999195337296,"subfield":{"id":"https://openalex.org/subfields/1708","display_name":"Hardware and Architecture"},"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":0.9599999785423279},{"id":"https://openalex.org/keywords/inference-throughput","display_name":"inference throughput","score":0.5889999866485596},{"id":"https://openalex.org/keywords/large-language-models","display_name":"large language models","score":0.5289999842643738},{"id":"https://openalex.org/keywords/deepseek-v3","display_name":"DeepSeek-V3","score":0.7730000019073486},{"id":"https://openalex.org/keywords/accuracy-efficiency-trade-off","display_name":"accuracy-efficiency trade-off","score":0.7160000205039978}],"concepts":[{"id":"https://openalex.org/C41008148","wikidata":"https://www.wikidata.org/wiki/Q21198","display_name":"Computer science","level":0,"score":0.5794000029563904},{"id":"https://openalex.org/C2776214188","wikidata":"https://www.wikidata.org/wiki/Q408386","display_name":"Inference","level":2,"score":0.5733000040054321},{"id":"https://openalex.org/C111335779","wikidata":"https://www.wikidata.org/wiki/Q3454686","display_name":"Reduction (mathematics)","level":2,"score":0.5145999789237976},{"id":"https://openalex.org/C157764524","wikidata":"https://www.wikidata.org/wiki/Q1383412","display_name":"Throughput","level":3,"score":0.4499000012874603},{"id":"https://openalex.org/C119857082","wikidata":"https://www.wikidata.org/wiki/Q2539","display_name":"Machine learning","level":1,"score":0.3540000021457672},{"id":"https://openalex.org/C154945302","wikidata":"https://www.wikidata.org/wiki/Q11660","display_name":"Artificial intelligence","level":1,"score":0.31310001015663147},{"id":"https://openalex.org/C2780378061","wikidata":"https://www.wikidata.org/wiki/Q25351891","display_name":"Service (business)","level":2,"score":0.3050000071525574},{"id":"https://openalex.org/C126255220","wikidata":"https://www.wikidata.org/wiki/Q141495","display_name":"Mathematical optimization","level":1,"score":0.29899999499320984},{"id":"https://openalex.org/C127413603","wikidata":"https://www.wikidata.org/wiki/Q11023","display_name":"Engineering","level":0,"score":0.27300000190734863},{"id":"https://openalex.org/C165838908","wikidata":"https://www.wikidata.org/wiki/Q736777","display_name":"Calibration","level":2,"score":0.2671000063419342},{"id":"https://openalex.org/C61797465","wikidata":"https://www.wikidata.org/wiki/Q1188986","display_name":"Term (time)","level":2,"score":0.2662000060081482},{"id":"https://openalex.org/C42475967","wikidata":"https://www.wikidata.org/wiki/Q194292","display_name":"Operations research","level":1,"score":0.2624000012874603}],"mesh":[],"locations_count":3,"locations":[{"id":"doi:10.18653/v1/2026.findings-acl.2140","doi":"https://doi.org/10.18653/v1/2026.findings-acl.2140","is_oa":true,"landing_page_url":"https://doi.org/10.18653/v1/2026.findings-acl.2140","pdf_url":"https://aclanthology.org/2026.findings-acl.2140.pdf","source":{"id":"https://openalex.org/S4363605144","display_name":"Findings of the Association for Computational Linguistics: ACL (Findings of ACL)","issn_l":null,"issn":null,"is_oa":false,"is_in_doaj":false,"is_core":false,"listed_in":[],"host_organization":null,"host_organization_name":null,"host_organization_lineage":[],"host_organization_lineage_names":[],"type":"conference"},"license":"cc-by","license_id":"https://openalex.org/licenses/cc-by","version":"publishedVersion","is_accepted":true,"is_published":true,"raw_source_name":"Findings of the Association for Computational Linguistics: ACL 2026","raw_type":"proceedings-article"},{"id":"pmh:oai:arXiv.org:2505.03531","doi":null,"is_oa":true,"landing_page_url":"http://arxiv.org/abs/2505.03531","pdf_url":"https://arxiv.org/pdf/2505.03531","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.2505.03531","doi":"https://doi.org/10.48550/arxiv.2505.03531","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2505.03531","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":"doi:10.18653/v1/2026.findings-acl.2140","doi":"https://doi.org/10.18653/v1/2026.findings-acl.2140","is_oa":true,"landing_page_url":"https://doi.org/10.18653/v1/2026.findings-acl.2140","pdf_url":"https://aclanthology.org/2026.findings-acl.2140.pdf","source":{"id":"https://openalex.org/S4363605144","display_name":"Findings of the Association for Computational Linguistics: ACL (Findings of ACL)","issn_l":null,"issn":null,"is_oa":false,"is_in_doaj":false,"is_core":false,"listed_in":[],"host_organization":null,"host_organization_name":null,"host_organization_lineage":[],"host_organization_lineage_names":[],"type":"conference"},"license":"cc-by","license_id":"https://openalex.org/licenses/cc-by","version":"publishedVersion","is_accepted":true,"is_published":true,"raw_source_name":"Findings of the Association for Computational Linguistics: ACL 2026","raw_type":"proceedings-article"},"sustainable_development_goals":[],"sustainable_development_goals_aurora":[],"study_designs":[],"awards":[{"id":"https://openalex.org/G123088199","display_name":null,"funder_award_id":"24&ZD186","funder_id":"https://openalex.org/F4320335869","funder_display_name":"National Social Science Fund of China"},{"id":"https://openalex.org/G3551126291","display_name":null,"funder_award_id":"24&amp","funder_id":"https://openalex.org/F4320335869","funder_display_name":"National Social Science Fund of China"},{"id":"https://openalex.org/G4610279030","display_name":null,"funder_award_id":"2023ZD0121502","funder_id":"https://openalex.org/F4320329860","funder_display_name":"National Science and Technology Major Project"},{"id":"https://openalex.org/G7131482003","display_name":null,"funder_award_id":"24&amp","funder_id":"https://openalex.org/F4320321001","funder_display_name":"National Natural Science Foundation of China"},{"id":"https://openalex.org/G7898550485","display_name":null,"funder_award_id":"62306216","funder_id":"https://openalex.org/F4320321001","funder_display_name":"National Natural Science Foundation of China"},{"id":"https://openalex.org/G872165318","display_name":null,"funder_award_id":"ZD186","funder_id":"https://openalex.org/F4320335869","funder_display_name":"National Social Science Fund of China"}],"funders":[{"id":"https://openalex.org/F4320321001","display_name":"National Natural Science Foundation of China","ror":"https://ror.org/01h0zpd94"},{"id":"https://openalex.org/F4320326153","display_name":"Xinjiang Production and Construction Corps","ror":"https://ror.org/03hcmxw73"},{"id":"https://openalex.org/F4320329860","display_name":"National Science and Technology Major Project","ror":null},{"id":"https://openalex.org/F4320335869","display_name":"National Social Science Fund of China","ror":null}],"has_content":{"pdf":true,"grobid_xml":true},"content_urls":{"pdf":"https://content.openalex.org/works/W4415248342.pdf","grobid_xml":"https://content.openalex.org/works/W4415248342.grobid-xml"},"referenced_works_count":0,"referenced_works":[],"related_works":["https://openalex.org/W7153671113","https://openalex.org/W4417409151","https://openalex.org/W7161255858","https://openalex.org/W7128353775","https://openalex.org/W7223139849","https://openalex.org/W7160898458","https://openalex.org/W4416035414","https://openalex.org/W7162538093","https://openalex.org/W7164891892","https://openalex.org/W7211920064"],"abstract_inverted_index":{"In":[0],"fine-grained":[1,47,125],"sparse":[2],"Mixture-of-Experts":[3],"(MoE)":[4],"models,":[5],"a":[6,13,34,39,105,153,168],"large":[7],"pool":[8],"of":[9,61],"specialized":[10],"experts":[11],"replaces":[12],"small":[14],"homogeneous":[15],"set,":[16],"shifting":[17],"performance":[18,138],"and":[19,66,94,115,159],"throughput":[20,84,134],"to":[21],"be":[22],"governed":[23],"by":[24,73],"inference-time":[25],"expert":[26,165],"activation.Yet":[27],"most":[28],"existing":[29],"optimization":[30],"recipes":[31],"implicitly":[32],"assume":[33],"fixed":[35,64,181],"activation":[36,65,111,158,178],"budget":[37],"(e.g.,":[38],"constant":[40],"Top-k":[41,70],"per":[42],"layer),":[43],"whose":[44],"behavior":[45],"in":[46],"MoEs":[48,126],"is":[49],"poorly":[50],"understood.We":[51],"first":[52],"characterize":[53],"runtime":[54],"skipping":[55,79,130,166],"strategies,":[56],"quantifying":[57],"the":[58,120,124,148],"accuracy-efficiency":[59],"trade-off":[60],"(i)":[62],"uniform":[63],"(ii)":[67],"static":[68,78,88,129],"layer-wise":[69],"allocation":[71],"found":[72],"search.Our":[74],"analysis":[75],"reveals":[76],"that":[77,108],"can":[80],"already":[81],"provide":[82],"substantial":[83],"gains,":[85],"but":[86],"optimal":[87],"schedules":[89],"vary":[90],"significantly":[91],"across":[92],"models":[93],"routing":[95],"mechanisms.We":[96],"therefore":[97],"introduce":[98],"Adaptive":[99],"Skipping":[100],"with":[101,136,176],"Entropy-Penalized":[102],"Thresholding":[103],"(ASET),":[104],"training-free":[106],"policy":[107],"adapts":[109],"token-level":[110],"using":[112],"router":[113],"confidence":[114],"entropy":[116],"while":[117],"remaining":[118],"within":[119],"model's":[121],"original":[122],"budget.Across":[123],"we":[127],"study,":[128],"policies":[131],"yield":[132],"10-78%":[133],"gains":[135],"minimal":[137],"degradation,":[139],"including":[140],"\u226510%":[141],"improvement":[142],"on":[143],"DeepSeek-V3":[144],"without":[145],"measurable":[146],"loss.On":[147],"OLMoE":[149],"testbed,":[150],"ASET":[151],"yields":[152],"Pareto":[154],"frontier":[155],"between":[156],"average":[157],"task":[160],"quality.Overall,":[161],"these":[162],"results":[163],"identify":[164],"as":[167],"practical":[169],"lever":[170],"for":[171],"faster":[172],"finegrained":[173],"MoE":[174],"inference,":[175],"adaptive":[177],"helping":[179],"when":[180],"budgets":[182],"are":[183],"too":[184],"rigid.":[185]},"counts_by_year":[],"updated_date":"2026-10-29T09:56:32.521200","created_date":"2025-10-16T00:00:00"}