{"id":"https://openalex.org/W7166831075","doi":"https://doi.org/10.18653/v1/2026.findings-acl.857","title":"From Pseudo-Balancing to True Specialization: Memory-Aware Routing for Mixture-of-Experts","display_name":"From Pseudo-Balancing to True Specialization: Memory-Aware Routing for Mixture-of-Experts","publication_year":2026,"publication_date":"2026-01-01","ids":{"openalex":"https://openalex.org/W7166831075","doi":"https://doi.org/10.18653/v1/2026.findings-acl.857"},"language":null,"primary_location":{"id":"doi:10.18653/v1/2026.findings-acl.857","is_oa":true,"landing_page_url":"https://doi.org/10.18653/v1/2026.findings-acl.857","pdf_url":"https://aclanthology.org/2026.findings-acl.857.pdf","source":null,"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"},"type":"conference-paper","indexed_in":["crossref"],"open_access":{"is_oa":true,"oa_status":"gold","oa_url":"https://aclanthology.org/2026.findings-acl.857.pdf","any_repository_has_fulltext":null},"authorships":[{"author_position":"first","author":{"id":"https://openalex.org/A5139783894","display_name":"Peixuan Hou","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Peixuan Hou","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5100525788","display_name":"Yunbo Hou","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Yunbo Hou","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5139767519","display_name":"Bin Chen","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Bin Chen","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5139812318","display_name":"LI He","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"LI He","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5139726016","display_name":"Jian Xu","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Jian Xu","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5139750343","display_name":"Weiping Li","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Weiping Li","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5139732377","display_name":"Bo Zheng","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Bo Zheng","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"last","author":{"id":"https://openalex.org/A5139838555","display_name":"Guojie Song","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Guojie Song","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":0.0,"has_fulltext":true,"cited_by_count":0,"citation_normalized_percentile":{"value":0.8543883,"is_in_top_1_percent":false,"is_in_top_10_percent":false},"cited_by_percentile_year":null,"biblio":{"volume":null,"issue":null,"first_page":"17320","last_page":"17337"},"is_retracted":false,"is_paratext":false,"is_xpac":false,"primary_topic":{"id":"https://openalex.org/T10064","display_name":"Complex Network Analysis Techniques","score":0.08780000358819962,"subfield":{"id":"https://openalex.org/subfields/3109","display_name":"Statistical and Nonlinear Physics"},"field":{"id":"https://openalex.org/fields/31","display_name":"Physics and Astronomy"},"domain":{"id":"https://openalex.org/domains/3","display_name":"Physical Sciences"}},"topics":[{"id":"https://openalex.org/T10064","display_name":"Complex Network Analysis Techniques","score":0.08780000358819962,"subfield":{"id":"https://openalex.org/subfields/3109","display_name":"Statistical and Nonlinear Physics"},"field":{"id":"https://openalex.org/fields/31","display_name":"Physics and Astronomy"},"domain":{"id":"https://openalex.org/domains/3","display_name":"Physical Sciences"}},{"id":"https://openalex.org/T11704","display_name":"Mobile Crowdsensing and Crowdsourcing","score":0.07090000063180923,"subfield":{"id":"https://openalex.org/subfields/1706","display_name":"Computer Science Applications"},"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":0.06960000097751617,"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/routing","display_name":"Routing (electronic design automation)","score":0.46160000562667847},{"id":"https://openalex.org/keywords/key","display_name":"Key (lock)","score":0.24070000648498535},{"id":"https://openalex.org/keywords/identification","display_name":"Identification (biology)","score":0.24040000140666962},{"id":"https://openalex.org/keywords/work","display_name":"Work (physics)","score":0.23919999599456787},{"id":"https://openalex.org/keywords/matching","display_name":"Matching (statistics)","score":0.23810000717639923}],"concepts":[{"id":"https://openalex.org/C41008148","wikidata":"https://www.wikidata.org/wiki/Q21198","display_name":"Computer science","level":0,"score":0.5936999917030334},{"id":"https://openalex.org/C74172769","wikidata":"https://www.wikidata.org/wiki/Q1446839","display_name":"Routing (electronic design automation)","level":2,"score":0.46160000562667847},{"id":"https://openalex.org/C31258907","wikidata":"https://www.wikidata.org/wiki/Q1301371","display_name":"Computer network","level":1,"score":0.41920000314712524},{"id":"https://openalex.org/C11413529","wikidata":"https://www.wikidata.org/wiki/Q8366","display_name":"Algorithm","level":1,"score":0.2418999969959259},{"id":"https://openalex.org/C26517878","wikidata":"https://www.wikidata.org/wiki/Q228039","display_name":"Key (lock)","level":2,"score":0.24070000648498535},{"id":"https://openalex.org/C116834253","wikidata":"https://www.wikidata.org/wiki/Q2039217","display_name":"Identification (biology)","level":2,"score":0.24040000140666962},{"id":"https://openalex.org/C18762648","wikidata":"https://www.wikidata.org/wiki/Q42213","display_name":"Work (physics)","level":2,"score":0.23919999599456787},{"id":"https://openalex.org/C165064840","wikidata":"https://www.wikidata.org/wiki/Q1321061","display_name":"Matching (statistics)","level":2,"score":0.23810000717639923},{"id":"https://openalex.org/C99498987","wikidata":"https://www.wikidata.org/wiki/Q2210247","display_name":"Noise (video)","level":3,"score":0.23100000619888306},{"id":"https://openalex.org/C177264268","wikidata":"https://www.wikidata.org/wiki/Q1514741","display_name":"Set (abstract data type)","level":2,"score":0.2257000058889389}],"mesh":[],"locations_count":1,"locations":[{"id":"doi:10.18653/v1/2026.findings-acl.857","is_oa":true,"landing_page_url":"https://doi.org/10.18653/v1/2026.findings-acl.857","pdf_url":"https://aclanthology.org/2026.findings-acl.857.pdf","source":null,"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"}],"best_oa_location":{"id":"doi:10.18653/v1/2026.findings-acl.857","is_oa":true,"landing_page_url":"https://doi.org/10.18653/v1/2026.findings-acl.857","pdf_url":"https://aclanthology.org/2026.findings-acl.857.pdf","source":null,"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":[],"awards":[{"id":"https://openalex.org/G7830412908","display_name":null,"funder_award_id":"62276006","funder_id":"https://openalex.org/F4320321001","funder_display_name":"National Natural Science Foundation of China"}],"funders":[{"id":"https://openalex.org/F4320321001","display_name":"National Natural Science Foundation of China","ror":"https://ror.org/01h0zpd94"}],"has_content":{"grobid_xml":true,"pdf":true},"content_urls":{"pdf":"https://content.openalex.org/works/W7166831075.pdf","grobid_xml":"https://content.openalex.org/works/W7166831075.grobid-xml"},"referenced_works_count":0,"referenced_works":[],"related_works":[],"abstract_inverted_index":{"Mixture-of-Experts":[0],"(MoE)":[1],"efficiently":[2],"trains":[3],"large":[4,119],"models":[5,121,149],"by":[6,189,194],"using":[7],"sparse":[8],"activation":[9],"to":[10,28,42,60,75,156,166],"lower":[11],"costs,":[12],"selecting":[13],"a":[14,61,131,143],"few":[15],"experts":[16,49,77],"based":[17],"on":[18],"data":[19],"characteristics.For":[20],"MoE,":[21],"an":[22,38],"unbalanced":[23],"expert":[24,30,66,110,141,180,187],"load":[25,67,176],"will":[26],"lead":[27,59],"inefficient":[29,102],"utilization":[31,47],"and":[32,55,101,108,178,191,199],"routing":[33],"collapse.Existing":[34],"methods":[35],"commonly":[36],"achieve":[37],"expert-centered":[39],"balancing":[40],"strategy":[41],"solve":[43],"it,":[44],"prioritizing":[45],"equal":[46],"of":[48,82,116],"over":[50],"semantic":[51],"alignment":[52],"between":[53],"tokens":[54,161],"experts.However,":[56],"this":[57],"can":[58],"pseudo-balance":[62,171],"phenomenon:":[63],"To":[64],"ensure":[65],"balancing,":[68],"the":[69,83,114,170,206],"same":[70],"input":[71],"is":[72],"randomly":[73],"routed":[74,163],"different":[76],"across":[78],"training":[79],"steps":[80],"instead":[81],"most":[84],"matching":[85],"one.It":[86],"introduces":[87],"two":[88],"critical":[89],"issues:":[90],"(1)":[91],"Severe":[92],"knowledge":[93],"overlap":[94],"among":[95],"experts,":[96,168],"resulting":[97],"in":[98,106],"redundant":[99],"representations":[100],"parameter":[103,197],"utilization.(2)":[104],"Difficulty":[105],"forming":[107],"stabilizing":[109],"specialization.These":[111],"issues":[112],"limit":[113],"scalability":[115],"models,":[117],"especially":[118],"language":[120],"(LLM).To":[122],"address":[123],"these":[124],"limitations,":[125],"we":[126],"introduce":[127],"Memory-Aware":[128],"Routing":[129],"(MAR),":[130],"training-phase":[132],"approach":[133],"that":[134,160,184],"enhances":[135],"existing":[136],"load-balancing":[137],"strategies.By":[138],"equipping":[139],"each":[140],"with":[142,203],"memory":[144],"buffer,":[145],"our":[146],"method":[147],"explicitly":[148],"their":[150],"long-term":[151],"preferences,":[152],"allowing":[153],"historical":[154],"experience":[155],"guide":[157],"routing.This":[158],"ensures":[159],"are":[162],"more":[164],"consistently":[165],"compatible":[167],"mitigating":[169],"problem":[172],"while":[173],"maintaining":[174],"global":[175],"balance":[177],"fostering":[179],"specialization.Experimental":[181],"results":[182],"show":[183],"MAR":[185],"improves":[186],"specialization":[188],"35%":[190],"downstream":[192],"accuracy":[193],"2%-25%,":[195],"doubles":[196],"efficiency,":[198],"matches":[200],"baseline":[201],"performance":[202],"only":[204],"half":[205],"experts.":[207]},"counts_by_year":[],"updated_date":"2026-07-29T14:22:42.915294","created_date":"2026-07-02T00:00:00"}