{"id":"https://openalex.org/W7166802417","doi":"https://doi.org/10.18653/v1/2026.findings-acl.1944","title":"Improved Policy Optimization for Mixture-of-Experts Models: Importance Sampling and Rewarding from an Expert-Centric Perspective","display_name":"Improved Policy Optimization for Mixture-of-Experts Models: Importance Sampling and Rewarding from an Expert-Centric Perspective","publication_year":2026,"publication_date":"2026-01-01","ids":{"openalex":"https://openalex.org/W7166802417","doi":"https://doi.org/10.18653/v1/2026.findings-acl.1944"},"language":"en","primary_location":{"id":"doi:10.18653/v1/2026.findings-acl.1944","doi":"https://doi.org/10.18653/v1/2026.findings-acl.1944","is_oa":true,"landing_page_url":"https://doi.org/10.18653/v1/2026.findings-acl.1944","pdf_url":"https://aclanthology.org/2026.findings-acl.1944.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":["crossref"],"open_access":{"is_oa":true,"oa_status":"hybrid","oa_url":"https://aclanthology.org/2026.findings-acl.1944.pdf","any_repository_has_fulltext":false},"authorships":[{"author_position":"first","author":{"id":"https://openalex.org/A5139797199","display_name":"Yining Qian","orcid":null,"observed_orcids":[]},"institutions":[{"id":"https://openalex.org/I76569877","display_name":"Southeast University","ror":"https://ror.org/04ct4d772","country_code":"CN","type":"education","lineage":["https://openalex.org/I76569877"]}],"countries":["CN"],"is_corresponding":false,"raw_author_name":"Yining Qian","raw_affiliation_strings":["Southeast University ,"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"Southeast University ,","institution_ids":["https://openalex.org/I76569877"]}]},{"author_position":"middle","author":{"id":"https://openalex.org/A5139797770","display_name":"Jinpeng Li","orcid":null,"observed_orcids":[]},"institutions":[{"id":"https://openalex.org/I2250955327","display_name":"Huawei Technologies (China)","ror":"https://ror.org/00cmhce21","country_code":"CN","type":"company","lineage":["https://openalex.org/I2250955327"]}],"countries":["CN"],"is_corresponding":false,"raw_author_name":"Jinpeng Li","raw_affiliation_strings":["Huawei Technologies"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"Huawei Technologies","institution_ids":["https://openalex.org/I2250955327"]}]},{"author_position":"middle","author":{"id":"https://openalex.org/A5140949489","display_name":"Fei Mi","orcid":null,"observed_orcids":[]},"institutions":[{"id":"https://openalex.org/I2250955327","display_name":"Huawei Technologies (China)","ror":"https://ror.org/00cmhce21","country_code":"CN","type":"company","lineage":["https://openalex.org/I2250955327"]}],"countries":["CN"],"is_corresponding":false,"raw_author_name":"Fei Mi","raw_affiliation_strings":["Huawei Technologies"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"Huawei Technologies","institution_ids":["https://openalex.org/I2250955327"]}]},{"author_position":"middle","author":{"id":"https://openalex.org/A5046228314","display_name":"Lifeng Shang","orcid":null,"observed_orcids":[]},"institutions":[{"id":"https://openalex.org/I2250955327","display_name":"Huawei Technologies (China)","ror":"https://ror.org/00cmhce21","country_code":"CN","type":"company","lineage":["https://openalex.org/I2250955327"]}],"countries":["CN"],"is_corresponding":false,"raw_author_name":"Lifeng Shang","raw_affiliation_strings":["Huawei Technologies"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"Huawei Technologies","institution_ids":["https://openalex.org/I2250955327"]}]},{"author_position":"last","author":{"id":"https://openalex.org/A5139767696","display_name":"Xiang Zhang","orcid":null,"observed_orcids":[]},"institutions":[{"id":"https://openalex.org/I76569877","display_name":"Southeast University","ror":"https://ror.org/04ct4d772","country_code":"CN","type":"education","lineage":["https://openalex.org/I76569877"]}],"countries":["CN"],"is_corresponding":false,"raw_author_name":"Xiang Zhang","raw_affiliation_strings":["Southeast University ,"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"Southeast University ,","institution_ids":["https://openalex.org/I76569877"]}]}],"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":0.43730703,"is_in_top_1_percent":false,"is_in_top_10_percent":false},"cited_by_percentile_year":null,"biblio":{"volume":null,"issue":null,"first_page":"39029","last_page":"39038"},"is_retracted":false,"is_paratext":false,"is_xpac":false,"primary_topic":{"id":"https://openalex.org/T10462","display_name":"Reinforcement Learning in Robotics","score":0.888700008392334,"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/T10462","display_name":"Reinforcement Learning in Robotics","score":0.888700008392334,"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/T10028","display_name":"Topic Modeling","score":0.024800000712275505,"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/T11307","display_name":"Domain Adaptation and Few-Shot Learning","score":0.021199999377131462,"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":0.9459999799728394},{"id":"https://openalex.org/keywords/policy-optimization","display_name":"policy optimization","score":0.7310000061988831},{"id":"https://openalex.org/keywords/importance-sampling","display_name":"importance sampling","score":0.8450000286102295},{"id":"https://openalex.org/keywords/training-instability","display_name":"training instability","score":0.6570000052452087},{"id":"https://openalex.org/keywords/large-language-models","display_name":"large language models","score":0.875},{"id":"https://openalex.org/keywords/reasoning-tasks","display_name":"reasoning tasks","score":0.49399998784065247}],"concepts":[{"id":"https://openalex.org/C12713177","wikidata":"https://www.wikidata.org/wiki/Q1900281","display_name":"Perspective (graphical)","level":2,"score":0.696399986743927},{"id":"https://openalex.org/C140779682","wikidata":"https://www.wikidata.org/wiki/Q210868","display_name":"Sampling (signal processing)","level":3,"score":0.5174000263214111},{"id":"https://openalex.org/C41008148","wikidata":"https://www.wikidata.org/wiki/Q21198","display_name":"Computer science","level":0,"score":0.46650001406669617},{"id":"https://openalex.org/C112930515","wikidata":"https://www.wikidata.org/wiki/Q4389547","display_name":"Risk analysis (engineering)","level":1,"score":0.3287999927997589},{"id":"https://openalex.org/C162324750","wikidata":"https://www.wikidata.org/wiki/Q8134","display_name":"Economics","level":0,"score":0.32179999351501465},{"id":"https://openalex.org/C26517878","wikidata":"https://www.wikidata.org/wiki/Q228039","display_name":"Key (lock)","level":2,"score":0.31619998812675476},{"id":"https://openalex.org/C539667460","wikidata":"https://www.wikidata.org/wiki/Q2414942","display_name":"Management science","level":1,"score":0.2948000133037567},{"id":"https://openalex.org/C96250715","wikidata":"https://www.wikidata.org/wiki/Q965330","display_name":"Estimation","level":2,"score":0.2842000126838684},{"id":"https://openalex.org/C42475967","wikidata":"https://www.wikidata.org/wiki/Q194292","display_name":"Operations research","level":1,"score":0.2736000120639801},{"id":"https://openalex.org/C109986646","wikidata":"https://www.wikidata.org/wiki/Q546113","display_name":"Public policy","level":2,"score":0.26899999380111694},{"id":"https://openalex.org/C2778137410","wikidata":"https://www.wikidata.org/wiki/Q2732820","display_name":"Government (linguistics)","level":2,"score":0.26339998841285706},{"id":"https://openalex.org/C162118730","wikidata":"https://www.wikidata.org/wiki/Q1128453","display_name":"Actuarial science","level":1,"score":0.26170000433921814},{"id":"https://openalex.org/C144133560","wikidata":"https://www.wikidata.org/wiki/Q4830453","display_name":"Business","level":0,"score":0.26109999418258667},{"id":"https://openalex.org/C149782125","wikidata":"https://www.wikidata.org/wiki/Q160039","display_name":"Econometrics","level":1,"score":0.2581999897956848}],"mesh":[],"locations_count":1,"locations":[{"id":"doi:10.18653/v1/2026.findings-acl.1944","doi":"https://doi.org/10.18653/v1/2026.findings-acl.1944","is_oa":true,"landing_page_url":"https://doi.org/10.18653/v1/2026.findings-acl.1944","pdf_url":"https://aclanthology.org/2026.findings-acl.1944.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"}],"best_oa_location":{"id":"doi:10.18653/v1/2026.findings-acl.1944","doi":"https://doi.org/10.18653/v1/2026.findings-acl.1944","is_oa":true,"landing_page_url":"https://doi.org/10.18653/v1/2026.findings-acl.1944","pdf_url":"https://aclanthology.org/2026.findings-acl.1944.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":[],"funders":[],"has_content":{"grobid_xml":true,"pdf":true},"content_urls":{"pdf":"https://content.openalex.org/works/W7166802417.pdf","grobid_xml":"https://content.openalex.org/works/W7166802417.grobid-xml"},"referenced_works_count":0,"referenced_works":[],"related_works":["https://openalex.org/W7101804538","https://openalex.org/W7212687009","https://openalex.org/W7222964004","https://openalex.org/W4399353875","https://openalex.org/W7117685307","https://openalex.org/W7130685168","https://openalex.org/W4402683890","https://openalex.org/W4416033980","https://openalex.org/W4416035879","https://openalex.org/W4416054566"],"abstract_inverted_index":{"Reinforcement":[0],"learning":[1],"(RL)":[2],"has":[3],"demonstrated":[4],"considerable":[5],"promise":[6],"in":[7,29],"enhancing":[8],"large":[9],"language":[10],"models.However,":[11],"its":[12],"application":[13],"to":[14,77,103,157],"Mixture-of-Experts":[15],"(MoE)":[16],"architectures":[17],"is":[18],"frequently":[19],"hindered":[20],"by":[21,133],"training":[22],"instability,":[23],"primarily":[24],"stemming":[25],"from":[26],"token-level":[27,87],"misalignment":[28],"expert":[30],"assignments":[31],"between":[32,40,124],"current":[33],"and":[34,50,127],"behavior":[35],"policies.Existing":[36],"approaches":[37],"often":[38],"oscillate":[39],"overly":[41],"coarse":[42],"sequence-level":[43,96],"importance":[44,72,88,135],"sampling,":[45],"which":[46,69,109],"ignores":[47],"token-specific":[48],"discrepancies,":[49],"restrictive":[51],"expert-selection":[52],"constraints":[53],"that":[54,140],"suppress":[55],"beneficial":[56],"policy":[57,128],"exploration.To":[58],"bridge":[59],"this":[60,100],"gap,":[61],"we":[62],"propose":[63],"Expert":[64],"Relative":[65],"Policy":[66],"Optimization":[67],"(ERPO),":[68],"introduces":[70],"expert-level":[71],"sampling.By":[73],"grouping":[74],"tokens":[75],"according":[76],"their":[78],"routing":[79,112],"assignments,":[80],"ERPO":[81,98,141],"mitigates":[82],"the":[83,92,151,158],"high":[84],"variance":[85],"of":[86,95,153,162],"sampling":[89,136],"while":[90],"overcoming":[91],"token-agnostic":[93],"limitations":[94],"methods.Furthermore,":[97],"leverages":[99],"expert-centric":[101],"granularity":[102],"introduce":[104],"an":[105],"Expert-Selection":[106],"Entropy":[107],"Reward,":[108],"dynamically":[110],"adjusts":[111],"uncertainty":[113],"based":[114],"on":[115],"task-specific":[116],"feedback.This":[117],"unique":[118],"mechanism":[119],"ensures":[120],"a":[121],"rigorous":[122],"alignment":[123],"reward":[125],"signals":[126],"updates-a":[129],"capability":[130],"inherently":[131],"unattainable":[132],"traditional":[134],"methods.Experimental":[137],"results":[138],"demonstrate":[139],"significantly":[142],"outperforms":[143],"strong":[144],"baselines":[145],"across":[146],"multiple":[147],"reasoning":[148],"tasks,":[149],"highlighting":[150],"efficacy":[152],"tailoring":[154],"RL":[155],"objectives":[156],"structural":[159],"inductive":[160],"biases":[161],"MoE":[163],"models.":[164]},"counts_by_year":[],"updated_date":"2026-10-28T10:33:24.094632","created_date":"2026-07-02T00:00:00"}