{"id":"https://openalex.org/W7163715208","doi":"https://doi.org/10.48550/arxiv.2606.05538","title":"Less is MoE: Trimming Experts in Domain-Specialist Language Models","display_name":"Less is MoE: Trimming Experts in Domain-Specialist Language Models","publication_year":2026,"publication_date":"2026-06-04","ids":{"openalex":"https://openalex.org/W7163715208","doi":"https://doi.org/10.48550/arxiv.2606.05538"},"language":null,"primary_location":{"id":"doi:10.48550/arxiv.2606.05538","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2606.05538","pdf_url":null,"source":{"id":"https://openalex.org/S4306400194","display_name":"arXiv (Cornell University)","issn_l":null,"issn":null,"is_oa":true,"is_in_doaj":false,"is_core":false,"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"},"type":"preprint","indexed_in":["datacite"],"open_access":{"is_oa":true,"oa_status":"green","oa_url":"https://doi.org/10.48550/arxiv.2606.05538","any_repository_has_fulltext":true},"authorships":[{"author_position":"first","author":{"id":"https://openalex.org/A5138001803","display_name":"Haoze He","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"He, Haoze","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5133625368","display_name":"Xinkai Zou","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Zou, Xinkai","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5137934080","display_name":"Xuan Jiang","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Jiang, Xuan","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5066294648","display_name":"\u4e01\u674f\u6c85","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Ding, Xingyuan","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5137961754","display_name":"Ao Qu","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Qu, Ao","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5111213544","display_name":"Juncheng Billy Li","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Li, Juncheng Billy","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"last","author":{"id":"https://openalex.org/A5137966363","display_name":"Heather Miller","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Miller, Heather","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":false,"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/T10028","display_name":"Topic Modeling","score":0.2572999894618988,"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":0.2572999894618988,"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":0.19269999861717224,"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/T11704","display_name":"Mobile Crowdsensing and Crowdsourcing","score":0.08079999685287476,"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"}}],"keywords":[{"id":"https://openalex.org/keywords/trimming","display_name":"Trimming","score":0.6419000029563904},{"id":"https://openalex.org/keywords/granularity","display_name":"Granularity","score":0.6413999795913696},{"id":"https://openalex.org/keywords/dimension","display_name":"Dimension (graph theory)","score":0.5483999848365784},{"id":"https://openalex.org/keywords/inference","display_name":"Inference","score":0.51910001039505},{"id":"https://openalex.org/keywords/trace","display_name":"TRACE (psycholinguistics)","score":0.5174999833106995},{"id":"https://openalex.org/keywords/rank","display_name":"Rank (graph theory)","score":0.5156999826431274},{"id":"https://openalex.org/keywords/compression","display_name":"Compression (physics)","score":0.49320000410079956},{"id":"https://openalex.org/keywords/ranking","display_name":"Ranking (information retrieval)","score":0.4675999879837036},{"id":"https://openalex.org/keywords/key","display_name":"Key (lock)","score":0.3982999920845032}],"concepts":[{"id":"https://openalex.org/C41008148","wikidata":"https://www.wikidata.org/wiki/Q21198","display_name":"Computer science","level":0,"score":0.6944000124931335},{"id":"https://openalex.org/C56951928","wikidata":"https://www.wikidata.org/wiki/Q3539213","display_name":"Trimming","level":2,"score":0.6419000029563904},{"id":"https://openalex.org/C177774035","wikidata":"https://www.wikidata.org/wiki/Q1246948","display_name":"Granularity","level":2,"score":0.6413999795913696},{"id":"https://openalex.org/C33676613","wikidata":"https://www.wikidata.org/wiki/Q13415176","display_name":"Dimension (graph theory)","level":2,"score":0.5483999848365784},{"id":"https://openalex.org/C2776214188","wikidata":"https://www.wikidata.org/wiki/Q408386","display_name":"Inference","level":2,"score":0.51910001039505},{"id":"https://openalex.org/C75291252","wikidata":"https://www.wikidata.org/wiki/Q1315756","display_name":"TRACE (psycholinguistics)","level":2,"score":0.5174999833106995},{"id":"https://openalex.org/C164226766","wikidata":"https://www.wikidata.org/wiki/Q7293202","display_name":"Rank (graph theory)","level":2,"score":0.5156999826431274},{"id":"https://openalex.org/C180016635","wikidata":"https://www.wikidata.org/wiki/Q2712821","display_name":"Compression (physics)","level":2,"score":0.49320000410079956},{"id":"https://openalex.org/C189430467","wikidata":"https://www.wikidata.org/wiki/Q7293293","display_name":"Ranking (information retrieval)","level":2,"score":0.4675999879837036},{"id":"https://openalex.org/C154945302","wikidata":"https://www.wikidata.org/wiki/Q11660","display_name":"Artificial intelligence","level":1,"score":0.44530001282691956},{"id":"https://openalex.org/C124101348","wikidata":"https://www.wikidata.org/wiki/Q172491","display_name":"Data mining","level":1,"score":0.4171000123023987},{"id":"https://openalex.org/C26517878","wikidata":"https://www.wikidata.org/wiki/Q228039","display_name":"Key (lock)","level":2,"score":0.3982999920845032},{"id":"https://openalex.org/C132943942","wikidata":"https://www.wikidata.org/wiki/Q2562511","display_name":"Footprint","level":2,"score":0.3944000005722046},{"id":"https://openalex.org/C86037889","wikidata":"https://www.wikidata.org/wiki/Q4330127","display_name":"Learning to rank","level":3,"score":0.3849000036716461},{"id":"https://openalex.org/C105339364","wikidata":"https://www.wikidata.org/wiki/Q2297740","display_name":"Software deployment","level":2,"score":0.38269999623298645},{"id":"https://openalex.org/C119857082","wikidata":"https://www.wikidata.org/wiki/Q2539","display_name":"Machine learning","level":1,"score":0.37700000405311584},{"id":"https://openalex.org/C78548338","wikidata":"https://www.wikidata.org/wiki/Q2493","display_name":"Data compression","level":2,"score":0.37549999356269836},{"id":"https://openalex.org/C74912251","wikidata":"https://www.wikidata.org/wiki/Q6815727","display_name":"Memory footprint","level":2,"score":0.36899998784065247},{"id":"https://openalex.org/C137293760","wikidata":"https://www.wikidata.org/wiki/Q3621696","display_name":"Language model","level":2,"score":0.3671000003814697},{"id":"https://openalex.org/C157764524","wikidata":"https://www.wikidata.org/wiki/Q1383412","display_name":"Throughput","level":3,"score":0.3499999940395355},{"id":"https://openalex.org/C125411270","wikidata":"https://www.wikidata.org/wiki/Q18653","display_name":"Encoding (memory)","level":2,"score":0.29919999837875366},{"id":"https://openalex.org/C2777402240","wikidata":"https://www.wikidata.org/wiki/Q6783436","display_name":"Masking (illustration)","level":2,"score":0.29829999804496765},{"id":"https://openalex.org/C79581498","wikidata":"https://www.wikidata.org/wiki/Q1367530","display_name":"Suite","level":2,"score":0.29109999537467957},{"id":"https://openalex.org/C81669768","wikidata":"https://www.wikidata.org/wiki/Q2359161","display_name":"Precision and recall","level":2,"score":0.27619999647140503},{"id":"https://openalex.org/C25797200","wikidata":"https://www.wikidata.org/wiki/Q828137","display_name":"Compression ratio","level":3,"score":0.26809999346733093},{"id":"https://openalex.org/C184898388","wikidata":"https://www.wikidata.org/wiki/Q1435712","display_name":"Pairwise comparison","level":2,"score":0.2563999891281128},{"id":"https://openalex.org/C2779960059","wikidata":"https://www.wikidata.org/wiki/Q7113681","display_name":"Overhead (engineering)","level":2,"score":0.25290000438690186},{"id":"https://openalex.org/C165838908","wikidata":"https://www.wikidata.org/wiki/Q736777","display_name":"Calibration","level":2,"score":0.2522999942302704}],"mesh":[],"locations_count":1,"locations":[{"id":"doi:10.48550/arxiv.2606.05538","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2606.05538","pdf_url":null,"source":{"id":"https://openalex.org/S4306400194","display_name":"arXiv (Cornell University)","issn_l":null,"issn":null,"is_oa":true,"is_in_doaj":false,"is_core":false,"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":null,"raw_source_name":null,"raw_type":"Preprint"}],"best_oa_location":{"id":"doi:10.48550/arxiv.2606.05538","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2606.05538","pdf_url":null,"source":{"id":"https://openalex.org/S4306400194","display_name":"arXiv (Cornell University)","issn_l":null,"issn":null,"is_oa":true,"is_in_doaj":false,"is_core":false,"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"},"sustainable_development_goals":[],"awards":[],"funders":[],"has_content":{"grobid_xml":false,"pdf":false},"content_urls":null,"referenced_works_count":0,"referenced_works":[],"related_works":[],"abstract_inverted_index":{"Mixture-of-Experts":[0],"(MoE)":[1],"models":[2],"achieve":[3],"strong":[4],"performance":[5],"through":[6],"conditional":[7],"computation,":[8],"but":[9,46],"their":[10],"large":[11],"parameter":[12],"footprint":[13],"poses":[14],"deployment":[15],"challenges.":[16],"Prior":[17],"MoE":[18,117,155],"compression":[19,118,148],"approaches":[20],"catastrophically":[21],"fail":[22],"when":[23],"evaluated":[24],"on":[25,96],"general-purpose":[26],"benchmarks":[27],"beyond":[28],"commonsense":[29],"reasoning.":[30],"We":[31],"trace":[32],"this":[33],"failure":[34],"to":[35,105],"the":[36,114],"granularity":[37,141],"of":[38,72,82],"compression:":[39],"important":[40],"capabilities":[41],"are":[42],"distributed":[43],"across":[44],"experts":[45],"concentrated":[47],"in":[48,75,154],"FFN":[49,104],"sparse":[50],"intermediate":[51,85,107,139],"dimensions.":[52],"To":[53],"identify":[54],"these":[55],"dimensions,":[56],"we":[57,98],"use":[58],"Fisher":[59,111],"importance":[60],"which":[61,101],"outperforms":[62],"activation-,":[63],"router-score-,":[64],"and":[65,68,130,149],"magnitude-based":[66],"alternatives,":[67],"identifies":[69],"tiny":[70],"sets":[71],"task-critical":[73],"dimensions:":[74],"Qwen1.5-MoE,":[76],"removing":[77],"as":[78,80],"few":[79],"12":[81],"1.35M":[83],"routed-FFN":[84],"dimensions":[86,108],"collapses":[87],"GSM8K":[88],"accuracy":[89],"while":[90,124],"largely":[91],"preserving":[92],"factual-knowledge":[93],"performance.":[94],"Building":[95],"this,":[97],"propose":[99],"Fisher-MoE,":[100],"operates":[102],"within":[103],"remove":[106],"ranked":[109],"by":[110,128,134],"importance.":[112],"At":[113],"same":[115],"50%":[116],"ratio,":[119],"Fisher-MoE":[120],"preserves":[121],"model":[122],"capability,":[123],"reducing":[125],"weight":[126],"memory":[127],"~45%":[129],"improving":[131],"inference":[132],"throughput":[133],"21%.":[135],"These":[136],"findings":[137],"suggest":[138],"dimension":[140],"is":[142],"an":[143],"effective":[144],"unit":[145],"for":[146],"both":[147],"ranking":[150],"where":[151],"capability":[152],"concentrates":[153],"models.":[156]},"counts_by_year":[],"updated_date":"2026-07-01T06:00:48.157686","created_date":"2026-06-06T00:00:00"}