{"id":"https://openalex.org/W7212002995","doi":"https://doi.org/10.18653/2026.emnlp-main.1463","title":"Beyond Retraining-Free MoE Compression: A Cost-Normalized Study of Post-Compression Adjustment","display_name":"Beyond Retraining-Free MoE Compression: A Cost-Normalized Study of Post-Compression Adjustment","publication_year":2026,"publication_date":"2026-01-01","ids":{"openalex":"https://openalex.org/W7212002995","doi":"https://doi.org/10.18653/2026.emnlp-main.1463"},"language":"en","primary_location":{"id":"doi:10.18653/2026.emnlp-main.1463","doi":"https://doi.org/10.18653/2026.emnlp-main.1463","is_oa":true,"landing_page_url":"https://doi.org/10.18653/2026.emnlp-main.1463","pdf_url":"https://aclanthology.org/2026.emnlp-main.1463.pdf","source":{"id":"https://openalex.org/S4363608991","display_name":"Conference on Empirical Methods in Natural Language Processing (EMNLP)","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":null,"license_id":null,"version":"publishedVersion","is_accepted":true,"is_published":true,"raw_source_name":"Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing","raw_type":"proceedings-article"},"volume":null,"type":"conference-paper","indexed_in":["arxiv","crossref","datacite"],"open_access":{"is_oa":true,"oa_status":"bronze","oa_url":"https://aclanthology.org/2026.emnlp-main.1463.pdf","any_repository_has_fulltext":true},"authorships":[{"author_position":"first","author":{"id":"https://openalex.org/A5128086975","display_name":"Sieun Hyeon","orcid":null,"observed_orcids":[]},"institutions":[{"id":"https://openalex.org/I139264467","display_name":"Seoul National University","ror":"https://ror.org/04h9pn542","country_code":"KR","type":"education","lineage":["https://openalex.org/I139264467"]}],"countries":["KR"],"is_corresponding":false,"raw_author_name":"Sieun Hyeon","raw_affiliation_strings":["Department of Electrical and Computer Engineering , Seoul National University ,"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"Department of Electrical and Computer Engineering , Seoul National University ,","institution_ids":["https://openalex.org/I139264467"]}]},{"author_position":"last","author":{"id":"https://openalex.org/A5151013663","display_name":"Jaeyoung Do","orcid":null,"observed_orcids":[]},"institutions":[{"id":"https://openalex.org/I139264467","display_name":"Seoul National University","ror":"https://ror.org/04h9pn542","country_code":"KR","type":"education","lineage":["https://openalex.org/I139264467"]}],"countries":["KR"],"is_corresponding":false,"raw_author_name":"Jaeyoung Do","raw_affiliation_strings":["Department of Electrical and Computer Engineering , Seoul National University ,","Interdisciplinary Program in Artificial Intelligence , Seoul National University"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"Department of Electrical and Computer Engineering , Seoul National University ,","institution_ids":["https://openalex.org/I139264467"]},{"raw_affiliation_string":"Interdisciplinary Program in Artificial Intelligence , Seoul National University","institution_ids":["https://openalex.org/I139264467"]}]}],"institutions":[],"countries_distinct_count":1,"institutions_distinct_count":1,"corresponding_author_ids":[],"corresponding_institution_ids":["https://openalex.org/I139264467"],"apc_list":null,"apc_paid":null,"fwci":0.0,"has_fulltext":true,"cited_by_count":0,"citation_normalized_percentile":{"value":0.67722127,"is_in_top_1_percent":false,"is_in_top_10_percent":false},"cited_by_percentile_year":null,"biblio":{"volume":null,"issue":null,"first_page":"33444","last_page":"33479"},"is_retracted":false,"is_paratext":false,"is_xpac":false,"primary_topic":{"id":"https://openalex.org/T10036","display_name":"Advanced Neural Network Applications","score":0.9394999742507935,"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":0.9394999742507935,"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/T11307","display_name":"Domain Adaptation and Few-Shot Learning","score":0.027300000190734863,"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":0.017000000923871994,"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":0.7319999933242798},{"id":"https://openalex.org/keywords/expert-pruning","display_name":"expert pruning","score":0.7850000262260437},{"id":"https://openalex.org/keywords/knowledge-distillation","display_name":"knowledge distillation","score":0.9739999771118164},{"id":"https://openalex.org/keywords/large-language-models","display_name":"large language models","score":0.5929999947547913},{"id":"https://openalex.org/keywords/performance-recovery","display_name":"performance recovery","score":0.37599998712539673}],"concepts":[{"id":"https://openalex.org/C180016635","wikidata":"https://www.wikidata.org/wiki/Q2712821","display_name":"Compression (physics)","level":2,"score":0.7211999893188477},{"id":"https://openalex.org/C41008148","wikidata":"https://www.wikidata.org/wiki/Q21198","display_name":"Computer science","level":0,"score":0.44620001316070557},{"id":"https://openalex.org/C108010975","wikidata":"https://www.wikidata.org/wiki/Q500094","display_name":"Pruning","level":2,"score":0.3790999948978424},{"id":"https://openalex.org/C78548338","wikidata":"https://www.wikidata.org/wiki/Q2493","display_name":"Data compression","level":2,"score":0.3772999942302704},{"id":"https://openalex.org/C149782125","wikidata":"https://www.wikidata.org/wiki/Q160039","display_name":"Econometrics","level":1,"score":0.3111000061035156},{"id":"https://openalex.org/C33923547","wikidata":"https://www.wikidata.org/wiki/Q395","display_name":"Mathematics","level":0,"score":0.3003999888896942},{"id":"https://openalex.org/C25797200","wikidata":"https://www.wikidata.org/wiki/Q828137","display_name":"Compression ratio","level":3,"score":0.28859999775886536},{"id":"https://openalex.org/C94375191","wikidata":"https://www.wikidata.org/wiki/Q11205","display_name":"Arithmetic","level":1,"score":0.27149999141693115},{"id":"https://openalex.org/C105795698","wikidata":"https://www.wikidata.org/wiki/Q12483","display_name":"Statistics","level":1,"score":0.26409998536109924},{"id":"https://openalex.org/C200601418","wikidata":"https://www.wikidata.org/wiki/Q2193887","display_name":"Reliability engineering","level":1,"score":0.2515999972820282}],"mesh":[],"locations_count":3,"locations":[{"id":"doi:10.18653/2026.emnlp-main.1463","doi":"https://doi.org/10.18653/2026.emnlp-main.1463","is_oa":true,"landing_page_url":"https://doi.org/10.18653/2026.emnlp-main.1463","pdf_url":"https://aclanthology.org/2026.emnlp-main.1463.pdf","source":{"id":"https://openalex.org/S4363608991","display_name":"Conference on Empirical Methods in Natural Language Processing (EMNLP)","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":null,"license_id":null,"version":"publishedVersion","is_accepted":true,"is_published":true,"raw_source_name":"Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing","raw_type":"proceedings-article"},{"id":"pmh:oai:arXiv.org:2609.06076","doi":null,"is_oa":true,"landing_page_url":"https://arxiv.org/abs/2609.06076","pdf_url":"https://arxiv.org/pdf/2609.06076","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":null,"license_id":null,"version":"submittedVersion","is_accepted":false,"is_published":false,"raw_source_name":null,"raw_type":"text"},{"id":"doi:10.48550/arxiv.2609.06076","doi":"https://doi.org/10.48550/arxiv.2609.06076","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2609.06076","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":null,"license_id":null,"version":null,"is_accepted":false,"is_published":false,"raw_source_name":null,"raw_type":"Preprint"}],"best_oa_location":{"id":"doi:10.18653/2026.emnlp-main.1463","doi":"https://doi.org/10.18653/2026.emnlp-main.1463","is_oa":true,"landing_page_url":"https://doi.org/10.18653/2026.emnlp-main.1463","pdf_url":"https://aclanthology.org/2026.emnlp-main.1463.pdf","source":{"id":"https://openalex.org/S4363608991","display_name":"Conference on Empirical Methods in Natural Language Processing (EMNLP)","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":null,"license_id":null,"version":"publishedVersion","is_accepted":true,"is_published":true,"raw_source_name":"Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing","raw_type":"proceedings-article"},"sustainable_development_goals":[],"study_designs":[],"awards":[{"id":"https://openalex.org/G2947324332","display_name":null,"funder_award_id":"RS-2021II211343","funder_id":"https://openalex.org/F4320335489","funder_display_name":"Institute for Information and Communications Technology Promotion"},{"id":"https://openalex.org/G4475359334","display_name":null,"funder_award_id":"RS2021-II211343","funder_id":"https://openalex.org/F4320321292","funder_display_name":"Seoul National University"},{"id":"https://openalex.org/G6120464043","display_name":null,"funder_award_id":"RS2021-II211343","funder_id":"https://openalex.org/F4320328359","funder_display_name":"Ministry of Science and ICT, South Korea"},{"id":"https://openalex.org/G8978267545","display_name":null,"funder_award_id":"RS-2021-II211343","funder_id":"https://openalex.org/F4320322120","funder_display_name":"National Research Foundation of Korea"}],"funders":[{"id":"https://openalex.org/F4320320671","display_name":"National Research Foundation","ror":"https://ror.org/05s0g1g46"},{"id":"https://openalex.org/F4320321292","display_name":"Seoul National University","ror":"https://ror.org/04h9pn542"},{"id":"https://openalex.org/F4320322120","display_name":"National Research Foundation of Korea","ror":"https://ror.org/013aysd81"},{"id":"https://openalex.org/F4320328359","display_name":"Ministry of Science and ICT, South Korea","ror":"https://ror.org/01wpjm123"},{"id":"https://openalex.org/F4320332195","display_name":"Samsung","ror":"https://ror.org/04w3jy968"},{"id":"https://openalex.org/F4320335489","display_name":"Institute for Information and Communications Technology Promotion","ror":"https://ror.org/01g0hqq23"}],"has_content":{"grobid_xml":true,"pdf":true},"content_urls":{"pdf":"https://content.openalex.org/works/W7212002995.pdf","grobid_xml":"https://content.openalex.org/works/W7212002995.grobid-xml"},"referenced_works_count":0,"referenced_works":[],"related_works":["https://openalex.org/W7133503800","https://openalex.org/W4404792840","https://openalex.org/W6910774473","https://openalex.org/W7221812033","https://openalex.org/W7137282234","https://openalex.org/W7214747416","https://openalex.org/W7160893354","https://openalex.org/W7202174808","https://openalex.org/W4403344761","https://openalex.org/W7128630931"],"abstract_inverted_index":{"Retraining-free":[0],"MoE":[1,28,45],"compression":[2,118],"reduces":[3],"deployment":[4],"memory":[5],"by":[6],"pruning":[7],"or":[8],"merging":[9],"experts,":[10],"but":[11],"often":[12],"treats":[13],"the":[14,18,87,106,111,132],"compressed":[15,27,34],"checkpoint":[16],"as":[17,33],"final":[19],"artifact.We":[20],"argue":[21],"that":[22,36,116],"this":[23],"view":[24],"is":[25,95],"incomplete:":[26],"checkpoints":[29],"are":[30],"better":[31],"understood":[32],"initializations":[35],"benefit":[37],"from":[38],"a":[39,77,128],"tiny":[40],"postcompression":[41,124],"adjustment":[42,104,125],"stage.Across":[43],"two":[44],"LLM":[46],"backbones,":[47],"four":[48],"pruning/merging":[49],"methods,":[50],"three":[51],"expert-retention":[52],"ratios,":[53],"and":[54,61,68,76,102],"28":[55],"benchmarks,":[56],"we":[57],"compare":[58],"LM":[59,93],"fine-tuning":[60,94],"teacher-based":[62],"KD":[63],"under":[64],"matched":[65],"small-data":[66],"budgets":[67],"measured":[69],"GPU":[70],"costs.Using":[71],"only":[72],"3,000":[73],"C4":[74],"examples":[75],"single":[78],"epoch":[79],"of":[80,86,131],"adjustment,":[81],"FULL":[82],"FT":[83],"recovers":[84],"37.3%":[85],"original-to-compressed":[88],"performance":[89,133],"gap":[90],"on":[91],"average.Moreover,":[92],"more":[96],"costeffective":[97],"than":[98],"standard":[99],"token-level":[100],"KD,":[101],"fullparameter":[103],"gives":[105],"strongest":[107],"costrecovery":[108],"trade-off":[109],"among":[110],"tested":[112],"scopes.These":[113],"results":[114],"suggest":[115],"retraining-free":[117],"should":[119],"be":[120],"paired":[121],"with":[122],"small":[123],"to":[126],"recover":[127],"substantial":[129],"portion":[130],"lost":[134],"during":[135],"compression.":[136]},"counts_by_year":[],"updated_date":"2026-11-06T12:05:46.655746","created_date":"2026-09-10T00:00:00"}