{"id":"https://openalex.org/W7162562476","doi":"https://doi.org/10.48550/arxiv.2605.26496","title":"Dense2MoE: Pushing the Pareto Frontier of On-Device LLMs via Unified Pruning and Upcycling","display_name":"Dense2MoE: Pushing the Pareto Frontier of On-Device LLMs via Unified Pruning and Upcycling","publication_year":2026,"publication_date":"2026-05-26","ids":{"openalex":"https://openalex.org/W7162562476","doi":"https://doi.org/10.48550/arxiv.2605.26496"},"language":null,"primary_location":{"id":"doi:10.48550/arxiv.2605.26496","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2605.26496","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.2605.26496","any_repository_has_fulltext":true},"authorships":[{"author_position":"first","author":{"id":"https://openalex.org/A5137136928","display_name":"Fengfa Li","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Li, Fengfa","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5137090281","display_name":"Hongjin Ji","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Ji, Hongjin","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5137127676","display_name":"Yifeng Ding","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Ding, Yifeng","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5125706493","display_name":"Lei Ren","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Ren, Lei","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"last","author":{"id":"https://openalex.org/A5137086013","display_name":"Chen Wei","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Wei, Chen","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/T10036","display_name":"Advanced Neural Network Applications","score":0.39969998598098755,"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":0.39969998598098755,"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":0.0723000019788742,"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/T13553","display_name":"Age of Information Optimization","score":0.06360000371932983,"subfield":{"id":"https://openalex.org/subfields/1705","display_name":"Computer Networks and Communications"},"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/inference","display_name":"Inference","score":0.6295999884605408},{"id":"https://openalex.org/keywords/pruning","display_name":"Pruning","score":0.5544999837875366},{"id":"https://openalex.org/keywords/pareto-principle","display_name":"Pareto principle","score":0.5454999804496765},{"id":"https://openalex.org/keywords/redundancy","display_name":"Redundancy (engineering)","score":0.5203999876976013},{"id":"https://openalex.org/keywords/benchmark","display_name":"Benchmark (surveying)","score":0.3614000082015991},{"id":"https://openalex.org/keywords/layer","display_name":"Layer (electronics)","score":0.33889999985694885},{"id":"https://openalex.org/keywords/robustness","display_name":"Robustness (evolution)","score":0.31850001215934753},{"id":"https://openalex.org/keywords/artificial-neural-network","display_name":"Artificial neural network","score":0.3174000084400177}],"concepts":[{"id":"https://openalex.org/C2776214188","wikidata":"https://www.wikidata.org/wiki/Q408386","display_name":"Inference","level":2,"score":0.6295999884605408},{"id":"https://openalex.org/C41008148","wikidata":"https://www.wikidata.org/wiki/Q21198","display_name":"Computer science","level":0,"score":0.5882999897003174},{"id":"https://openalex.org/C108010975","wikidata":"https://www.wikidata.org/wiki/Q500094","display_name":"Pruning","level":2,"score":0.5544999837875366},{"id":"https://openalex.org/C137635306","wikidata":"https://www.wikidata.org/wiki/Q182667","display_name":"Pareto principle","level":2,"score":0.5454999804496765},{"id":"https://openalex.org/C152124472","wikidata":"https://www.wikidata.org/wiki/Q1204361","display_name":"Redundancy (engineering)","level":2,"score":0.5203999876976013},{"id":"https://openalex.org/C119857082","wikidata":"https://www.wikidata.org/wiki/Q2539","display_name":"Machine learning","level":1,"score":0.41359999775886536},{"id":"https://openalex.org/C154945302","wikidata":"https://www.wikidata.org/wiki/Q11660","display_name":"Artificial intelligence","level":1,"score":0.39500001072883606},{"id":"https://openalex.org/C185798385","wikidata":"https://www.wikidata.org/wiki/Q1161707","display_name":"Benchmark (surveying)","level":2,"score":0.3614000082015991},{"id":"https://openalex.org/C2779227376","wikidata":"https://www.wikidata.org/wiki/Q6505497","display_name":"Layer (electronics)","level":2,"score":0.33889999985694885},{"id":"https://openalex.org/C63479239","wikidata":"https://www.wikidata.org/wiki/Q7353546","display_name":"Robustness (evolution)","level":3,"score":0.31850001215934753},{"id":"https://openalex.org/C50644808","wikidata":"https://www.wikidata.org/wiki/Q192776","display_name":"Artificial neural network","level":2,"score":0.3174000084400177},{"id":"https://openalex.org/C50454189","wikidata":"https://www.wikidata.org/wiki/Q2075125","display_name":"Karush\u2013Kuhn\u2013Tucker conditions","level":2,"score":0.3059999942779541},{"id":"https://openalex.org/C188441871","wikidata":"https://www.wikidata.org/wiki/Q7554146","display_name":"Softmax function","level":3,"score":0.30219998955726624},{"id":"https://openalex.org/C186108316","wikidata":"https://www.wikidata.org/wiki/Q352530","display_name":"Adaptive neuro fuzzy inference system","level":4,"score":0.2996000051498413},{"id":"https://openalex.org/C60908668","wikidata":"https://www.wikidata.org/wiki/Q690207","display_name":"Perceptron","level":3,"score":0.29649999737739563},{"id":"https://openalex.org/C48145219","wikidata":"https://www.wikidata.org/wiki/Q1335365","display_name":"Security token","level":2,"score":0.28459998965263367},{"id":"https://openalex.org/C2776904630","wikidata":"https://www.wikidata.org/wiki/Q356336","display_name":"Adept","level":3,"score":0.27869999408721924},{"id":"https://openalex.org/C68781425","wikidata":"https://www.wikidata.org/wiki/Q2052203","display_name":"Multi-objective optimization","level":2,"score":0.2653000056743622},{"id":"https://openalex.org/C193415008","wikidata":"https://www.wikidata.org/wiki/Q639681","display_name":"Network architecture","level":2,"score":0.25519999861717224},{"id":"https://openalex.org/C46686674","wikidata":"https://www.wikidata.org/wiki/Q466303","display_name":"Boosting (machine learning)","level":2,"score":0.25060001015663147}],"mesh":[],"locations_count":1,"locations":[{"id":"doi:10.48550/arxiv.2605.26496","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2605.26496","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.2605.26496","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2605.26496","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":[{"score":0.5237464308738708,"id":"https://metadata.un.org/sdg/9","display_name":"Industry, innovation and infrastructure"}],"awards":[],"funders":[],"has_content":{"pdf":false,"grobid_xml":false},"content_urls":null,"referenced_works_count":0,"referenced_works":[],"related_works":[],"abstract_inverted_index":{"The":[0],"Mixture":[1],"of":[2,168],"Experts":[3],"MoE":[4,107,144],"architecture":[5],"is":[6],"highly":[7],"promising":[8],"for":[9,156],"resource":[10],"constrained":[11],"on":[12,141,157],"device":[13,142,158],"deployments":[14],"yet":[15],"training":[16,131],"these":[17],"models":[18,33,114,145],"from":[19,96],"scratch":[20],"incurs":[21],"prohibitive":[22],"costs":[23],"Current":[24],"methods":[25,175],"attempt":[26],"to":[27],"alleviate":[28],"this":[29,59],"by":[30,79,90],"upcycling":[31,71,174],"dense":[32,138,165],"into":[34,106,140],"MoEs":[35],"however":[36],"they":[37],"often":[38],"introduce":[39],"parameter":[40],"redundancy":[41,51],"that":[42,67,149],"degrades":[43],"inference":[44,87,159],"efficiency":[45],"Alternatively":[46],"standard":[47,173],"layer":[48],"pruning":[49,69,91],"mitigates":[50],"but":[52],"inevitably":[53],"compromises":[54],"model":[55,162],"accuracy":[56,163],"To":[57],"resolve":[58],"dilemma":[60],"we":[61],"propose":[62],"Dense2MoE":[63,83,133,150],"a":[64,127],"novel":[65],"framework":[66],"unifies":[68],"and":[70,117,172],"through":[72],"Layer":[73,103],"Fusion":[74],"UpCycling":[75],"LF":[76],"UC":[77],"Guided":[78],"hardware":[80],"Roofline":[81],"theory":[82],"systematically":[84],"overcomes":[85],"the":[86,113,153,169],"memory":[88],"wall":[89],"bandwidth":[92],"heavy":[93],"attention":[94],"modules":[95],"redundant":[97],"layers":[98],"while":[99],"repurposing":[100],"their":[101],"Multi":[102],"Perceptrons":[104],"MLPs":[105],"experts":[108],"This":[109],"structural":[110],"innovation":[111],"preserves":[112],"core":[115],"capabilities":[116],"strictly":[118],"limits":[119],"active":[120],"parameters":[121],"via":[122],"selective":[123],"token":[124],"routing":[125],"With":[126],"modest":[128],"continual":[129],"pre":[130],"budget":[132],"efficiently":[134],"converts":[135],"publicly":[136],"available":[137],"LLMs":[139],"ready":[143],"Extensive":[146],"experiments":[147],"demonstrate":[148],"significantly":[151],"advances":[152],"Pareto":[154],"frontier":[155],"latency":[160],"versus":[161],"outperforming":[164],"baselines":[166],"state":[167],"art":[170],"compression":[171]},"counts_by_year":[],"updated_date":"2026-07-01T06:00:48.157686","created_date":"2026-05-28T00:00:00"}