{"id":"https://openalex.org/W7164166617","doi":"https://doi.org/10.48550/arxiv.2606.10722","title":"Continual LLM Upcycling: A Predictor-Gated Bank-Wise Sparsity Training Recipe for Dense-to-Sparse LLMs","display_name":"Continual LLM Upcycling: A Predictor-Gated Bank-Wise Sparsity Training Recipe for Dense-to-Sparse LLMs","publication_year":2026,"publication_date":"2026-06-09","ids":{"openalex":"https://openalex.org/W7164166617","doi":"https://doi.org/10.48550/arxiv.2606.10722"},"language":null,"primary_location":{"id":"doi:10.48550/arxiv.2606.10722","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2606.10722","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.2606.10722","any_repository_has_fulltext":true},"authorships":[{"author_position":"first","author":{"id":"https://openalex.org/A5102447561","display_name":"Ruixuan Huang","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Huang, Ruixuan","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5138337978","display_name":"Jinyuan Shi","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Shi, Jinyuan","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5078800069","display_name":"Hantao Huang","orcid":"https://orcid.org/0000-0002-2533-2082"},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Huang, Hantao","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5138393961","display_name":"Yifan Huang","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Huang, Yifan","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5138338454","display_name":"Ziyi Guan","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Guan, Ziyi","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5138293432","display_name":"Hao Zeng","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Zeng, Hao","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5073803011","display_name":"Ian En-Hsu Yen","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Yen, Ian En-Hsu","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"last","author":{"id":"https://openalex.org/A5138386952","display_name":"Minghui Yu","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Yu, Minghui","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/T12326","display_name":"Network Packet Processing and Optimization","score":0.12849999964237213,"subfield":{"id":"https://openalex.org/subfields/1708","display_name":"Hardware and Architecture"},"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/T12326","display_name":"Network Packet Processing and Optimization","score":0.12849999964237213,"subfield":{"id":"https://openalex.org/subfields/1708","display_name":"Hardware and Architecture"},"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/T10181","display_name":"Natural Language Processing Techniques","score":0.1128000020980835,"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/T10054","display_name":"Parallel Computing and Optimization Techniques","score":0.08730000257492065,"subfield":{"id":"https://openalex.org/subfields/1708","display_name":"Hardware and Architecture"},"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/training","display_name":"Training (meteorology)","score":0.6682000160217285},{"id":"https://openalex.org/keywords/path","display_name":"Path (computing)","score":0.6290000081062317},{"id":"https://openalex.org/keywords/context","display_name":"Context (archaeology)","score":0.6032999753952026},{"id":"https://openalex.org/keywords/benchmark","display_name":"Benchmark (surveying)","score":0.5968999862670898},{"id":"https://openalex.org/keywords/construct","display_name":"Construct (python library)","score":0.5616000294685364},{"id":"https://openalex.org/keywords/inference","display_name":"Inference","score":0.5605000257492065},{"id":"https://openalex.org/keywords/routing","display_name":"Routing (electronic design automation)","score":0.5267999768257141}],"concepts":[{"id":"https://openalex.org/C41008148","wikidata":"https://www.wikidata.org/wiki/Q21198","display_name":"Computer science","level":0,"score":0.6794999837875366},{"id":"https://openalex.org/C2777211547","wikidata":"https://www.wikidata.org/wiki/Q17141490","display_name":"Training (meteorology)","level":2,"score":0.6682000160217285},{"id":"https://openalex.org/C2777735758","wikidata":"https://www.wikidata.org/wiki/Q817765","display_name":"Path (computing)","level":2,"score":0.6290000081062317},{"id":"https://openalex.org/C2779343474","wikidata":"https://www.wikidata.org/wiki/Q3109175","display_name":"Context (archaeology)","level":2,"score":0.6032999753952026},{"id":"https://openalex.org/C185798385","wikidata":"https://www.wikidata.org/wiki/Q1161707","display_name":"Benchmark (surveying)","level":2,"score":0.5968999862670898},{"id":"https://openalex.org/C2780801425","wikidata":"https://www.wikidata.org/wiki/Q5164392","display_name":"Construct (python library)","level":2,"score":0.5616000294685364},{"id":"https://openalex.org/C2776214188","wikidata":"https://www.wikidata.org/wiki/Q408386","display_name":"Inference","level":2,"score":0.5605000257492065},{"id":"https://openalex.org/C119857082","wikidata":"https://www.wikidata.org/wiki/Q2539","display_name":"Machine learning","level":1,"score":0.5476999878883362},{"id":"https://openalex.org/C74172769","wikidata":"https://www.wikidata.org/wiki/Q1446839","display_name":"Routing (electronic design automation)","level":2,"score":0.5267999768257141},{"id":"https://openalex.org/C154945302","wikidata":"https://www.wikidata.org/wiki/Q11660","display_name":"Artificial intelligence","level":1,"score":0.5141000151634216},{"id":"https://openalex.org/C2778671685","wikidata":"https://www.wikidata.org/wiki/Q219239","display_name":"Recipe","level":2,"score":0.39480000734329224},{"id":"https://openalex.org/C51632099","wikidata":"https://www.wikidata.org/wiki/Q3985153","display_name":"Training set","level":2,"score":0.39469999074935913},{"id":"https://openalex.org/C137293760","wikidata":"https://www.wikidata.org/wiki/Q3621696","display_name":"Language model","level":2,"score":0.3476000130176544},{"id":"https://openalex.org/C48145219","wikidata":"https://www.wikidata.org/wiki/Q1335365","display_name":"Security token","level":2,"score":0.3102000057697296},{"id":"https://openalex.org/C124101348","wikidata":"https://www.wikidata.org/wiki/Q172491","display_name":"Data mining","level":1,"score":0.3098999857902527},{"id":"https://openalex.org/C2780009758","wikidata":"https://www.wikidata.org/wiki/Q6804172","display_name":"Measure (data warehouse)","level":2,"score":0.298799991607666},{"id":"https://openalex.org/C48677424","wikidata":"https://www.wikidata.org/wiki/Q6888088","display_name":"Mode (computer interface)","level":2,"score":0.2944999933242798},{"id":"https://openalex.org/C56372850","wikidata":"https://www.wikidata.org/wiki/Q1050404","display_name":"Sparse matrix","level":3,"score":0.2906999886035919}],"mesh":[],"locations_count":1,"locations":[{"id":"doi:10.48550/arxiv.2606.10722","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2606.10722","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.2606.10722","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2606.10722","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":{"We":[0,55,111,122],"study":[1],"dense-to-sparse":[2],"continual":[3,97],"training":[4,25,115,120],"as":[5],"a":[6,19,31,47,58,107,125,134],"way":[7],"to":[8,50,62,103],"construct":[9],"channel-sparse":[10],"large":[11],"language":[12,91],"models":[13],"from":[14,18],"dense":[15,21,101],"checkpoints.":[16],"Starting":[17],"Qwen2.5-8B":[20],"backbone,":[22],"we":[23,45],"continue":[24],"at":[26],"32K":[27,38],"context":[28],"and":[29,43,94,119,132],"introduce":[30],"predictor-gated":[32],"sparse":[33,80,109],"SwiGLU":[34],"FFN":[35,75],"in":[36,66,73],"the":[37,74,83,89,100,113,141],"stage.":[39],"For":[40],"each":[41],"token":[42],"layer,":[44],"use":[46],"low-rank":[48],"predictor":[49],"produce":[51],"FFN-channel":[52],"routing":[53,84],"logits.":[54],"then":[56],"apply":[57],"bank-wise":[59],"top-k":[60],"rule":[61],"retain":[63],"16":[64],"channels":[65],"every":[67],"64-channel":[68],"bank,":[69],"yielding":[70],"4x":[71],"sparsity":[72],"intermediate":[76],"activation.":[77],"Unlike":[78],"post-hoc":[79],"inference":[81],"methods,":[82],"module":[85],"is":[86],"placed":[87],"on":[88,130],"main":[90],"modeling":[92],"path":[93],"optimized":[95],"during":[96],"training,":[98],"enabling":[99],"model":[102],"be":[104],"upcycled":[105],"into":[106],"hardware-oriented":[108],"model.":[110],"report":[112],"architecture,":[114],"recipe,":[116],"benchmark":[117],"performance,":[118],"lessons.":[121],"also":[123],"identify":[124],"layer-local":[126],"long-context":[127],"failure":[128],"mode":[129],"RULER-CWE":[131],"propose":[133],"single-layer":[135],"repair":[136],"algorithm":[137],"that":[138],"substantially":[139],"improves":[140],"affected":[142],"length":[143],"range.":[144]},"counts_by_year":[],"updated_date":"2026-07-01T06:00:48.157686","created_date":"2026-06-11T00:00:00"}