{"id":"https://openalex.org/W7164013730","doi":"https://doi.org/10.48550/arxiv.2606.09131","title":"Late-Layer Fusion is Enough: Dual-Path Vision Token Routing for Multimodal Large Language Models under Visual Saturation","display_name":"Late-Layer Fusion is Enough: Dual-Path Vision Token Routing for Multimodal Large Language Models under Visual Saturation","publication_year":2026,"publication_date":"2026-06-08","ids":{"openalex":"https://openalex.org/W7164013730","doi":"https://doi.org/10.48550/arxiv.2606.09131"},"language":null,"primary_location":{"id":"doi:10.48550/arxiv.2606.09131","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2606.09131","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":null,"license_id":null,"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.2606.09131","any_repository_has_fulltext":true},"authorships":[{"author_position":"first","author":{"id":"https://openalex.org/A5138268165","display_name":"Siyuan Liu","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Liu, Siyuan","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"last","author":{"id":"https://openalex.org/A5138249977","display_name":"Jinyang Wu","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Wu, Jinyang","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/T11714","display_name":"Multimodal Machine Learning Applications","score":0.8091999888420105,"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/T11714","display_name":"Multimodal Machine Learning Applications","score":0.8091999888420105,"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/T10775","display_name":"Generative Adversarial Networks and Image Synthesis","score":0.03739999979734421,"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/T10036","display_name":"Advanced Neural Network Applications","score":0.03189999982714653,"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"}}],"keywords":[{"id":"https://openalex.org/keywords/security-token","display_name":"Security token","score":0.6796000003814697},{"id":"https://openalex.org/keywords/computation","display_name":"Computation","score":0.5853000283241272},{"id":"https://openalex.org/keywords/transformer","display_name":"Transformer","score":0.5005000233650208},{"id":"https://openalex.org/keywords/language-model","display_name":"Language model","score":0.3919999897480011},{"id":"https://openalex.org/keywords/traverse","display_name":"Traverse","score":0.3756999969482422},{"id":"https://openalex.org/keywords/edge-device","display_name":"Edge device","score":0.32429999113082886},{"id":"https://openalex.org/keywords/routing","display_name":"Routing (electronic design automation)","score":0.32260000705718994},{"id":"https://openalex.org/keywords/layer","display_name":"Layer (electronics)","score":0.3075999915599823}],"concepts":[{"id":"https://openalex.org/C41008148","wikidata":"https://www.wikidata.org/wiki/Q21198","display_name":"Computer science","level":0,"score":0.7457000017166138},{"id":"https://openalex.org/C48145219","wikidata":"https://www.wikidata.org/wiki/Q1335365","display_name":"Security token","level":2,"score":0.6796000003814697},{"id":"https://openalex.org/C45374587","wikidata":"https://www.wikidata.org/wiki/Q12525525","display_name":"Computation","level":2,"score":0.5853000283241272},{"id":"https://openalex.org/C154945302","wikidata":"https://www.wikidata.org/wiki/Q11660","display_name":"Artificial intelligence","level":1,"score":0.5415999889373779},{"id":"https://openalex.org/C66322947","wikidata":"https://www.wikidata.org/wiki/Q11658","display_name":"Transformer","level":3,"score":0.5005000233650208},{"id":"https://openalex.org/C137293760","wikidata":"https://www.wikidata.org/wiki/Q3621696","display_name":"Language model","level":2,"score":0.3919999897480011},{"id":"https://openalex.org/C176809094","wikidata":"https://www.wikidata.org/wiki/Q15401496","display_name":"Traverse","level":2,"score":0.3756999969482422},{"id":"https://openalex.org/C31972630","wikidata":"https://www.wikidata.org/wiki/Q844240","display_name":"Computer vision","level":1,"score":0.3357999920845032},{"id":"https://openalex.org/C138236772","wikidata":"https://www.wikidata.org/wiki/Q25098575","display_name":"Edge device","level":3,"score":0.32429999113082886},{"id":"https://openalex.org/C74172769","wikidata":"https://www.wikidata.org/wiki/Q1446839","display_name":"Routing (electronic design automation)","level":2,"score":0.32260000705718994},{"id":"https://openalex.org/C2779227376","wikidata":"https://www.wikidata.org/wiki/Q6505497","display_name":"Layer (electronics)","level":2,"score":0.3075999915599823},{"id":"https://openalex.org/C26760741","wikidata":"https://www.wikidata.org/wiki/Q160402","display_name":"Perception","level":2,"score":0.3073999881744385},{"id":"https://openalex.org/C28490314","wikidata":"https://www.wikidata.org/wiki/Q189436","display_name":"Speech recognition","level":1,"score":0.2985000014305115},{"id":"https://openalex.org/C108583219","wikidata":"https://www.wikidata.org/wiki/Q197536","display_name":"Deep learning","level":2,"score":0.287200003862381},{"id":"https://openalex.org/C195324797","wikidata":"https://www.wikidata.org/wiki/Q33742","display_name":"Natural language","level":2,"score":0.2847000062465668},{"id":"https://openalex.org/C118505674","wikidata":"https://www.wikidata.org/wiki/Q42586063","display_name":"Encoder","level":2,"score":0.2818000018596649},{"id":"https://openalex.org/C178253425","wikidata":"https://www.wikidata.org/wiki/Q162668","display_name":"Visual perception","level":3,"score":0.28040000796318054},{"id":"https://openalex.org/C26517878","wikidata":"https://www.wikidata.org/wiki/Q228039","display_name":"Key (lock)","level":2,"score":0.27959999442100525},{"id":"https://openalex.org/C2780226545","wikidata":"https://www.wikidata.org/wiki/Q6888030","display_name":"Modality (human\u2013computer interaction)","level":2,"score":0.2784000039100647},{"id":"https://openalex.org/C204321447","wikidata":"https://www.wikidata.org/wiki/Q30642","display_name":"Natural language processing","level":1,"score":0.27639999985694885},{"id":"https://openalex.org/C2984842247","wikidata":"https://www.wikidata.org/wiki/Q197536","display_name":"Deep neural networks","level":3,"score":0.25440001487731934}],"mesh":[],"locations_count":1,"locations":[{"id":"doi:10.48550/arxiv.2606.09131","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2606.09131","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":null,"license_id":null,"version":null,"is_accepted":false,"is_published":null,"raw_source_name":null,"raw_type":"Preprint"}],"best_oa_location":{"id":"doi:10.48550/arxiv.2606.09131","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2606.09131","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":null,"license_id":null,"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":{"Multimodal":[0],"large":[1],"language":[2,26],"models":[3],"(MLLMs)":[4],"commonly":[5],"inherit":[6],"the":[7,19,64,152,171,176,183,204,211],"deep,":[8],"symmetric":[9],"Transformer":[10,206],"backbone":[11],"designed":[12],"for":[13,139,234],"unimodal":[14],"text":[15,37,89],"modeling,":[16],"and":[17,25,36,45,81,106,115,174,178,223],"apply":[18],"same":[20],"computation":[21,114,202],"uniformly":[22],"to":[23,61,76,92],"image":[24,35,168],"tokens.":[27],"This":[28],"design":[29],"overlooks":[30],"a":[31,50,101,135,156,162,226],"key":[32],"modality":[33,108],"asymmetry:":[34],"tokens":[38,59,90,150,216],"differ":[39],"substantially":[40],"in":[41,63,111,118,170,203,239],"information":[42],"density,":[43],"redundancy,":[44],"required":[46],"reasoning":[47],"depth.":[48],"Through":[49],"layer-wise":[51],"analysis":[52],"of":[53],"LLaVA-1.5,":[54],"we":[55,128],"observe":[56],"that":[57,166,214,225],"vision":[58,149,215],"tend":[60],"saturate":[62],"middle":[65],"layers.":[66],"Specifically,":[67],"text-to-image":[68],"attention":[69],"decreases":[70],"from":[71,94],"0.68":[72],"at":[73,151,182],"layer":[74,79,86,230],"0":[75],"0.07":[77],"by":[78,126],"4,":[80],"stabilizes":[82],"near":[83],"0.04":[84],"after":[85],"18,":[87],"whereas":[88],"continue":[91],"benefit":[93],"deep":[95,122,172,205,220],"semantic":[96],"processing.":[97],"These":[98],"findings":[99],"suggest":[100],"mismatch":[102],"between":[103],"architectural":[104],"symmetry":[105],"depth-asynchronous":[107],"evolution,":[109],"resulting":[110],"redundant":[112],"visual":[113,177,201],"possible":[116],"drift":[117],"perceptual":[119,237],"representations":[120],"during":[121],"task-specific":[123],"adaptation.":[124],"Motivated":[125],"this,":[127],"propose":[129],"Dual-Path":[130],"Vision":[131],"Token":[132],"Routing":[133],"(DPVR),":[134],"modality-asymmetric":[136],"routing":[137],"framework":[138],"efficient":[140],"MLLMs.":[141,241],"Its":[142],"core":[143],"instantiation,":[144],"DPVR-LF":[145,191],"(Late-Layer":[146],"Fusion),":[147],"routes":[148],"saturation":[153],"point":[154],"into":[155],"one-layer":[157],"trainable":[158,189],"side":[159],"branch,":[160],"runs":[161],"thirteen-layer":[163],"text-only":[164],"forward":[165],"skips":[167],"positions":[169],"stack,":[173],"re-fuses":[175],"textual":[179],"streams":[180],"only":[181],"final":[184],"layer.":[185],"With":[186],"approximately":[187],"3%":[188],"parameters,":[190],"preserves":[192],"competitive":[193],"multimodal":[194],"performance":[195],"on":[196],"standard":[197],"benchmarks":[198],"while":[199],"reducing":[200],"stack.":[207],"The":[208],"results":[209],"challenge":[210],"conventional":[212],"assumption":[213],"must":[217],"traverse":[218],"all":[219],"language-model":[221],"layers,":[222],"indicate":[224],"single":[227],"late":[228],"fusion":[229],"can":[231],"be":[232],"sufficient":[233],"maintaining":[235],"strong":[236],"competence":[238],"LLaVA-style":[240]},"counts_by_year":[],"updated_date":"2026-07-01T06:00:48.157686","created_date":"2026-06-10T00:00:00"}