{"id":"https://openalex.org/W7133360479","doi":"https://doi.org/10.48550/arxiv.2603.00510","title":"What Do Visual Tokens Really Encode? Uncovering Sparsity and Redundancy in Multimodal Large Language Models","display_name":"What Do Visual Tokens Really Encode? Uncovering Sparsity and Redundancy in Multimodal Large Language Models","publication_year":2026,"publication_date":"2026-02-28","ids":{"openalex":"https://openalex.org/W7133360479","doi":"https://doi.org/10.48550/arxiv.2603.00510"},"language":null,"primary_location":{"id":"doi:10.48550/arxiv.2603.00510","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2603.00510","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.2603.00510","any_repository_has_fulltext":true},"authorships":[{"author_position":"first","author":{"id":"https://openalex.org/A5127900756","display_name":"Yingqi Fan","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Fan, Yingqi","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5127883786","display_name":"Junlong Tong","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Tong, Junlong","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5127902329","display_name":"Anhao Zhao","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Zhao, Anhao","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"last","author":{"id":"https://openalex.org/A5127877804","display_name":"Xiaoyu Shen","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Shen, Xiaoyu","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":0.8543000221252441,"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":0.8543000221252441,"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":0.016499999910593033,"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":0.014800000004470348,"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/embedding","display_name":"Embedding","score":0.732699990272522},{"id":"https://openalex.org/keywords/redundancy","display_name":"Redundancy (engineering)","score":0.6295999884605408},{"id":"https://openalex.org/keywords/security-token","display_name":"Security token","score":0.6248000264167786},{"id":"https://openalex.org/keywords/encode","display_name":"ENCODE","score":0.5874000191688538},{"id":"https://openalex.org/keywords/semantics","display_name":"Semantics (computer science)","score":0.4830000102519989},{"id":"https://openalex.org/keywords/language-model","display_name":"Language model","score":0.42570000886917114},{"id":"https://openalex.org/keywords/visualization","display_name":"Visualization","score":0.3874000012874603},{"id":"https://openalex.org/keywords/computation","display_name":"Computation","score":0.38359999656677246},{"id":"https://openalex.org/keywords/encoding","display_name":"Encoding (memory)","score":0.3490999937057495}],"concepts":[{"id":"https://openalex.org/C41008148","wikidata":"https://www.wikidata.org/wiki/Q21198","display_name":"Computer science","level":0,"score":0.8181999921798706},{"id":"https://openalex.org/C41608201","wikidata":"https://www.wikidata.org/wiki/Q980509","display_name":"Embedding","level":2,"score":0.732699990272522},{"id":"https://openalex.org/C152124472","wikidata":"https://www.wikidata.org/wiki/Q1204361","display_name":"Redundancy (engineering)","level":2,"score":0.6295999884605408},{"id":"https://openalex.org/C48145219","wikidata":"https://www.wikidata.org/wiki/Q1335365","display_name":"Security token","level":2,"score":0.6248000264167786},{"id":"https://openalex.org/C66746571","wikidata":"https://www.wikidata.org/wiki/Q1134833","display_name":"ENCODE","level":3,"score":0.5874000191688538},{"id":"https://openalex.org/C154945302","wikidata":"https://www.wikidata.org/wiki/Q11660","display_name":"Artificial intelligence","level":1,"score":0.5239999890327454},{"id":"https://openalex.org/C184337299","wikidata":"https://www.wikidata.org/wiki/Q1437428","display_name":"Semantics (computer science)","level":2,"score":0.4830000102519989},{"id":"https://openalex.org/C137293760","wikidata":"https://www.wikidata.org/wiki/Q3621696","display_name":"Language model","level":2,"score":0.42570000886917114},{"id":"https://openalex.org/C204321447","wikidata":"https://www.wikidata.org/wiki/Q30642","display_name":"Natural language processing","level":1,"score":0.40950000286102295},{"id":"https://openalex.org/C36464697","wikidata":"https://www.wikidata.org/wiki/Q451553","display_name":"Visualization","level":2,"score":0.3874000012874603},{"id":"https://openalex.org/C45374587","wikidata":"https://www.wikidata.org/wiki/Q12525525","display_name":"Computation","level":2,"score":0.38359999656677246},{"id":"https://openalex.org/C125411270","wikidata":"https://www.wikidata.org/wiki/Q18653","display_name":"Encoding (memory)","level":2,"score":0.3490999937057495},{"id":"https://openalex.org/C2775945657","wikidata":"https://www.wikidata.org/wiki/Q381442","display_name":"Structuring","level":2,"score":0.33889999985694885},{"id":"https://openalex.org/C2776760102","wikidata":"https://www.wikidata.org/wiki/Q5139990","display_name":"Code (set theory)","level":3,"score":0.33489999175071716},{"id":"https://openalex.org/C2777508537","wikidata":"https://www.wikidata.org/wiki/Q7936620","display_name":"Visual reasoning","level":2,"score":0.3246000111103058},{"id":"https://openalex.org/C2778251979","wikidata":"https://www.wikidata.org/wiki/Q7936617","display_name":"Visual processing","level":3,"score":0.3149000108242035},{"id":"https://openalex.org/C160086991","wikidata":"https://www.wikidata.org/wiki/Q5939193","display_name":"Human visual system model","level":3,"score":0.2939999997615814},{"id":"https://openalex.org/C207363949","wikidata":"https://www.wikidata.org/wiki/Q462915","display_name":"Visual space","level":3,"score":0.2897999882698059},{"id":"https://openalex.org/C2780878386","wikidata":"https://www.wikidata.org/wiki/Q1659648","display_name":"Visual language","level":2,"score":0.2824999988079071},{"id":"https://openalex.org/C44291984","wikidata":"https://www.wikidata.org/wiki/Q1074173","display_name":"Question answering","level":2,"score":0.27950000762939453},{"id":"https://openalex.org/C178253425","wikidata":"https://www.wikidata.org/wiki/Q162668","display_name":"Visual perception","level":3,"score":0.27059999108314514},{"id":"https://openalex.org/C2983448237","wikidata":"https://www.wikidata.org/wiki/Q1078276","display_name":"Language understanding","level":2,"score":0.26750001311302185},{"id":"https://openalex.org/C2780330621","wikidata":"https://www.wikidata.org/wiki/Q7936609","display_name":"Visual modeling","level":4,"score":0.2623000144958496},{"id":"https://openalex.org/C165696696","wikidata":"https://www.wikidata.org/wiki/Q11287","display_name":"Exploit","level":2,"score":0.25440001487731934},{"id":"https://openalex.org/C2776214188","wikidata":"https://www.wikidata.org/wiki/Q408386","display_name":"Inference","level":2,"score":0.2531000077724457},{"id":"https://openalex.org/C80444323","wikidata":"https://www.wikidata.org/wiki/Q2878974","display_name":"Theoretical computer science","level":1,"score":0.2524000108242035}],"mesh":[],"locations_count":1,"locations":[{"id":"doi:10.48550/arxiv.2603.00510","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2603.00510","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.2603.00510","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2603.00510","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":{"Multimodal":[0],"large":[1],"language":[2,13],"models":[3],"(MLLMs)":[4],"project":[5],"visual":[6,22,57,109,113,178,196],"tokens":[7,58,92,142],"into":[8,61],"the":[9,16,54,69,75,106,125,151,182],"embedding":[10,153],"space":[11],"of":[12,21,74,128,177],"models,":[14],"yet":[15],"internal":[17,136],"structuring":[18],"and":[19,64,101,115,161,187,198],"processing":[20,158],"semantics":[23],"remain":[24],"poorly":[25],"understood.":[26],"In":[27],"this":[28],"work,":[29],"we":[30,87,138],"introduce":[31],"a":[32,37,44,49,83,173],"two-fold":[33],"analytical":[34],"framework":[35],"featuring":[36],"novel":[38],"probing":[39],"tool,":[40],"$\\textbf{EmbedLens}$,":[41],"to":[42,104],"conduct":[43],"fine-grained":[45,96],"analysis.":[46],"We":[47],"uncover":[48],"pronounced":[50],"semantic":[51],"sparsity":[52],"at":[53],"input":[55],"level:":[56],"consistently":[59],"partition":[60],"sink,":[62],"dead,":[63],"alive":[65,70,91,141],"categories.":[66],"Remarkably,":[67],"only":[68],"tokens,":[71],"comprising":[72],"$\\approx60\\%$":[73],"total":[76],"input,":[77],"carry":[78],"image-specific":[79],"meaning.":[80],"Furthermore,":[81],"using":[82],"targeted":[84],"patch-compression":[85],"benchmark,":[86],"demonstrate":[88],"that":[89,132,140,156,162],"these":[90],"already":[93],"encode":[94],"rich,":[95],"cues":[97],"(e.g.,":[98],"objects,":[99],"colors,":[100],"OCR)":[102],"prior":[103],"entering":[105],"LLM.":[107],"Internal":[108],"computations":[110],"(such":[111],"as":[112],"attention":[114],"feed-forward":[116],"networks)":[117],"are":[118],"redundant":[119],"for":[120,184],"most":[121],"standard":[122],"tasks.":[123],"For":[124],"small":[126],"subset":[127],"highly":[129],"vision-centric":[130],"tasks":[131],"actually":[133],"benefit":[134],"from":[135],"processing,":[137,180],"reveal":[139],"naturally":[143],"align":[144],"with":[145],"intermediate":[146],"LLM":[147],"layers":[148],"rather":[149],"than":[150],"initial":[152],"space,":[154],"indicating":[155],"shallow-layer":[157],"is":[159,166,203],"unnecessary":[160],"direct":[163],"mid-layer":[164,199],"injection":[165],"both":[167],"sufficient.":[168],"Ultimately,":[169],"our":[170],"findings":[171],"provide":[172],"unified":[174],"mechanistic":[175],"view":[176],"token":[179,193],"paving":[181],"way":[183],"more":[185],"efficient":[186],"interpretable":[188],"MLLM":[189],"architectures":[190],"through":[191],"selective":[192],"pruning,":[194],"minimized":[195],"computation,":[197],"injection.":[200],"The":[201],"code":[202],"released":[204],"at:":[205],"https://github.com/EIT-NLP/EmbedLens.":[206]},"counts_by_year":[],"updated_date":"2026-07-01T06:00:48.157686","created_date":"2026-03-04T00:00:00"}