{"id":"https://openalex.org/W7166090344","doi":"https://doi.org/10.48550/arxiv.2606.27161","title":"TOPS: First-Principles Visual Token Pruning via Constructing Token Optimal Preservation Sets for Efficient MLLM Inference","display_name":"TOPS: First-Principles Visual Token Pruning via Constructing Token Optimal Preservation Sets for Efficient MLLM Inference","publication_year":2026,"publication_date":"2026-06-25","ids":{"openalex":"https://openalex.org/W7166090344","doi":"https://doi.org/10.48550/arxiv.2606.27161"},"language":null,"primary_location":{"id":"doi:10.48550/arxiv.2606.27161","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2606.27161","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.2606.27161","any_repository_has_fulltext":true},"authorships":[{"author_position":"first","author":{"id":"https://openalex.org/A5136718043","display_name":"Tinghao Wang","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Wang, Tinghao","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5101893660","display_name":"Yi Guo","orcid":"https://orcid.org/0009-0006-1890-6726"},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Guo, Yichen","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5139456836","display_name":"Rui Huang","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Huang, Rui","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5139415491","display_name":"Zheng Lu","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Lu, Zheng","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5036390522","display_name":"Q. Zhang","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Zhang, Qizhe","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5139406885","display_name":"Chenxi Li","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Li, Chenxi","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5139454868","display_name":"Yuan Zhang","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Zhang, Yuan","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5139397235","display_name":"Jiajun Cao","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Cao, Jiajun","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5139453578","display_name":"Zhirong Shen","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Shen, Zhirong","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5134808768","display_name":"Yaosong Du","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Du, Yaosong","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5032441902","display_name":"Guangyan Gan","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Gan, Guangyan","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5139440623","display_name":"Wenya Wang","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Wang, Wenya","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5080547297","display_name":"Lin William Cong","orcid":"https://orcid.org/0000-0002-2617-2367"},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Cong, Lin William","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"last","author":{"id":"https://openalex.org/A5139395583","display_name":"Shanghang Zhang","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Zhang, Shanghang","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":0.8903999924659729,"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":0.8903999924659729,"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/T10028","display_name":"Topic Modeling","score":0.02630000002682209,"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/T10181","display_name":"Natural Language Processing Techniques","score":0.01489999983459711,"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/security-token","display_name":"Security token","score":0.8644000291824341},{"id":"https://openalex.org/keywords/pruning","display_name":"Pruning","score":0.8105000257492065},{"id":"https://openalex.org/keywords/inference","display_name":"Inference","score":0.5133000016212463},{"id":"https://openalex.org/keywords/task","display_name":"Task (project management)","score":0.46399998664855957},{"id":"https://openalex.org/keywords/perspective","display_name":"Perspective (graphical)","score":0.32589998841285706},{"id":"https://openalex.org/keywords/visualization","display_name":"Visualization","score":0.3181999921798706}],"concepts":[{"id":"https://openalex.org/C48145219","wikidata":"https://www.wikidata.org/wiki/Q1335365","display_name":"Security token","level":2,"score":0.8644000291824341},{"id":"https://openalex.org/C108010975","wikidata":"https://www.wikidata.org/wiki/Q500094","display_name":"Pruning","level":2,"score":0.8105000257492065},{"id":"https://openalex.org/C41008148","wikidata":"https://www.wikidata.org/wiki/Q21198","display_name":"Computer science","level":0,"score":0.8090999722480774},{"id":"https://openalex.org/C154945302","wikidata":"https://www.wikidata.org/wiki/Q11660","display_name":"Artificial intelligence","level":1,"score":0.5634999871253967},{"id":"https://openalex.org/C2776214188","wikidata":"https://www.wikidata.org/wiki/Q408386","display_name":"Inference","level":2,"score":0.5133000016212463},{"id":"https://openalex.org/C2780451532","wikidata":"https://www.wikidata.org/wiki/Q759676","display_name":"Task (project management)","level":2,"score":0.46399998664855957},{"id":"https://openalex.org/C119857082","wikidata":"https://www.wikidata.org/wiki/Q2539","display_name":"Machine learning","level":1,"score":0.40540000796318054},{"id":"https://openalex.org/C12713177","wikidata":"https://www.wikidata.org/wiki/Q1900281","display_name":"Perspective (graphical)","level":2,"score":0.32589998841285706},{"id":"https://openalex.org/C36464697","wikidata":"https://www.wikidata.org/wiki/Q451553","display_name":"Visualization","level":2,"score":0.3181999921798706},{"id":"https://openalex.org/C115067241","wikidata":"https://www.wikidata.org/wiki/Q1639854","display_name":"Token passing","level":3,"score":0.31130000948905945},{"id":"https://openalex.org/C80444323","wikidata":"https://www.wikidata.org/wiki/Q2878974","display_name":"Theoretical computer science","level":1,"score":0.2890999913215637},{"id":"https://openalex.org/C179799912","wikidata":"https://www.wikidata.org/wiki/Q205084","display_name":"Computational complexity theory","level":2,"score":0.27790001034736633},{"id":"https://openalex.org/C195324797","wikidata":"https://www.wikidata.org/wiki/Q33742","display_name":"Natural language","level":2,"score":0.2775000035762787},{"id":"https://openalex.org/C2777675136","wikidata":"https://www.wikidata.org/wiki/Q835642","display_name":"TOPS","level":3,"score":0.26809999346733093},{"id":"https://openalex.org/C2777508537","wikidata":"https://www.wikidata.org/wiki/Q7936620","display_name":"Visual reasoning","level":2,"score":0.26739999651908875},{"id":"https://openalex.org/C175154964","wikidata":"https://www.wikidata.org/wiki/Q380077","display_name":"Task analysis","level":3,"score":0.26440000534057617}],"mesh":[],"locations_count":1,"locations":[{"id":"doi:10.48550/arxiv.2606.27161","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2606.27161","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.2606.27161","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2606.27161","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":[1,18],"language":[2],"models":[3],"(MLLMs)":[4],"have":[5],"achieved":[6],"strong":[7],"multimodal":[8],"reasoning":[9],"capabilities,":[10],"but":[11],"their":[12],"efficiency":[13],"is":[14],"limited":[15],"by":[16],"the":[17,68],"number":[19],"of":[20,67,71,161],"visual":[21,79,162,181],"tokens,":[22,46],"which":[23],"introduces":[24],"substantial":[25],"computational":[26],"overhead.":[27],"Visual":[28],"token":[29,72,80,107],"pruning":[30,81,127,153,179],"offers":[31],"a":[32,64,83,96,123],"natural":[33],"solution,":[34],"yet":[35],"existing":[36],"methods":[37,57,150],"are":[38,50],"imperfect:":[39],"attention-based":[40],"criteria":[41,49,61],"tend":[42],"to":[43,53,133],"retain":[44],"redundant":[45,180],"while":[47,164],"diversity-based":[48],"often":[51],"agnostic":[52],"user":[54],"instructions.":[55],"Even":[56],"that":[58,129,146,178],"combine":[59],"multiple":[60],"still":[62],"lack":[63],"principled":[65],"formulation":[66],"intrinsic":[69],"objective":[70],"pruning.":[73],"In":[74],"this":[75],"paper,":[76],"we":[77,100,120],"revisit":[78],"from":[82],"first-principles":[84],"perspective":[85],"and":[86,113,125,142,167,173,187],"formulate":[87],"it":[88],"as":[89],"constructing":[90],"Token":[91],"Optimal":[92],"Preservation":[93],"Sets.":[94],"Through":[95],"top-down":[97],"information-theoretic":[98],"analysis,":[99],"identify":[101],"three":[102],"fundamental":[103],"principles":[104],"for":[105],"effective":[106],"selection:":[108],"Task":[109],"Relevance,":[110],"Information":[111],"Coverage,":[112],"Semantic":[114],"Diversity.":[115],"Based":[116],"on":[117,138,156,170],"these":[118],"principles,":[119],"propose":[121],"TOPS,":[122],"training-free":[124],"model-agnostic":[126],"module":[128],"can":[130,183],"be":[131],"applied":[132],"various":[134],"MLLMs.":[135],"Extensive":[136],"experiments":[137],"7":[139],"MLLM":[140,191],"backbones":[141],"14":[143],"benchmarks":[144],"demonstrate":[145],"TOPS":[147,158],"outperforms":[148],"prior":[149],"under":[151],"diverse":[152],"settings.":[154],"Notably,":[155],"LLaVA-NeXT,":[157],"removes":[159],"77.8%":[160],"tokens":[163,182],"preserving":[165],"100.0%":[166],"100.6%":[168],"performance":[169],"its":[171],"7B":[172],"13B":[174],"models,":[175],"respectively,":[176],"suggesting":[177],"sometimes":[184],"mitigate":[185],"hallucination":[186],"inspire":[188],"future":[189],"lightweight":[190],"design.":[192]},"counts_by_year":[],"updated_date":"2026-07-01T06:00:48.157686","created_date":"2026-06-27T00:00:00"}