{"id":"https://openalex.org/W7155559730","doi":"https://doi.org/10.48550/arxiv.2604.20937","title":"Sink-Token-Aware Pruning for Fine-Grained Video Understanding in Efficient Video LLMs","display_name":"Sink-Token-Aware Pruning for Fine-Grained Video Understanding in Efficient Video LLMs","publication_year":2026,"publication_date":"2026-04-22","ids":{"openalex":"https://openalex.org/W7155559730","doi":"https://doi.org/10.48550/arxiv.2604.20937"},"language":null,"primary_location":{"id":"doi:10.48550/arxiv.2604.20937","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2604.20937","pdf_url":null,"source":{"id":"https://openalex.org/S4306400194","display_name":"arXiv (Cornell University)","issn_l":"2331-8422","issn":["2331-8422"],"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.2604.20937","any_repository_has_fulltext":true},"authorships":[{"author_position":"first","author":{"id":"https://openalex.org/A5134497435","display_name":"Kibum Kim","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Kim, Kibum","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5134502915","display_name":"Jiwan Kim","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Kim, Jiwan","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5012926287","display_name":"Kyle Min","orcid":"https://orcid.org/0000-0002-3978-918X"},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Min, Kyle","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5134539879","display_name":"Yueqi Wang","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Wang, Yueqi","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5047583621","display_name":"Jinyoung Moon","orcid":"https://orcid.org/0000-0002-6616-824X"},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Moon, Jinyoung","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5134553429","display_name":"Julian McAuley","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"McAuley, Julian","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"last","author":{"id":"https://openalex.org/A5134563896","display_name":"Chanyoung Park","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Park, Chanyoung","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":0.9779000282287598,"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":0.9779000282287598,"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":0.006399999838322401,"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/T10775","display_name":"Generative Adversarial Networks and Image Synthesis","score":0.0026000000070780516,"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/pruning","display_name":"Pruning","score":0.7166000008583069},{"id":"https://openalex.org/keywords/security-token","display_name":"Security token","score":0.6730999946594238},{"id":"https://openalex.org/keywords/inference","display_name":"Inference","score":0.6334999799728394},{"id":"https://openalex.org/keywords/obstacle","display_name":"Obstacle","score":0.4523000121116638},{"id":"https://openalex.org/keywords/latency","display_name":"Latency (audio)","score":0.42179998755455017},{"id":"https://openalex.org/keywords/key","display_name":"Key (lock)","score":0.37450000643730164},{"id":"https://openalex.org/keywords/blindness","display_name":"Blindness","score":0.36579999327659607}],"concepts":[{"id":"https://openalex.org/C41008148","wikidata":"https://www.wikidata.org/wiki/Q21198","display_name":"Computer science","level":0,"score":0.7806000113487244},{"id":"https://openalex.org/C108010975","wikidata":"https://www.wikidata.org/wiki/Q500094","display_name":"Pruning","level":2,"score":0.7166000008583069},{"id":"https://openalex.org/C48145219","wikidata":"https://www.wikidata.org/wiki/Q1335365","display_name":"Security token","level":2,"score":0.6730999946594238},{"id":"https://openalex.org/C2776214188","wikidata":"https://www.wikidata.org/wiki/Q408386","display_name":"Inference","level":2,"score":0.6334999799728394},{"id":"https://openalex.org/C154945302","wikidata":"https://www.wikidata.org/wiki/Q11660","display_name":"Artificial intelligence","level":1,"score":0.5964000225067139},{"id":"https://openalex.org/C119857082","wikidata":"https://www.wikidata.org/wiki/Q2539","display_name":"Machine learning","level":1,"score":0.5648999810218811},{"id":"https://openalex.org/C2776650193","wikidata":"https://www.wikidata.org/wiki/Q264661","display_name":"Obstacle","level":2,"score":0.4523000121116638},{"id":"https://openalex.org/C82876162","wikidata":"https://www.wikidata.org/wiki/Q17096504","display_name":"Latency (audio)","level":2,"score":0.42179998755455017},{"id":"https://openalex.org/C26517878","wikidata":"https://www.wikidata.org/wiki/Q228039","display_name":"Key (lock)","level":2,"score":0.37450000643730164},{"id":"https://openalex.org/C2780929884","wikidata":"https://www.wikidata.org/wiki/Q737460","display_name":"Blindness","level":2,"score":0.36579999327659607},{"id":"https://openalex.org/C36464697","wikidata":"https://www.wikidata.org/wiki/Q451553","display_name":"Visualization","level":2,"score":0.30730000138282776},{"id":"https://openalex.org/C143050476","wikidata":"https://www.wikidata.org/wiki/Q194502","display_name":"Sink (geography)","level":2,"score":0.27129998803138733},{"id":"https://openalex.org/C2776401178","wikidata":"https://www.wikidata.org/wiki/Q12050496","display_name":"Feature (linguistics)","level":2,"score":0.26820001006126404},{"id":"https://openalex.org/C31972630","wikidata":"https://www.wikidata.org/wiki/Q844240","display_name":"Computer vision","level":1,"score":0.26649999618530273},{"id":"https://openalex.org/C46637626","wikidata":"https://www.wikidata.org/wiki/Q6693015","display_name":"Low latency (capital markets)","level":2,"score":0.26489999890327454},{"id":"https://openalex.org/C137293760","wikidata":"https://www.wikidata.org/wiki/Q3621696","display_name":"Language model","level":2,"score":0.25929999351501465},{"id":"https://openalex.org/C2780586882","wikidata":"https://www.wikidata.org/wiki/Q7520643","display_name":"Simple (philosophy)","level":2,"score":0.2563999891281128},{"id":"https://openalex.org/C111370547","wikidata":"https://www.wikidata.org/wiki/Q7451120","display_name":"Sensory cue","level":2,"score":0.25209999084472656},{"id":"https://openalex.org/C158495155","wikidata":"https://www.wikidata.org/wiki/Q2369151","display_name":"Visual search","level":2,"score":0.2500999867916107}],"mesh":[],"locations_count":1,"locations":[{"id":"doi:10.48550/arxiv.2604.20937","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2604.20937","pdf_url":null,"source":{"id":"https://openalex.org/S4306400194","display_name":"arXiv (Cornell University)","issn_l":"2331-8422","issn":["2331-8422"],"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.2604.20937","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2604.20937","pdf_url":null,"source":{"id":"https://openalex.org/S4306400194","display_name":"arXiv (Cornell University)","issn_l":"2331-8422","issn":["2331-8422"],"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":{"pdf":false,"grobid_xml":false},"content_urls":null,"referenced_works_count":0,"referenced_works":[],"related_works":[],"abstract_inverted_index":{"Video":[0],"Large":[1],"Language":[2],"Models":[3],"(Video":[4],"LLMs)":[5],"incur":[6],"high":[7],"inference":[8],"latency":[9],"due":[10],"to":[11,19,33,101,141,146,155,162,178,212],"a":[12,31,63,85,98,130,138,149],"large":[13],"number":[14],"of":[15,173,214],"visual":[16,25,73,115,215],"tokens":[17,93,108],"provided":[18],"LLMs.":[20],"To":[21,79,169],"address":[22],"this,":[23],"training-free":[24],"token":[26],"pruning":[27,160,180,210],"has":[28],"emerged":[29],"as":[30,76,148],"solution":[32],"reduce":[34],"computational":[35],"costs;":[36],"however,":[37],"existing":[38,156],"methods":[39,61,161,181],"are":[40],"primarily":[41],"validated":[42],"on":[43,67],"Multiple-Choice":[44],"Question":[45],"Answering":[46],"(MCQA)":[47],"benchmarks,":[48],"where":[49],"coarse-grained":[50],"cues":[51],"often":[52],"suffice.":[53],"In":[54],"this":[55,81,153],"work,":[56],"we":[57,83,125,175],"reveal":[58],"that":[59,94,136,203],"these":[60,106,123],"suffer":[62],"sharp":[64],"performance":[65],"collapse":[66],"fine-grained":[68,102,119],"understanding":[69],"tasks":[70],"requiring":[71],"precise":[72],"grounding,":[74],"such":[75],"hallucination":[77],"evaluation.":[78],"explore":[80],"gap,":[82],"conduct":[84],"systematic":[86],"analysis":[87],"and":[88,117,151,158,184,186,198],"identify":[89],"sink":[90,107,139,150],"tokens--semantically":[91],"uninformative":[92],"attract":[95],"excessive":[96],"attention--as":[97],"key":[99],"obstacle":[100],"video":[103,167],"understanding.":[104,120,168],"When":[105],"survive":[109],"pruning,":[110],"they":[111],"distort":[112],"the":[113,171],"model's":[114],"evidence":[116],"hinder":[118],"Motivated":[121],"by":[122],"insights,":[124],"propose":[126],"Sink-Token-aware":[127],"Pruning":[128],"(SToP),":[129],"simple":[131],"yet":[132],"effective":[133],"plug-and-play":[134],"method":[135],"introduces":[137],"score":[140,154],"quantify":[142],"each":[143],"token's":[144],"tendency":[145],"behave":[147],"applies":[152],"spatial":[157],"temporal":[159],"suppress":[163],"them,":[164],"thereby":[165],"enhancing":[166],"validate":[170],"effectiveness":[172],"SToP,":[174],"apply":[176],"it":[177,188],"state-of-the-art":[179],"(VisionZip,":[182],"FastVid,":[183],"Holitom)":[185],"evaluate":[187],"across":[189],"diverse":[190],"benchmarks":[191],"covering":[192],"hallucination,":[193],"open-ended":[194],"generation,":[195],"compositional":[196],"reasoning,":[197],"MCQA.":[199],"Our":[200],"results":[201],"demonstrate":[202],"SToP":[204],"significantly":[205],"boosts":[206],"performance,":[207],"even":[208],"when":[209],"up":[211],"90%":[213],"tokens.":[216]},"counts_by_year":[],"updated_date":"2026-07-28T07:46:37.118299","created_date":"2026-04-25T00:00:00"}