{"id":"https://openalex.org/W4417536759","doi":"https://doi.org/10.48550/arxiv.2506.15724","title":"MadaKV: Adaptive Modality-Perception KV Cache Eviction for Efficient Multimodal Long-Context Inference","display_name":"MadaKV: Adaptive Modality-Perception KV Cache Eviction for Efficient Multimodal Long-Context Inference","publication_year":2025,"publication_date":"2025-06-06","ids":{"openalex":"https://openalex.org/W4417536759","doi":"https://doi.org/10.48550/arxiv.2506.15724"},"language":"en","primary_location":{"id":"pmh:oai:arXiv.org:2506.15724","doi":null,"is_oa":true,"landing_page_url":"http://arxiv.org/abs/2506.15724","pdf_url":"https://arxiv.org/pdf/2506.15724","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,"listed_in":[],"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":"submittedVersion","is_accepted":false,"is_published":false,"raw_source_name":null,"raw_type":"text"},"volume":null,"type":"preprint","indexed_in":["arxiv","datacite"],"open_access":{"is_oa":true,"oa_status":"green","oa_url":"https://arxiv.org/pdf/2506.15724","any_repository_has_fulltext":true},"authorships":[{"author_position":"first","author":{"id":null,"display_name":"Li, Kunxi","orcid":null,"observed_orcids":[]},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Li, Kunxi","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":null,"display_name":"Jiang, Zhonghua","orcid":null,"observed_orcids":[]},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Jiang, Zhonghua","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":null,"display_name":"Shen, Zhouzhou","orcid":null,"observed_orcids":[]},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Shen, Zhouzhou","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":null,"display_name":"Wang, Zhaode","orcid":null,"observed_orcids":[]},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Wang, Zhaode","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":null,"display_name":"Lv, Chengfei","orcid":null,"observed_orcids":[]},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Lv, Chengfei","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":null,"display_name":"Zhang, Shengyu","orcid":null,"observed_orcids":[]},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Zhang, Shengyu","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5059190563","display_name":"Fan Wu","orcid":"https://orcid.org/0000-0003-0965-9058","observed_orcids":["https://orcid.org/0000-0003-0965-9058","https://orcid.org/0009-0007-0544-235X","https://orcid.org/0000-0001-5064-7062","https://orcid.org/0000-0001-9392-2597"]},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Wu, Fan","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"last","author":{"id":null,"display_name":"Wu, Fei","orcid":null,"observed_orcids":[]},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Wu, Fei","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":0.9672999978065491,"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":0.9672999978065491,"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":0.014600000344216824,"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/T10181","display_name":"Natural Language Processing Techniques","score":0.008299999870359898,"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/multimodal-large-language-models","display_name":"multimodal large language models","score":0.9739999771118164},{"id":"https://openalex.org/keywords/long-context-inference","display_name":"long-context inference","score":0.972000002861023},{"id":"https://openalex.org/keywords/modality-preference","display_name":"modality preference","score":0.7699999809265137},{"id":"https://openalex.org/keywords/attention-heads","display_name":"attention heads","score":0.7900000214576721},{"id":"https://openalex.org/keywords/hierarchical-compression","display_name":"hierarchical compression","score":0.6919999718666077},{"id":"https://openalex.org/keywords/decoding-delay","display_name":"decoding latency","score":0.7699999809265137}],"concepts":[{"id":"https://openalex.org/C115537543","wikidata":"https://www.wikidata.org/wiki/Q165596","display_name":"Cache","level":2,"score":0.7706000208854675},{"id":"https://openalex.org/C41008148","wikidata":"https://www.wikidata.org/wiki/Q21198","display_name":"Computer science","level":0,"score":0.7566999793052673},{"id":"https://openalex.org/C2776214188","wikidata":"https://www.wikidata.org/wiki/Q408386","display_name":"Inference","level":2,"score":0.6517999768257141},{"id":"https://openalex.org/C185798385","wikidata":"https://www.wikidata.org/wiki/Q1161707","display_name":"Benchmark (surveying)","level":2,"score":0.4968999922275543},{"id":"https://openalex.org/C38556500","wikidata":"https://www.wikidata.org/wiki/Q13404475","display_name":"Cache algorithms","level":4,"score":0.46939998865127563},{"id":"https://openalex.org/C2778227907","wikidata":"https://www.wikidata.org/wiki/Q1893186","display_name":"Eviction","level":2,"score":0.435699999332428},{"id":"https://openalex.org/C74912251","wikidata":"https://www.wikidata.org/wiki/Q6815727","display_name":"Memory footprint","level":2,"score":0.4205999970436096},{"id":"https://openalex.org/C82876162","wikidata":"https://www.wikidata.org/wiki/Q17096504","display_name":"Latency (audio)","level":2,"score":0.4187999963760376},{"id":"https://openalex.org/C26517878","wikidata":"https://www.wikidata.org/wiki/Q228039","display_name":"Key (lock)","level":2,"score":0.4099000096321106},{"id":"https://openalex.org/C25536678","wikidata":"https://www.wikidata.org/wiki/Q5015977","display_name":"Cache invalidation","level":5,"score":0.3628999888896942},{"id":"https://openalex.org/C189783530","wikidata":"https://www.wikidata.org/wiki/Q352090","display_name":"CPU cache","level":3,"score":0.3601999878883362},{"id":"https://openalex.org/C57273362","wikidata":"https://www.wikidata.org/wiki/Q576722","display_name":"Decoding methods","level":2,"score":0.35760000348091125},{"id":"https://openalex.org/C113166858","wikidata":"https://www.wikidata.org/wiki/Q5015981","display_name":"Cache pollution","level":5,"score":0.3562000095844269},{"id":"https://openalex.org/C79403827","wikidata":"https://www.wikidata.org/wiki/Q3988","display_name":"Real-time computing","level":1,"score":0.33169999718666077},{"id":"https://openalex.org/C201148951","wikidata":"https://www.wikidata.org/wiki/Q5015976","display_name":"Cache coloring","level":4,"score":0.3131999969482422},{"id":"https://openalex.org/C113775141","wikidata":"https://www.wikidata.org/wiki/Q428691","display_name":"Computer engineering","level":1,"score":0.3109000027179718},{"id":"https://openalex.org/C2780226545","wikidata":"https://www.wikidata.org/wiki/Q6888030","display_name":"Modality (human\u2013computer interaction)","level":2,"score":0.3052999973297119},{"id":"https://openalex.org/C139807058","wikidata":"https://www.wikidata.org/wiki/Q352374","display_name":"Adaptation (eye)","level":2,"score":0.29179999232292175},{"id":"https://openalex.org/C154945302","wikidata":"https://www.wikidata.org/wiki/Q11660","display_name":"Artificial intelligence","level":1,"score":0.28519999980926514},{"id":"https://openalex.org/C173608175","wikidata":"https://www.wikidata.org/wiki/Q232661","display_name":"Parallel computing","level":1,"score":0.27489998936653137},{"id":"https://openalex.org/C119857082","wikidata":"https://www.wikidata.org/wiki/Q2539","display_name":"Machine learning","level":1,"score":0.27459999918937683},{"id":"https://openalex.org/C774472","wikidata":"https://www.wikidata.org/wiki/Q6760393","display_name":"Margin (machine learning)","level":2,"score":0.26330000162124634},{"id":"https://openalex.org/C120314980","wikidata":"https://www.wikidata.org/wiki/Q180634","display_name":"Distributed computing","level":1,"score":0.25279998779296875}],"mesh":[],"locations_count":2,"locations":[{"id":"pmh:oai:arXiv.org:2506.15724","doi":null,"is_oa":true,"landing_page_url":"http://arxiv.org/abs/2506.15724","pdf_url":"https://arxiv.org/pdf/2506.15724","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,"listed_in":[],"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":"submittedVersion","is_accepted":false,"is_published":false,"raw_source_name":null,"raw_type":"text"},{"id":"doi:10.48550/arxiv.2506.15724","doi":"https://doi.org/10.48550/arxiv.2506.15724","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2506.15724","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,"listed_in":[],"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":"pmh:oai:arXiv.org:2506.15724","doi":null,"is_oa":true,"landing_page_url":"http://arxiv.org/abs/2506.15724","pdf_url":"https://arxiv.org/pdf/2506.15724","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,"listed_in":[],"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":"submittedVersion","is_accepted":false,"is_published":false,"raw_source_name":null,"raw_type":"text"},"sustainable_development_goals":[],"sustainable_development_goals_aurora":[],"study_designs":[],"awards":[],"funders":[],"has_content":{"grobid_xml":false,"pdf":false},"content_urls":null,"referenced_works_count":0,"referenced_works":[],"related_works":[],"abstract_inverted_index":{"This":[0],"paper":[1],"introduces":[2],"MadaKV,":[3],"a":[4],"modality-adaptive":[5],"key-value":[6],"(KV)":[7],"cache":[8,48,100,140],"eviction":[9,49,141],"strategy":[10],"designed":[11],"to":[12,58,109,137],"enhance":[13],"the":[14,128,132],"efficiency":[15],"of":[16,134],"multimodal":[17,26,119],"large":[18],"language":[19],"models":[20],"(MLLMs)":[21],"in":[22,37,40,98],"long-context":[23,120],"inference.":[24],"In":[25],"scenarios,":[27],"attention":[28,44,87],"heads":[29,88],"exhibit":[30],"varying":[31],"preferences":[32],"for":[33,54],"different":[34],"modalities,":[35],"resulting":[36],"significant":[38],"disparities":[39],"modality":[41,74,84],"importance":[42],"across":[43,117],"heads.":[45],"Traditional":[46],"KV":[47,99,139],"methods,":[50],"which":[51],"are":[52],"tailored":[53],"unimodal":[55],"settings,":[56],"fail":[57],"capture":[59],"modality-specific":[60],"information,":[61],"thereby":[62],"yielding":[63],"suboptimal":[64],"performance.":[65],"MadaKV":[66,94,135],"addresses":[67],"these":[68],"challenges":[69],"through":[70],"two":[71],"key":[72],"components:":[73],"preference":[75],"adaptation":[76],"and":[77,89,103,127],"hierarchical":[78],"compression":[79],"compensation.":[80],"By":[81],"dynamically":[82],"sensing":[83],"information":[85],"within":[86],"adaptively":[90],"retaining":[91],"critical":[92],"tokens,":[93],"achieves":[95],"substantial":[96],"reductions":[97],"memory":[101],"footprint":[102],"model":[104],"inference":[105],"decoding":[106],"latency":[107],"(1.3":[108],"1.5":[110],"times":[111],"improvement)":[112],"while":[113],"maintaining":[114],"high":[115],"accuracy":[116],"various":[118],"tasks.":[121],"Extensive":[122],"experiments":[123],"on":[124],"representative":[125],"MLLMs":[126],"MileBench":[129],"benchmark":[130],"demonstrate":[131],"effectiveness":[133],"compared":[136],"existing":[138],"methods.":[142]},"counts_by_year":[],"updated_date":"2026-10-16T09:59:22.235173","created_date":"2025-10-10T00:00:00"}