{"id":"https://openalex.org/W4400104293","doi":"https://doi.org/10.48550/arxiv.2406.17808","title":"Training-Free Exponential Context Extension via Cascading KV Cache","display_name":"Training-Free Exponential Context Extension via Cascading KV Cache","publication_year":2024,"publication_date":"2024-06-24","ids":{"openalex":"https://openalex.org/W4400104293","doi":"https://doi.org/10.48550/arxiv.2406.17808"},"language":"en","primary_location":{"id":"pmh:oai:arXiv.org:2406.17808","doi":null,"is_oa":true,"landing_page_url":"http://arxiv.org/abs/2406.17808","pdf_url":"https://arxiv.org/pdf/2406.17808","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-nc-nd","license_id":"https://openalex.org/licenses/cc-by-nc-nd","version":"submittedVersion","is_accepted":false,"is_published":false,"raw_source_name":null,"raw_type":"text"},"parent_works":[],"lineage":["https://openalex.org/W4400104293"],"child_works_count":0,"type":"preprint","is_traditional":true,"indexed_in":["arxiv","datacite"],"open_access":{"is_oa":true,"oa_status":"green","oa_url":"https://arxiv.org/pdf/2406.17808","any_repository_has_fulltext":true},"authorships":[{"author_position":"first","author":{"id":"https://openalex.org/A5055349429","display_name":"Jeffrey Willette","orcid":null,"observed_orcids":[]},"institutions":[{"id":"https://openalex.org/I157485424","display_name":"Korea Advanced Institute of Science and Technology","ror":"https://ror.org/05apxxy63","country_code":"KR","type":"education","lineage":["https://openalex.org/I157485424"]}],"countries":["KR"],"is_corresponding":false,"raw_author_name":"Willette, Jeffrey","raw_affiliation_strings":["KAIST AI ,"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"KAIST AI ,","institution_ids":["https://openalex.org/I157485424"]}]},{"author_position":"middle","author":{"id":null,"display_name":"Lee, Heejun","orcid":null,"observed_orcids":[]},"institutions":[{"id":"https://openalex.org/I142401562","display_name":"Electronics and Telecommunications Research Institute","ror":"https://ror.org/03ysstz10","country_code":"KR","type":"facility","lineage":["https://openalex.org/I142401562","https://openalex.org/I2801339556","https://openalex.org/I4210144908","https://openalex.org/I4387152098"]},{"id":"https://openalex.org/I157485424","display_name":"Korea Advanced Institute of Science and Technology","ror":"https://ror.org/05apxxy63","country_code":"KR","type":"education","lineage":["https://openalex.org/I157485424"]}],"countries":["KR"],"is_corresponding":false,"raw_author_name":"Lee, Heejun","raw_affiliation_strings":["ETRI , South Korea","KAIST AI ,"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"ETRI , South Korea","institution_ids":["https://openalex.org/I142401562"]},{"raw_affiliation_string":"KAIST AI ,","institution_ids":["https://openalex.org/I157485424"]}]},{"author_position":"middle","author":{"id":"https://openalex.org/A5020344753","display_name":"Youngwan Lee","orcid":"https://orcid.org/0009-0009-5145-8314","observed_orcids":["https://orcid.org/0009-0009-5145-8314","https://orcid.org/0000-0001-8644-155X"]},"institutions":[{"id":"https://openalex.org/I157485424","display_name":"Korea Advanced Institute of Science and Technology","ror":"https://ror.org/05apxxy63","country_code":"KR","type":"education","lineage":["https://openalex.org/I157485424"]}],"countries":["KR"],"is_corresponding":false,"raw_author_name":"Lee, Youngwan","raw_affiliation_strings":["KAIST AI , Deepauto.ai , South Korea"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"KAIST AI , Deepauto.ai , South Korea","institution_ids":["https://openalex.org/I157485424"]}]},{"author_position":"middle","author":{"id":"https://openalex.org/A5079046832","display_name":"Myeongjae Jeon","orcid":"https://orcid.org/0000-0002-0748-6627","observed_orcids":["https://orcid.org/0000-0002-0748-6627"]},"institutions":[{"id":"https://openalex.org/I123900574","display_name":"Pohang University of Science and Technology","ror":"https://ror.org/04xysgw12","country_code":"KR","type":"education","lineage":["https://openalex.org/I123900574"]}],"countries":["KR"],"is_corresponding":false,"raw_author_name":"Jeon, Myeongjae","raw_affiliation_strings":["POSTECH , South Korea"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"POSTECH , South Korea","institution_ids":["https://openalex.org/I123900574"]}]},{"author_position":"last","author":{"id":"https://openalex.org/A5070302452","display_name":"Sung Ju Hwang","orcid":"https://orcid.org/0000-0002-9675-2324","observed_orcids":["https://orcid.org/0000-0002-9675-2324"]},"institutions":[{"id":"https://openalex.org/I157485424","display_name":"Korea Advanced Institute of Science and Technology","ror":"https://ror.org/05apxxy63","country_code":"KR","type":"education","lineage":["https://openalex.org/I157485424"]}],"countries":["KR"],"is_corresponding":false,"raw_author_name":"Hwang, Sung Ju","raw_affiliation_strings":["KAIST AI , Deepauto.ai , South Korea"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"KAIST AI , Deepauto.ai , South Korea","institution_ids":["https://openalex.org/I157485424"]}]}],"institutions":[],"countries_distinct_count":1,"institutions_distinct_count":3,"corresponding_author_ids":[],"corresponding_institution_ids":[],"apc_list":null,"apc_paid":null,"fwci":null,"has_fulltext":true,"cited_by_count":7,"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/T10028","display_name":"Topic Modeling","score":0.6840999722480774,"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"}},"topics":[{"id":"https://openalex.org/T10028","display_name":"Topic Modeling","score":0.6840999722480774,"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":0.2833999991416931,"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/T10036","display_name":"Advanced Neural Network Applications","score":0.009700000286102295,"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/key-value-caching","display_name":"key-value caching","score":0.753000020980835},{"id":"https://openalex.org/keywords/long-context-inference","display_name":"long-context inference","score":0.5759999752044678},{"id":"https://openalex.org/keywords/large-language-models","display_name":"large language models","score":0.7639999985694885},{"id":"https://openalex.org/keywords/conditional-generation","display_name":"conditional generation","score":0.45500001311302185},{"id":"https://openalex.org/keywords/few-shot-learning","display_name":"few-shot learning","score":0.5929999947547913},{"id":"https://openalex.org/keywords/inference-efficiency","display_name":"inference efficiency","score":0.3400000035762787}],"concepts":[{"id":"https://openalex.org/C2778029271","wikidata":"https://www.wikidata.org/wiki/Q5421931","display_name":"Extension (predicate logic)","level":2,"score":0.7288649678230286},{"id":"https://openalex.org/C115537543","wikidata":"https://www.wikidata.org/wiki/Q165596","display_name":"Cache","level":2,"score":0.717319667339325},{"id":"https://openalex.org/C102392041","wikidata":"https://www.wikidata.org/wiki/Q592860","display_name":"Sliding window protocol","level":3,"score":0.6166450381278992},{"id":"https://openalex.org/C2779343474","wikidata":"https://www.wikidata.org/wiki/Q3109175","display_name":"Context (archaeology)","level":2,"score":0.5522659420967102},{"id":"https://openalex.org/C41008148","wikidata":"https://www.wikidata.org/wiki/Q21198","display_name":"Computer science","level":0,"score":0.547200620174408},{"id":"https://openalex.org/C151376022","wikidata":"https://www.wikidata.org/wiki/Q168698","display_name":"Exponential function","level":2,"score":0.5018386840820312},{"id":"https://openalex.org/C2778751112","wikidata":"https://www.wikidata.org/wiki/Q835016","display_name":"Window (computing)","level":2,"score":0.48243841528892517},{"id":"https://openalex.org/C2777211547","wikidata":"https://www.wikidata.org/wiki/Q17141490","display_name":"Training (meteorology)","level":2,"score":0.4551226794719696},{"id":"https://openalex.org/C173608175","wikidata":"https://www.wikidata.org/wiki/Q232661","display_name":"Parallel computing","level":1,"score":0.4001126289367676},{"id":"https://openalex.org/C11413529","wikidata":"https://www.wikidata.org/wiki/Q8366","display_name":"Algorithm","level":1,"score":0.32402920722961426},{"id":"https://openalex.org/C33923547","wikidata":"https://www.wikidata.org/wiki/Q395","display_name":"Mathematics","level":0,"score":0.2811654210090637},{"id":"https://openalex.org/C111919701","wikidata":"https://www.wikidata.org/wiki/Q9135","display_name":"Operating system","level":1,"score":0.12943655252456665},{"id":"https://openalex.org/C127313418","wikidata":"https://www.wikidata.org/wiki/Q1069","display_name":"Geology","level":0,"score":0.12220826745033264},{"id":"https://openalex.org/C199360897","wikidata":"https://www.wikidata.org/wiki/Q9143","display_name":"Programming language","level":1,"score":0.0961022675037384},{"id":"https://openalex.org/C134306372","wikidata":"https://www.wikidata.org/wiki/Q7754","display_name":"Mathematical analysis","level":1,"score":0.09405088424682617},{"id":"https://openalex.org/C121332964","wikidata":"https://www.wikidata.org/wiki/Q413","display_name":"Physics","level":0,"score":0.07906228303909302},{"id":"https://openalex.org/C153294291","wikidata":"https://www.wikidata.org/wiki/Q25261","display_name":"Meteorology","level":1,"score":0.0},{"id":"https://openalex.org/C151730666","wikidata":"https://www.wikidata.org/wiki/Q7205","display_name":"Paleontology","level":1,"score":0.0}],"mesh":[],"locations_count":2,"locations":[{"id":"pmh:oai:arXiv.org:2406.17808","doi":null,"is_oa":true,"landing_page_url":"http://arxiv.org/abs/2406.17808","pdf_url":"https://arxiv.org/pdf/2406.17808","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-nc-nd","license_id":"https://openalex.org/licenses/cc-by-nc-nd","version":"submittedVersion","is_accepted":false,"is_published":false,"raw_source_name":null,"raw_type":"text"},{"id":"doi:10.48550/arxiv.2406.17808","doi":"https://doi.org/10.48550/arxiv.2406.17808","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2406.17808","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-nc-nd","license_id":"https://openalex.org/licenses/cc-by-nc-nd","version":null,"is_accepted":false,"is_published":false,"raw_source_name":null,"raw_type":"Preprint"}],"best_oa_location":{"id":"pmh:oai:arXiv.org:2406.17808","doi":null,"is_oa":true,"landing_page_url":"http://arxiv.org/abs/2406.17808","pdf_url":"https://arxiv.org/pdf/2406.17808","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-nc-nd","license_id":"https://openalex.org/licenses/cc-by-nc-nd","version":"submittedVersion","is_accepted":false,"is_published":false,"raw_source_name":null,"raw_type":"text"},"sustainable_development_goals":[],"study_designs":[],"awards":[{"id":"https://openalex.org/G1398259286","display_name":null,"funder_award_id":"RS-2019II191906","funder_id":"https://openalex.org/F4320328359","funder_display_name":"Ministry of Science and ICT, South Korea"},{"id":"https://openalex.org/G3736587842","display_name":null,"funder_award_id":"RS-2019-II190075","funder_id":"https://openalex.org/F4320335489","funder_display_name":"Institute for Information and Communications Technology Promotion"},{"id":"https://openalex.org/G4468082379","display_name":null,"funder_award_id":"RS-2019-II190075","funder_id":"https://openalex.org/F4320328359","funder_display_name":"Ministry of Science and ICT, South Korea"},{"id":"https://openalex.org/G4700831490","display_name":null,"funder_award_id":"2022-","funder_id":"https://openalex.org/F4320335489","funder_display_name":"Institute for Information and Communications Technology Promotion"},{"id":"https://openalex.org/G5051746381","display_name":null,"funder_award_id":"2019-II191906","funder_id":"https://openalex.org/F4320328359","funder_display_name":"Ministry of Science and ICT, South Korea"},{"id":"https://openalex.org/G7051958709","display_name":null,"funder_award_id":"RS-2022-00187238","funder_id":"https://openalex.org/F4320322120","funder_display_name":"National Research Foundation of Korea"},{"id":"https://openalex.org/G7294048361","display_name":null,"funder_award_id":"RS-2019-II191906","funder_id":"https://openalex.org/F4320335489","funder_display_name":"Institute for Information and Communications Technology Promotion"}],"funders":[{"id":"https://openalex.org/F4320320671","display_name":"National Research Foundation","ror":"https://ror.org/05s0g1g46"},{"id":"https://openalex.org/F4320321282","display_name":"Pohang University of Science and Technology","ror":"https://ror.org/04xysgw12"},{"id":"https://openalex.org/F4320322120","display_name":"National Research Foundation of Korea","ror":"https://ror.org/013aysd81"},{"id":"https://openalex.org/F4320324161","display_name":"Korea Advanced Institute of Science and Technology","ror":"https://ror.org/05apxxy63"},{"id":"https://openalex.org/F4320328359","display_name":"Ministry of Science and ICT, South Korea","ror":"https://ror.org/01wpjm123"},{"id":"https://openalex.org/F4320335489","display_name":"Institute for Information and Communications Technology Promotion","ror":"https://ror.org/01g0hqq23"}],"has_content":{"pdf":true,"grobid_xml":true},"content_urls":{"pdf":"https://content.openalex.org/works/W4400104293.pdf","grobid_xml":"https://content.openalex.org/works/W4400104293.grobid-xml"},"referenced_works_count":18,"referenced_works":["https://openalex.org/W2986922898","https://openalex.org/W2994673210","https://openalex.org/W3015468748","https://openalex.org/W3091156754","https://openalex.org/W4378505330","https://openalex.org/W4384648639","https://openalex.org/W4387225871","https://openalex.org/W4387294617","https://openalex.org/W4387356026","https://openalex.org/W4387561528","https://openalex.org/W4389471205","https://openalex.org/W4389518760","https://openalex.org/W4394737116","https://openalex.org/W4399414358","https://openalex.org/W4401306886","https://openalex.org/W4402671766","https://openalex.org/W4415797534","https://openalex.org/W4415798413"],"related_works":["https://openalex.org/W7166877006","https://openalex.org/W4407806278","https://openalex.org/W6910722644","https://openalex.org/W7134837349","https://openalex.org/W6891900780","https://openalex.org/W7135011577","https://openalex.org/W4405029395","https://openalex.org/W7155398280","https://openalex.org/W4417287129","https://openalex.org/W4414762093"],"abstract_inverted_index":{"The":[0],"transformer's":[1],"context":[2,26,90,117],"window":[3],"is":[4],"vital":[5],"for":[6,20,88,196],"tasks":[7],"such":[8],"as":[9,15,24],"few-shot":[10],"learning":[11],"and":[12,68,140],"conditional":[13],"generation":[14],"it":[16,144],"preserves":[17],"previous":[18],"tokens":[19,67,151],"active":[21],"memory.":[22],"However,":[23],"the":[25,29,35,62,107,112,121,156,186,194],"lengths":[27],"increase,":[28],"computational":[30,187],"costs":[31],"grow":[32],"quadratically,":[33],"hindering":[34],"deployment":[36,199],"of":[37,155,159,171,189],"large":[38],"language":[39],"models":[40],"(LLMs)":[41],"in":[42,81,200],"real-world,":[43],"long":[44],"sequence":[45],"scenarios.":[46],"Although":[47],"some":[48],"recent":[49],"key-value":[50],"caching":[51,128],"(KV":[52],"Cache)":[53],"methods":[54],"offer":[55],"linear":[56,127],"inference":[57],"complexity,":[58],"they":[59,73],"naively":[60],"manage":[61],"stored":[63],"context,":[64],"prematurely":[65],"evicting":[66],"losing":[69],"valuable":[70],"information.":[71],"Moreover,":[72],"lack":[74],"an":[75],"optimized":[76],"prefill/prompt":[77],"stage":[78,166],"strategy,":[79],"resulting":[80],"higher":[82],"latency":[83,167],"than":[84],"even":[85],"quadratic":[86],"attention":[87,177],"realistic":[89],"sizes.":[91],"In":[92],"response,":[93],"we":[94],"introduce":[95],"a":[96,169],"novel":[97],"mechanism":[98],"that":[99],"leverages":[100],"cascading":[101],"sub-cache":[102],"buffers":[103],"to":[104,114,175],"selectively":[105],"retain":[106],"most":[108],"relevant":[109],"tokens,":[110],"enabling":[111,203],"model":[113],"maintain":[115],"longer":[116],"histories":[118],"without":[119],"increasing":[120],"cache":[122,157],"size.":[123],"Our":[124],"approach":[125],"outperforms":[126],"baselines":[129],"across":[130],"key":[131],"benchmarks,":[132],"including":[133],"streaming":[134],"perplexity,":[135],"question":[136],"answering,":[137],"book":[138],"summarization,":[139],"passkey":[141],"retrieval,":[142],"where":[143],"retains":[145],"better":[146],"retrieval":[147],"accuracy":[148],"at":[149],"1M":[150,179],"after":[152],"four":[153],"doublings":[154],"size":[158],"65K.":[160],"Additionally,":[161],"our":[162],"method":[163],"reduces":[164],"prefill":[165],"by":[168],"factor":[170],"6.8":[172],"when":[173],"compared":[174],"flash":[176],"on":[178],"tokens.":[180],"These":[181],"innovations":[182],"not":[183],"only":[184],"enhance":[185],"efficiency":[188],"LLMs":[190],"but":[191],"also":[192],"pave":[193],"way":[195],"their":[197],"effective":[198],"resource-constrained":[201],"environments,":[202],"large-scale,":[204],"real-time":[205],"applications":[206],"with":[207],"significantly":[208],"reduced":[209],"latency.":[210]},"counts_by_year":[{"year":2026,"cited_by_count":3},{"year":2025,"cited_by_count":4}],"updated_date":"2026-11-14T11:04:58.165337","created_date":"2024-06-28T00:00:00"}