{"id":"https://openalex.org/W7166851427","doi":"https://doi.org/10.18653/v1/2026.findings-acl.494","title":"OjaKV: Context-Aware Online Low-Rank KV Cache Compression","display_name":"OjaKV: Context-Aware Online Low-Rank KV Cache Compression","publication_year":2026,"publication_date":"2026-01-01","ids":{"openalex":"https://openalex.org/W7166851427","doi":"https://doi.org/10.18653/v1/2026.findings-acl.494"},"language":null,"primary_location":{"id":"doi:10.18653/v1/2026.findings-acl.494","is_oa":true,"landing_page_url":"https://doi.org/10.18653/v1/2026.findings-acl.494","pdf_url":"https://aclanthology.org/2026.findings-acl.494.pdf","source":null,"license":"cc-by","license_id":"https://openalex.org/licenses/cc-by","version":"publishedVersion","is_accepted":true,"is_published":true,"raw_source_name":"Findings of the Association for Computational Linguistics: ACL 2026","raw_type":"proceedings-article"},"type":"conference-paper","indexed_in":["crossref"],"open_access":{"is_oa":true,"oa_status":"gold","oa_url":"https://aclanthology.org/2026.findings-acl.494.pdf","any_repository_has_fulltext":null},"authorships":[{"author_position":"first","author":{"id":"https://openalex.org/A5060302276","display_name":"Yuxuan Zhu","orcid":"https://orcid.org/0000-0001-6686-2736"},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Yuxuan Zhu","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5139812470","display_name":"David H. Yang","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"David H. Yang","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5010987707","display_name":"Mohammad Mohammadi Amiri","orcid":"https://orcid.org/0000-0002-7343-6628"},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Mohammad Mohammadi Amiri","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5139805749","display_name":"Keerthiram Murugesan","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Keerthiram Murugesan","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5066128057","display_name":"Tejaswini Pedapati","orcid":"https://orcid.org/0000-0002-5260-0951"},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Tejaswini Pedapati","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"last","author":{"id":"https://openalex.org/A5050344371","display_name":"Pin\u2010Yu Chen","orcid":"https://orcid.org/0000-0003-1039-8369"},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Pin-Yu Chen","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":0.0,"has_fulltext":true,"cited_by_count":0,"citation_normalized_percentile":{"value":0.85771522,"is_in_top_1_percent":false,"is_in_top_10_percent":false},"cited_by_percentile_year":null,"biblio":{"volume":null,"issue":null,"first_page":"10161","last_page":"10178"},"is_retracted":false,"is_paratext":false,"is_xpac":false,"primary_topic":{"id":"https://openalex.org/T11478","display_name":"Caching and Content Delivery","score":0.4945000112056732,"subfield":{"id":"https://openalex.org/subfields/1705","display_name":"Computer Networks and Communications"},"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/T11478","display_name":"Caching and Content Delivery","score":0.4945000112056732,"subfield":{"id":"https://openalex.org/subfields/1705","display_name":"Computer Networks and Communications"},"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/T10138","display_name":"Network Traffic and Congestion Control","score":0.09960000216960907,"subfield":{"id":"https://openalex.org/subfields/1705","display_name":"Computer Networks and Communications"},"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/T10054","display_name":"Parallel Computing and Optimization Techniques","score":0.05310000106692314,"subfield":{"id":"https://openalex.org/subfields/1708","display_name":"Hardware and Architecture"},"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/compression","display_name":"Compression (physics)","score":0.34599998593330383},{"id":"https://openalex.org/keywords/data-compression","display_name":"Data compression","score":0.31619998812675476},{"id":"https://openalex.org/keywords/cache","display_name":"Cache","score":0.29440000653266907},{"id":"https://openalex.org/keywords/noise","display_name":"Noise (video)","score":0.2597000002861023},{"id":"https://openalex.org/keywords/set","display_name":"Set (abstract data type)","score":0.24779999256134033}],"concepts":[{"id":"https://openalex.org/C41008148","wikidata":"https://www.wikidata.org/wiki/Q21198","display_name":"Computer science","level":0,"score":0.6313999891281128},{"id":"https://openalex.org/C180016635","wikidata":"https://www.wikidata.org/wiki/Q2712821","display_name":"Compression (physics)","level":2,"score":0.34599998593330383},{"id":"https://openalex.org/C78548338","wikidata":"https://www.wikidata.org/wiki/Q2493","display_name":"Data compression","level":2,"score":0.31619998812675476},{"id":"https://openalex.org/C115537543","wikidata":"https://www.wikidata.org/wiki/Q165596","display_name":"Cache","level":2,"score":0.29440000653266907},{"id":"https://openalex.org/C9390403","wikidata":"https://www.wikidata.org/wiki/Q3966","display_name":"Computer hardware","level":1,"score":0.2628999948501587},{"id":"https://openalex.org/C99498987","wikidata":"https://www.wikidata.org/wiki/Q2210247","display_name":"Noise (video)","level":3,"score":0.2597000002861023},{"id":"https://openalex.org/C177264268","wikidata":"https://www.wikidata.org/wiki/Q1514741","display_name":"Set (abstract data type)","level":2,"score":0.24779999256134033},{"id":"https://openalex.org/C43521106","wikidata":"https://www.wikidata.org/wiki/Q2165493","display_name":"Pipeline (software)","level":2,"score":0.24070000648498535},{"id":"https://openalex.org/C127413603","wikidata":"https://www.wikidata.org/wiki/Q11023","display_name":"Engineering","level":0,"score":0.23770000040531158},{"id":"https://openalex.org/C149635348","wikidata":"https://www.wikidata.org/wiki/Q193040","display_name":"Embedded system","level":1,"score":0.22949999570846558}],"mesh":[],"locations_count":1,"locations":[{"id":"doi:10.18653/v1/2026.findings-acl.494","is_oa":true,"landing_page_url":"https://doi.org/10.18653/v1/2026.findings-acl.494","pdf_url":"https://aclanthology.org/2026.findings-acl.494.pdf","source":null,"license":"cc-by","license_id":"https://openalex.org/licenses/cc-by","version":"publishedVersion","is_accepted":true,"is_published":true,"raw_source_name":"Findings of the Association for Computational Linguistics: ACL 2026","raw_type":"proceedings-article"}],"best_oa_location":{"id":"doi:10.18653/v1/2026.findings-acl.494","is_oa":true,"landing_page_url":"https://doi.org/10.18653/v1/2026.findings-acl.494","pdf_url":"https://aclanthology.org/2026.findings-acl.494.pdf","source":null,"license":"cc-by","license_id":"https://openalex.org/licenses/cc-by","version":"publishedVersion","is_accepted":true,"is_published":true,"raw_source_name":"Findings of the Association for Computational Linguistics: ACL 2026","raw_type":"proceedings-article"},"sustainable_development_goals":[],"awards":[],"funders":[],"has_content":{"grobid_xml":true,"pdf":true},"content_urls":{"pdf":"https://content.openalex.org/works/W7166851427.pdf","grobid_xml":"https://content.openalex.org/works/W7166851427.grobid-xml"},"referenced_works_count":0,"referenced_works":[],"related_works":[],"abstract_inverted_index":{"The":[0],"expanding":[1],"long-context":[2,198,239],"capabilities":[3],"of":[4,38,124,207],"large":[5],"language":[6],"models":[7],"are":[8,104],"constrained":[9],"by":[10,131],"a":[11,28,31,35,48,60,67,84,89,147,233],"significant":[12],"memory":[13,225],"bottleneck:":[14],"the":[15,51,111,122,134,159,164,205],"key-value":[16],"(KV)":[17],"cache":[18],"required":[19],"for":[20,26,44,107,119,121,140,237],"autoregressive":[21],"generation.This":[22],"bottleneck":[23],"is":[24,59,169,218],"substantial;":[25],"instance,":[27],"Llama-3.1-8Bmodel":[29],"processing":[30],"32K-token":[32],"prompt":[33],"at":[34,186],"batch":[36],"size":[37,49],"4":[39],"requires":[40],"approximately":[41],"16":[42],"GB":[43],"its":[45,193],"KV":[46],"cache,":[47],"exceeding":[50],"model's":[52],"weights.While":[53],"KV-cache":[54],"compression":[55,130,188],"via":[56],"low-rank":[57,129],"projection":[58,135],"promising":[61],"direction,":[62],"existing":[63],"methods'":[64],"rely":[65],"on":[66,196],"static,":[68],"offline-learned":[69],"subspace":[70,96,160,209],"that":[71,87,100,179,200],"performs":[72],"poorly":[73],"under":[74],"data":[75],"distribution":[76],"shifts.To":[77],"overcome":[78],"these":[79],"limitations,":[80],"we":[81],"introduce":[82],"OjaKV,":[83],"novel":[85],"framework":[86,168,231],"integrates":[88],"strategic":[90],"hybrid":[91,230],"storage":[92],"policy":[93],"with":[94,163,172,220],"online":[95,141,208],"adaptation.First,":[97],"OjaKV":[98,180,191],"recognizes":[99],"not":[101],"all":[102],"tokens":[103,113],"equally":[105],"important":[106],"compression;":[108],"it":[109,127],"preserves":[110],"crucial":[112],"in":[114,211],"full-rank,":[115],"maintaining":[116],"high-fidelity":[117],"anchors":[118],"attention.Second,":[120],"majority":[123],"intermediate":[125],"tokens,":[126],"applies":[128],"incrementally":[132],"adapting":[133],"basis":[136],"using":[137],"Oja's":[138],"algorithm":[139],"principal":[142],"component":[143],"analysis.This":[144],"adaptation":[145,210],"involves":[146],"comprehensive":[148],"update":[149],"during":[150,156],"prefilling":[151],"and":[152],"lightweight":[153],"periodic":[154],"updates":[155],"decoding,":[157],"ensuring":[158],"remains":[161],"aligned":[162],"evolving":[165],"context.Crucially,":[166],"our":[167,216,229],"fully":[170],"compatible":[171,219],"modern":[173],"attention":[174],"modules":[175],"like":[176],"FlashAttention.Experiments":[177],"show":[178],"maintains":[181],"or":[182],"even":[183],"improves":[184],"accuracy":[185],"high":[187],"ratios.In":[189],"particular,":[190],"achieves":[192],"strongest":[194],"gains":[195],"very":[197],"benchmarks":[199],"require":[201],"complex":[202],"reasoning,":[203],"highlighting":[204],"importance":[206],"dynamically":[212],"tracking":[213],"context":[214],"shifts.Furthermore,":[215],"approach":[217],"token-selection":[221],"methods,":[222],"enabling":[223],"compounded":[224],"savings.These":[226],"results":[227],"establish":[228],"as":[232],"practical,":[234],"plug-and-play":[235],"solution":[236],"memoryefficient":[238],"inference":[240],"without":[241],"requir-*":[242],"Equal":[243],"contribution.ing":[244],"model":[245],"fine-tuning.":[246]},"counts_by_year":[],"updated_date":"2026-07-29T14:22:42.915294","created_date":"2026-07-02T00:00:00"}