{"id":"https://openalex.org/W7166881899","doi":"https://doi.org/10.18653/v1/2026.acl-long.1811","title":"REAL: REtrieval-reAsoning and Logic-constructed Attention Behaviors for Long-Context KV Cache Compression","display_name":"REAL: REtrieval-reAsoning and Logic-constructed Attention Behaviors for Long-Context KV Cache Compression","publication_year":2026,"publication_date":"2026-01-01","ids":{"openalex":"https://openalex.org/W7166881899","doi":"https://doi.org/10.18653/v1/2026.acl-long.1811"},"language":null,"primary_location":{"id":"doi:10.18653/v1/2026.acl-long.1811","is_oa":true,"landing_page_url":"https://doi.org/10.18653/v1/2026.acl-long.1811","pdf_url":"https://aclanthology.org/2026.acl-long.1811.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":"Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)","raw_type":"proceedings-article"},"type":"conference-paper","indexed_in":["crossref"],"open_access":{"is_oa":true,"oa_status":"gold","oa_url":"https://aclanthology.org/2026.acl-long.1811.pdf","any_repository_has_fulltext":null},"authorships":[{"author_position":"first","author":{"id":"https://openalex.org/A5139810244","display_name":"Mengjie Li","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Mengjie Li","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5139827785","display_name":"Yuan Feng","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Yuan Feng","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5139819173","display_name":"Xike Xie","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Xike Xie","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"last","author":{"id":"https://openalex.org/A5028411609","display_name":"William J. Song","orcid":"https://orcid.org/0000-0001-9170-5986"},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"William J. Song","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":0.82934537,"is_in_top_1_percent":false,"is_in_top_10_percent":false},"cited_by_percentile_year":null,"biblio":{"volume":null,"issue":null,"first_page":"39035","last_page":"39052"},"is_retracted":false,"is_paratext":false,"is_xpac":false,"primary_topic":{"id":"https://openalex.org/T10054","display_name":"Parallel Computing and Optimization Techniques","score":0.29249998927116394,"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"}},"topics":[{"id":"https://openalex.org/T10054","display_name":"Parallel Computing and Optimization Techniques","score":0.29249998927116394,"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"}},{"id":"https://openalex.org/T11181","display_name":"Advanced Data Storage Technologies","score":0.1696999967098236,"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/T10317","display_name":"Advanced Database Systems and Queries","score":0.09210000187158585,"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"}}],"keywords":[{"id":"https://openalex.org/keywords/compression","display_name":"Compression (physics)","score":0.40070000290870667},{"id":"https://openalex.org/keywords/cache","display_name":"Cache","score":0.3718000054359436},{"id":"https://openalex.org/keywords/data-compression","display_name":"Data compression","score":0.33550000190734863},{"id":"https://openalex.org/keywords/noise","display_name":"Noise (video)","score":0.321399986743927}],"concepts":[{"id":"https://openalex.org/C41008148","wikidata":"https://www.wikidata.org/wiki/Q21198","display_name":"Computer science","level":0,"score":0.590499997138977},{"id":"https://openalex.org/C180016635","wikidata":"https://www.wikidata.org/wiki/Q2712821","display_name":"Compression (physics)","level":2,"score":0.40070000290870667},{"id":"https://openalex.org/C115537543","wikidata":"https://www.wikidata.org/wiki/Q165596","display_name":"Cache","level":2,"score":0.3718000054359436},{"id":"https://openalex.org/C24326235","wikidata":"https://www.wikidata.org/wiki/Q126095","display_name":"Electronic engineering","level":1,"score":0.35749998688697815},{"id":"https://openalex.org/C78548338","wikidata":"https://www.wikidata.org/wiki/Q2493","display_name":"Data compression","level":2,"score":0.33550000190734863},{"id":"https://openalex.org/C99498987","wikidata":"https://www.wikidata.org/wiki/Q2210247","display_name":"Noise (video)","level":3,"score":0.321399986743927},{"id":"https://openalex.org/C127413603","wikidata":"https://www.wikidata.org/wiki/Q11023","display_name":"Engineering","level":0,"score":0.2865999937057495},{"id":"https://openalex.org/C119599485","wikidata":"https://www.wikidata.org/wiki/Q43035","display_name":"Electrical engineering","level":1,"score":0.26080000400543213},{"id":"https://openalex.org/C145995521","wikidata":"https://www.wikidata.org/wiki/Q639897","display_name":"Pulse compression","level":3,"score":0.24779999256134033},{"id":"https://openalex.org/C149635348","wikidata":"https://www.wikidata.org/wiki/Q193040","display_name":"Embedded system","level":1,"score":0.24639999866485596}],"mesh":[],"locations_count":1,"locations":[{"id":"doi:10.18653/v1/2026.acl-long.1811","is_oa":true,"landing_page_url":"https://doi.org/10.18653/v1/2026.acl-long.1811","pdf_url":"https://aclanthology.org/2026.acl-long.1811.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":"Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)","raw_type":"proceedings-article"}],"best_oa_location":{"id":"doi:10.18653/v1/2026.acl-long.1811","is_oa":true,"landing_page_url":"https://doi.org/10.18653/v1/2026.acl-long.1811","pdf_url":"https://aclanthology.org/2026.acl-long.1811.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":"Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)","raw_type":"proceedings-article"},"sustainable_development_goals":[{"display_name":"Reduced inequalities","id":"https://metadata.un.org/sdg/10","score":0.4314236342906952}],"awards":[],"funders":[{"id":"https://openalex.org/F4320320671","display_name":"National Research Foundation","ror":"https://ror.org/05s0g1g46"}],"has_content":{"pdf":true,"grobid_xml":true},"content_urls":{"pdf":"https://content.openalex.org/works/W7166881899.pdf","grobid_xml":"https://content.openalex.org/works/W7166881899.grobid-xml"},"referenced_works_count":0,"referenced_works":[],"related_works":[],"abstract_inverted_index":{"The":[0,206],"growing":[1],"sequence":[2],"length":[3],"of":[4,24],"large":[5],"language":[6],"models":[7,124],"poses":[8],"significant":[9],"challenges":[10],"for":[11,52,158],"keyvalue":[12],"(KV)":[13],"caches.Existing":[14],"state-of-the-art":[15],"cache":[16,106],"eviction":[17,54],"methods":[18,166],"primarily":[19],"analyze":[20,68],"the":[21,45,57,79,108,136,156,182,192],"inference":[22],"behavior":[23,152],"attention":[25,69],"heads":[26],"in":[27,35,72,86,96,167],"successful":[28],"retrieval-reasoning":[29],"cases,":[30,37],"often":[31],"overlooking":[32],"diverse":[33],"behaviors":[34,51,71],"failure":[36,76,97],"such":[38],"as":[39],"bias":[40,93],"and":[41,75,94,102,125,188],"distraction.This":[42],"oversight":[43],"limits":[44],"potential":[46],"to":[47,66,111,135,163],"leverage":[48,112],"heterogeneous":[49],"head":[50,70],"improved":[53],"performance.Inspired":[55],"by":[56],"confusion":[58],"matrix,":[59],"we":[60,154],"introduce":[61],"an":[62],"Attention":[63],"Behavior":[64],"Matrix":[65],"comprehensively":[67],"both":[73],"success":[74,87],"scenarios.By":[77],"maximizing":[78],"signal-to-noise":[80],"ratio":[81],"-strengthening":[82],"valid":[83],"reasoning":[84],"pathways":[85],"cases":[88,98],"while":[89,140],"inhibiting":[90],"noise":[91],"from":[92,161],"distraction":[95],"-we":[99],"propose":[100],"REtrieval-reAsoning":[101],"Logic-constructed":[103],"(REAL)":[104],"KV":[105],"eviction,":[107],"first":[109],"method":[110],"multi-behavior":[113],"analysis.Comprehensive":[114],"evaluations":[115],"show":[116],"that":[117],"REAL":[118],"achieves":[119,132],"remarkable":[120],"performance":[121],"across":[122],"various":[123],"benchmarks;":[126],"notably,":[127],"on":[128,151,181],"LongBench":[129],"v2,":[130],"it":[131],"comparable":[133],"accuracy":[134],"strongest":[137],"baseline,":[138],"HeadKV-R2,":[139],"requiring":[141],"32x":[142],"less":[143],"space":[144],"(Figure":[145],"1).By":[146],"offering":[147],"a":[148,159],"novel":[149],"perspective":[150],"analysis,":[153],"pave":[155],"way":[157],"shift":[160],"success-only":[162],"comprehensive,":[164],"failureaware":[165],"long-context":[168],"modeling.Our":[169],"code":[170],"is":[171],"available":[172],"at":[173],"https://github.com/":[174],"yonseicasl/REAL.Question:":[175],"Which":[176],"operating":[177],"system":[178],"comes":[179,202,209],"installed":[180,203,210],"cheaper":[183],"device":[184],"between":[185],"black":[186],"laptop":[187],"silver":[189],"laptop?Needle:":[190],"\u2026\u2026":[191],"Silver":[193],"Laptop":[194],"costs":[195,200],"$1200":[196],"\u2026\u2026the":[197],"Black":[198],"Laptop\u2026\u2026":[199],"$800\u2026\u2026":[201],"with":[204,211],"Linux\u2026\u2026":[205],"Red":[207],"Tablet":[208],"Android\u2026\u2026":[212]},"counts_by_year":[],"updated_date":"2026-07-29T14:22:42.915294","created_date":"2026-07-02T00:00:00"}