[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"detail-sidebar-cat-1-en-105":3,"doc-seo-572640-105":53,"doc-detail-572640-en":126},{"code":4,"msg":5,"data":6},0,"success",[7,14,19,24,29,34,39,44,49],{"id":8,"doc_module":9,"doc_module_name":10,"category_name":11,"show_sort_weight":12,"slug":13},11,1,"Template","Presentations",90,"presentations",{"id":15,"doc_module":9,"doc_module_name":10,"category_name":16,"show_sort_weight":17,"slug":18},12,"Resumes",80,"resumes",{"id":20,"doc_module":9,"doc_module_name":10,"category_name":21,"show_sort_weight":22,"slug":23},14,"Invoices",70,"invoices",{"id":25,"doc_module":9,"doc_module_name":10,"category_name":26,"show_sort_weight":27,"slug":28},15,"Posters",60,"posters",{"id":30,"doc_module":9,"doc_module_name":10,"category_name":31,"show_sort_weight":32,"slug":33},16,"Social Media",50,"social-media",{"id":35,"doc_module":9,"doc_module_name":10,"category_name":36,"show_sort_weight":37,"slug":38},17,"Forms",40,"forms",{"id":40,"doc_module":9,"doc_module_name":10,"category_name":41,"show_sort_weight":42,"slug":43},18,"Letters",30,"letters",{"id":45,"doc_module":9,"doc_module_name":10,"category_name":46,"show_sort_weight":47,"slug":48},21,"Paper Templates",5,"papers-templates",{"id":50,"doc_module":9,"doc_module_name":10,"category_name":51,"show_sort_weight":4,"slug":52},158,"General","general-158",{"code":4,"msg":54,"data":55},"ok",{"site_id":56,"language":57,"slug":58,"title":59,"keywords":60,"description":61,"schema_data":62,"social_meta":119,"head_meta":121,"extra_data":123,"updated_unix":125},105,"en","cachenotes-task-aware-key-value-cache-compression-for-reasoning-intensive-knowledge-tasks","CACHENOTES - Task-Aware Key-Value Cache Compression for Reasoning-Intensive Knowledge Tasks","","Integrating external knowledge into Large Language Models (LLMs) is crucial, yet Retrieval-Augmented Generation (RAG) struggles with broad, multi-source queries and long-context inference is often computationally prohibitive. CACHENOTES addresses this by generating Compression-Planning-Tokens (CPTs) from a task description and corpus, then performing a one-time offline compression into a compact, reusable KV cache. At inference, this cache alone supports efficient reasoning-intensive question answering. On LongBench (QA), CACHENOTES exceeds RAG by 8+ F1 and cuts latency by 4×; on RULER, it surpasses query-agnostic compression by 55 points, narrowing the gap to query-aware methods.",{"@graph":63,"@context":118},[64,80,101],{"@type":65,"itemListElement":66},"BreadcrumbList",[67,71,74,77],{"item":68,"name":69,"@type":70,"position":9},"https://docshare.wps.com","Home","ListItem",{"item":72,"name":10,"@type":70,"position":73},"https://docshare.wps.com/template/",2,{"item":75,"name":11,"@type":70,"position":76},"https://docshare.wps.com/template/presentations/",3,{"item":78,"name":59,"@type":70,"position":79},"https://docshare.wps.com/template/cachenotes-task-aware-key-value-cache-compression-for-reasoning-intensive-knowledge-tasks/572640/",4,{"url":78,"name":59,"@type":81,"image":82,"author":87,"headline":59,"publisher":90,"fileFormat":93,"inLanguage":57,"description":61,"dateModified":94,"datePublished":95,"encodingFormat":93,"isAccessibleForFree":96,"interactionStatistic":97},"DigitalDocument",{"url":83,"@type":84,"width":85,"height":86},"https://docshare.wps.com/thumbnails/cachenotes-task-aware-key-value-cache-compression-for-reasoning-intensive-knowledge-tasks/572640.png","ImageObject",442,249,{"name":88,"@type":89},"cksankcs","Person",{"url":68,"name":91,"@type":92},"DocShare","Organization","application/pdf","2026-10-21","2026-10-13",true,{"@type":98,"interactionType":99,"userInteractionCount":47},"InteractionCounter",{"@type":100},"ViewAction",{"@type":102,"mainEntity":103},"FAQPage",[104,110,114],{"name":105,"@type":106,"acceptedAnswer":107},"What problem does CACHENOTES target in LLM knowledge tasks?","Question",{"text":108,"@type":109},"It targets the limitations of RAG on broad multi-source queries and the high computational cost of full-context inference for reasoning-intensive tasks over large corpora.","Answer",{"name":111,"@type":106,"acceptedAnswer":112},"How does CACHENOTES generate and use Compression-Planning-Tokens (CPTs)?",{"text":113,"@type":109},"Given a task description and a corpus, it first generates CPTs via an offline distillation pass, then uses CPTs to guide a one-time compression of the corpus into a compact KV cache.",{"name":115,"@type":106,"acceptedAnswer":116},"What performance gains does CACHENOTES show compared with RAG and prior compression methods?",{"text":117,"@type":109},"On LongBench QA, it outperforms RAG by over 8 F1 points and reduces latency by over 4×. On RULER, it exceeds previous query-agnostic compression by 55 points, narrowing the gap to query-aware compression approaches.","https://schema.org",{"og:url":78,"og:type":120,"og:title":59,"og:site_name":91,"og:description":61},"article",{"robots":122,"canonical":78},"index,follow",{"doc_id":124,"site_id":56},572640,1792294979,{"code":4,"msg":5,"data":127},{"doc_id":124,"user_id":128,"nickname":88,"user_avatar":129,"doc_module":9,"category_id":8,"category_name":11,"doc_title":59,"doc_description":61,"doc_content":130,"file_id":131,"file_url":132,"file_type":133,"file_size":134,"view_count":47,"is_deleted":4,"is_public":9,"is_downloadable":9,"audit_status":9,"page_count":135,"language":136,"language_code":57,"site_id":56,"html_lang":57,"table_of_contents":137,"faqs":138,"seo_title":139,"seo_description":61,"update_tm":140,"read_time":141},3985747882755,"https://ap-avatar.wpscdn.com/davatar_29158cc5080c5b710cf443261637dec0","CACHENOTES: Task-Aware Key-Value Cache Compression for Reasoning-Intensive Knowledge Tasks  \nGiulio Corallo1,2 Orion Weller3 Fabio Petroni4 Paolo Papotti2  \n1 SAP Labs, France 2EURECOM, France  \n3Johns Hopkins University, USA 4European Molecular Biology Laboratory, Italy  \nCorrespondence: [giulio.corallo@sap.com](giulio.corallo@sap.com)  \nAbstract  \nIntegrating external knowledge into Large Language Models (LLMs) is crucial for many realworld applications, yet current methods like Retrieval-Augmented Generation (RAG) face limitations with broad, multi-source queries, while long-context models are computationally prohibitive.  \nWe introduce CACHENOTES: Task-Aware KeyValue Cache Compression. Given a task description and a corpus, CACHENOTES first generates a sequence of Compression-PlanningTokens (CPTs), an offline task-focused distillation pass that identifies and organizes key information from the corpus. These CPTs are then used to guide a one-time compression of the corpus into a compact, reusable KV cache, which is then used alone at inference time to efficiently answer diverse, reasoning-intensive queries, eliminating repeated retrieval or context expansion.  \nExperiments on LongBench show that, on Question-Answering tasks at a 20 × compression, CACHENOTES outperforms RAG by over  \n8 F1 points and reduces latency by over 4 × . On RULER, it surpasses previous query-agnostic compression methods by 55 points, narrowing the gap to query-aware compression approaches. Additional results on real-world enterprise and synthetic datasets demonstrate its strong performance on multi-hop and broadcoverage queries.  \n1 Introduction  \nIncorporating external information into Large Language Models (LLMs) significantly enhances their utility across applications, enabling more informed and accurate outputs (Petroni et al., 2019) . Retrieval-Augmented Generation (RAG) (Lewis et al., 2020) effectively injects relevant evidence for narrow, well-scoped queries, but it struggles with questions that require synthesizing information spread across multiple documents or sections  \nof text (Barnett et al., 2024) . Such multi-hop or broad queries are hindered by retrieval’s reliance on local similarity, which fails to capture global relationships or latent links within the corpus, making it difficult to surface all relevant context (Weller et al., 2025) and often introducing noise or redundancy (Yu et al., 2024) .  \nRecent advances have extended the context length that LLMs can process (Reid et al., 2024 ; Liet al., 2025a), making it possible to feed the entire corpus as input and enabling more holistic reasoning. However, this approach comes with significant computational costs, as handling large inputs requires substantial memory resources, particularly on GPUs, creating a scalability bottleneck (Liu et al., 2023) . Furthermore, as context length grows, models often struggle to identify and exploit the truly relevant information buried within extensive text (Liu et al., 2024 ; Laban et al., 2025) . This leads to a fundamental challenge:  \nHow can LLMs efficiently and accurately perform reasoning-intensive tasks that require broad access to large external corpora, without the prohibitive costs of full-context inference or the limitations of retrieval-based methods?  \nWe address this challenge by introducing CACHENOTES: Task-Aware Key-Value Cache Compression, a query-agnostic yet task-guided framework that converts an entire corpus into areusable, compact Key-Value (KV) cache. Unlike prior query-aware compression, which recomputes a compressed KV cache for each query (Liet al., 2025b ; Rehg, 2024 ; Corallo and Papotti, 2024 ; Xu et al., 2025), or generic, prompt-based compression approaches (Kim et al., 2024), our method performs a one-time, offline compression guided by a human-written task description and a model-generated sequence of CompressionPlanning-Tokens (CPTs) . These CPTs distill key  \n6571  \nProceedings of the 19th Conference of the Europea","cbCaiktcwi8ab9vt","https://ap.wps.com/l/cbCaiktcwi8ab9vt","pdf",1233711,20,"English","# Abstract\n# Introduction\n# CPTs Generation\n# KV Compression\n# Downstream Query Answering\n# Experimental Results","[{\"question\":\"What problem does CACHENOTES target in LLM knowledge tasks?\",\"answer\":\"It targets the limitations of RAG on broad multi-source queries and the high computational cost of full-context inference for reasoning-intensive tasks over large corpora.\"},{\"question\":\"How does CACHENOTES generate and use Compression-Planning-Tokens (CPTs)?\",\"answer\":\"Given a task description and a corpus, it first generates CPTs via an offline distillation pass, then uses CPTs to guide a one-time compression of the corpus into a compact KV cache.\"},{\"question\":\"What performance gains does CACHENOTES show compared with RAG and prior compression methods?\",\"answer\":\"On LongBench QA, it outperforms RAG by over 8 F1 points and reduces latency by over 4×. On RULER, it exceeds previous query-agnostic compression by 55 points, narrowing the gap to query-aware compression approaches.\"}]","CACHENOTES - Task-Aware Key-Value Cache Compression for Reasoning-Intensive Knowledge Tasks | PDF",1791928583,7]