AI Paper Atlas

Research map / Architectures and efficiency / Attention and transformers

ProtoKV: Long-context Knowledges Are Already Well-Organized Before Your Query

Zhiyuan Yu, Shijian Xiao, Zhangyue Yin, Xiaoran Liu, Lekai Xing, Wenzhong Li, Nguyen Cam-Tu, Sanglu Lu

ICLR 2026 · Architectures and efficiency › Attention and transformers · approach: KV cache eviction and compression

Read the paperCheck your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
retain KV pairs by outlier degree only (SATKV) instead of SnapKV selectionLlama3-8B KV retention · NarrativeQA (budget 256)score: 23.32 -> 21.12Hurt
retain KV pairs by outlier degree only (SATKV) instead of SnapKV selectionLlama3-8B KV retention · LCC (budget 256)score: 60.27 -> 58.46Hurt
add semantic-prototype clustering (ProtoKV) over SnapKVLlama3-8B · RULER NIAH (budget 128)accuracy: 97.2 -> 98.2Improved
add semantic-prototype clustering (ProtoKV) over ChunkKVLlama3-8B · RULER QA (budget 512)accuracy: 42.2 -> 43.6Improved
add prototype clustering (ProtoKV) over SnapKVLlama3-8B · RULER QA (budget 512)accuracy: 43.8 -> 43.6No clear effect
use ProtoKV instead of K-means cluster methods (SqueezeAtt) with decoding-stage recallLlama3-8B · LongBench (budget 512)average score: 41.47 -> 42.59Improved
use ProtoKV instead of SqueezeAtt with prefilling-only retentionLlama3-8B · LongBench (budget 512)average score: 40.42 -> 40.76Improved
ProtoKV instead of SnapKV on Needle In A HaystackLlama3-8B-Instruct · Needle In A Haystack (budget 128)retrieval accuracy: 94.2% -> 97.3%Improved
combine ProtoKV with layer-wise budget allocationMistral-7B-Instruct · LongBench SDQA (budget 256)score: 32.87 -> 33.12Improved
combine ProtoKV with head-wise budget allocationMistral-7B-Instruct · LongBench Code (budget 256)score: 59.38 -> 60.16Improved
replace hybrid prototypes with chunk-only aggregation (remove SAT prototypes) (ablation)ProtoKV · LongBench average, budgets 128-512average scoreImproved
replace hybrid prototypes with LSH bucketing of all tokens (remove PDT chunks) (ablation)ProtoKV · LongBench average, budgets 128-512average scoreImproved

Earlier methods this paper compares with

Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.

Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy