AI Paper Atlas

Research map / Architectures and efficiency / Attention and transformers

Rethinking KV Cache Eviction via a Unified Information-Theoretic Objective

Jiaming Yang, Chenwei Tang, Liangli Zhen, Jiancheng Lv

ICML 2026 · Architectures and efficiency › Attention and transformers · approach: KV cache eviction and compression · 0 citations (OpenAlex)

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
Use CapKV instead of SnapKVQwen3-8B · LongBench avg, compression 0.5score: 45.36 -> 48.15Improved
Use CapKV instead of SnapKVQwen3-14B · LongBench avg, compression 0.9score: 37.13 -> 39.53Improved
Use CapKV instead of KeyDiffQwen3-8B · LongBench avg, compression 0.75score: 29.10 -> 44.88Improved
Use CapKV instead of Expected AttentionQwen3-8B · LongBench avg, compression 0.9score: 36.35 -> 36.57No clear effect
Use CapKV instead of SnapKV under ultra-long contextsQwen3-8B (YaRN) · LongBench v2, compression 0.5score: 46.2 -> 43.7Hurt
Use CapKV instead of SnapKV under ultra-long contextsQwen3-8B (YaRN) · LongBench v2, compression 0.75score: 38.7 -> 47.9Improved
Use exact output direction u = W_O v instead of u = v (ablation)CapKV · LongBench 5-subset, compression 0.5score: 57.00 -> 57.18No clear effect
Use covariance-based gating instead of mean-query gating (ablation)CapKV · LongBench 5-subset, compression 0.75score: 52.64 -> 49.22Hurt
Add query-aware weighting (tau=5 vs tau=0) (ablation)CapKV · LongBench avg, Llama3.1-8Bscore: 40.59 -> 41.73Improved
Use large tau (10) for query weighting (ablation)CapKV · LongBench avg, Llama3.1-8Bscore: 41.73 -> 40.53Hurt
Combine eviction with 4-bit HQQ KV quantizationCapKV · LongBench subsets, compression 0.5score: 55.47 avg, best among methodsNo clear effect

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