AI Paper Atlas

Research map / Architectures and efficiency / Attention and transformers

Predicting Future Utility: Global Combinatorial Optimization for Task-Agnostic KV Cache Eviction

Ziyao Tang, Pengkun Jiao, Xinhang Chen, Wei Liu, Shiyong Li, Jingjing Chen

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
replace AdaKV head allocation with LU-KV utility-based allocation (KeyDiff)Mistral-7B-Instruct-v0.3 · LongBench 80% compressionaverage score: 40.54 -> 46.21Improved
replace AdaKV allocation with LU-KV (SnapKV)Mistral-7B-Instruct-v0.3 · RULER 16K 80% compressionaverage accuracy: 37.48 -> 69.98Improved
replace AdaKV allocation with LU-KV (KeyDiff)Qwen2.5-32B-Instruct · RULER 16K 80% compressionaverage accuracy: 80.81 -> 88.88Improved
replace uniform allocation with LU-KV (SnapKV)Qwen2.5-32B-Instruct · LongBench 80% compressionaverage score: 41.21 -> 47.95Improved
use PyramidKV layer-wise allocation instead of uniformSnapKV/KeyDiff, Qwen2.5-32B · LongBench 80% compressionaverage score: 41.21 -> 38.68 (SnapKV)Hurt
use PyramidKV instead of uniformSnapKV, Mistral-7B · LongBench 80% compressionaverage score: 40.67 -> 40.94No clear effect
add optimality gap loss to allocation objective (ablation)LU-KV-SnapKV, Mistral-7B · LongBench 80% compressionaverage score: 43.95 -> 45.79Improved
add optimality gap loss to allocation objective (ablation)LU-KV-KeyDiff, Mistral-7B · LongBench 80% compressionaverage score: 41.98 -> 46.21Improved
add optimality gap loss to allocation objective (ablation)LU-KV-SnapKV, Qwen2.5-32B · LongBench 80% compressionaverage score: 45.00 -> 47.95Improved
add optimality gap loss to allocation objective (ablation)LU-KV-KeyDiff, Qwen2.5-32B · LongBench 80% compressionaverage score: 43.18 -> 48.26Improved
use greedy solver on convex-hull relaxation instead of exact DP (ablation)LU-KV · allocation optimalityallocation solutionNo clear effect
use LU-KV allocation instead of AdaKVSnapKV, Mistral-7B · total oracle eviction losseviction loss: 4.98e-04 -> 3.87e-04Improved

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