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
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 changed | Where | Measured | Effect |
|---|---|---|---|
| replace AdaKV head allocation with LU-KV utility-based allocation (KeyDiff) | Mistral-7B-Instruct-v0.3 · LongBench 80% compression | average score: 40.54 -> 46.21 | Improved |
| replace AdaKV allocation with LU-KV (SnapKV) | Mistral-7B-Instruct-v0.3 · RULER 16K 80% compression | average accuracy: 37.48 -> 69.98 | Improved |
| replace AdaKV allocation with LU-KV (KeyDiff) | Qwen2.5-32B-Instruct · RULER 16K 80% compression | average accuracy: 80.81 -> 88.88 | Improved |
| replace uniform allocation with LU-KV (SnapKV) | Qwen2.5-32B-Instruct · LongBench 80% compression | average score: 41.21 -> 47.95 | Improved |
| use PyramidKV layer-wise allocation instead of uniform | SnapKV/KeyDiff, Qwen2.5-32B · LongBench 80% compression | average score: 41.21 -> 38.68 (SnapKV) | Hurt |
| use PyramidKV instead of uniform | SnapKV, Mistral-7B · LongBench 80% compression | average score: 40.67 -> 40.94 | No clear effect |
| add optimality gap loss to allocation objective (ablation) | LU-KV-SnapKV, Mistral-7B · LongBench 80% compression | average score: 43.95 -> 45.79 | Improved |
| add optimality gap loss to allocation objective (ablation) | LU-KV-KeyDiff, Mistral-7B · LongBench 80% compression | average score: 41.98 -> 46.21 | Improved |
| add optimality gap loss to allocation objective (ablation) | LU-KV-SnapKV, Qwen2.5-32B · LongBench 80% compression | average score: 45.00 -> 47.95 | Improved |
| add optimality gap loss to allocation objective (ablation) | LU-KV-KeyDiff, Qwen2.5-32B · LongBench 80% compression | average score: 43.18 -> 48.26 | Improved |
| use greedy solver on convex-hull relaxation instead of exact DP (ablation) | LU-KV · allocation optimality | allocation solution | No clear effect |
| use LU-KV allocation instead of AdaKV | SnapKV, Mistral-7B · total oracle eviction loss | eviction loss: 4.98e-04 -> 3.87e-04 | Improved |
Closest papers by meaning
- CAKE: Cascading and Adaptive KV Cache Eviction with Layer Preferences (ICLR 2025 · similarity 0.91)
- Value-Aware Stochastic KV Cache Eviction for Reasoning Models (NeurIPS 2026 · similarity 0.90)
- Ada-KV: Optimizing KV Cache Eviction by Adaptive Budget Allocation for Efficient LLM Inference (NeurIPS 2025 · similarity 0.90)
- DropKV: Decoupling Residual-Output Perturbation for Near-Optimal KV-Cache Eviction (NeurIPS 2026 · similarity 0.90)
- AttentionPredictor: Temporal Patterns Matter for KV Cache Compression (NeurIPS 2025 · similarity 0.89)
- ForesightKV: Optimizing KV Cache Eviction for Reasoning Models by Learning Long-Term Contribution (ICML 2026 · similarity 0.89)
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.
