Research map / Architectures and efficiency / Attention and transformers
ReST-KV: Robust KV Cache Eviction with Layer-wise Output Reconstruction and Spatial-Temporal Smoothing
Yongqi An, Chang Lu, Kuan Zhu, Tao Yu, Chaoyang Zhao, Hong Wu, Ming Tang, Jinqiao Wang
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 |
|---|---|---|---|
| add layer-wise output reconstruction indicator (ablation) | ReST-KV · LongBench (128L) | avg score: 33.95 -> 35.86 | Improved |
| add EMA temporal smoothing (ablation) | ReST-KV · LongBench (128L) | avg score: 34.02 -> 35.86 | Improved |
| add adaptive window spatial smoothing (ablation) | ReST-KV · LongBench (128L) | avg score: 33.50 -> 35.86 | Improved |
| use full ReST-KV instead of attention-weight top-k (ablation) | KV eviction · LongBench (128L) | avg score: 32.98 -> 35.86 | Improved |
| combine ReST-KV with PyramidKV layer-wise budget allocation (ablation) | ReST-KV · LongBench (Llama2-7B-Chat, avg over budgets) | avg score: 30.48 -> 30.93 | Improved |
| combine ReST-KV with AdaKV head-wise budget allocation (ablation) | ReST-KV · LongBench (Llama2-7B-Chat, avg over budgets) | avg score: 30.48 -> 31.20 | Improved |
| use ReST-KV instead of SnapKV at tight budget | KV eviction · LongBench (Llama-3.1-8B, 64L) | avg score: 31.48 -> 33.54 | Improved |
| use ReST-KV instead of SnapKV at large budget | KV eviction · LongBench (Llama2-7B-Chat, 1024L, with Pyramid/Ada table) | avg score: 32.99 -> 32.97 | No clear effect |
| use ReST-KV instead of SnapKV on RULER | KV eviction · RULER (avg 4K-128K) | avg accuracy: 67.11 -> 82.27 | Improved |
| ReST-KV prefill overhead vs SnapKV | ReST-KV · Prefill TTFT 128k | TTFT (ms): 28,510 -> 29,082 | Hurt |
Closest papers by meaning
- CAKE: Cascading and Adaptive KV Cache Eviction with Layer Preferences (ICLR 2025 · similarity 0.92)
- Ada-KV: Optimizing KV Cache Eviction by Adaptive Budget Allocation for Efficient LLM Inference (NeurIPS 2025 · similarity 0.92)
- $\text{D}_{2}\text{O}$: Dynamic Discriminative Operations for Efficient Long-Context Inference of Large Language Models (ICLR 2025 · similarity 0.92)
- LaCache: Ladder-Shaped KV Caching for Efficient Long-Context Modeling of Large Language Models (ICML 2025 · similarity 0.91)
- MiniCache: KV Cache Compression in Depth Dimension for Large Language Models (NeurIPS 2024 · similarity 0.91)
- EchoKV: Efficient KV Cache Compression via Similarity-Based Reconstruction (NeurIPS 2026 · similarity 0.91)
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.
