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
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 |
|---|---|---|---|
| Use CapKV instead of SnapKV | Qwen3-8B · LongBench avg, compression 0.5 | score: 45.36 -> 48.15 | Improved |
| Use CapKV instead of SnapKV | Qwen3-14B · LongBench avg, compression 0.9 | score: 37.13 -> 39.53 | Improved |
| Use CapKV instead of KeyDiff | Qwen3-8B · LongBench avg, compression 0.75 | score: 29.10 -> 44.88 | Improved |
| Use CapKV instead of Expected Attention | Qwen3-8B · LongBench avg, compression 0.9 | score: 36.35 -> 36.57 | No clear effect |
| Use CapKV instead of SnapKV under ultra-long contexts | Qwen3-8B (YaRN) · LongBench v2, compression 0.5 | score: 46.2 -> 43.7 | Hurt |
| Use CapKV instead of SnapKV under ultra-long contexts | Qwen3-8B (YaRN) · LongBench v2, compression 0.75 | score: 38.7 -> 47.9 | Improved |
| Use exact output direction u = W_O v instead of u = v (ablation) | CapKV · LongBench 5-subset, compression 0.5 | score: 57.00 -> 57.18 | No clear effect |
| Use covariance-based gating instead of mean-query gating (ablation) | CapKV · LongBench 5-subset, compression 0.75 | score: 52.64 -> 49.22 | Hurt |
| Add query-aware weighting (tau=5 vs tau=0) (ablation) | CapKV · LongBench avg, Llama3.1-8B | score: 40.59 -> 41.73 | Improved |
| Use large tau (10) for query weighting (ablation) | CapKV · LongBench avg, Llama3.1-8B | score: 41.73 -> 40.53 | Hurt |
| Combine eviction with 4-bit HQQ KV quantization | CapKV · LongBench subsets, compression 0.5 | score: 55.47 avg, best among methods | No clear effect |
Closest papers by meaning
- DropKV: Decoupling Residual-Output Perturbation for Near-Optimal KV-Cache Eviction (NeurIPS 2026 · similarity 0.93)
- CAKE: Cascading and Adaptive KV Cache Eviction with Layer Preferences (ICLR 2025 · similarity 0.93)
- Taming the Fragility of KV Cache Eviction in LLM Inference (ICLR 2026 · similarity 0.93)
- Learning to Evict from Key-Value Cache (ICML 2026 · similarity 0.92)
- Ada-KV: Optimizing KV Cache Eviction by Adaptive Budget Allocation for Efficient LLM Inference (NeurIPS 2025 · similarity 0.92)
- Value-Aware Stochastic KV Cache Eviction for Reasoning Models (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.
