Research map / Architectures and efficiency / Attention and transformers
OBCache: Optimal Brain KV Cache Pruning for Efficient Long-Context LLM Inference
Yuzhe Gu, Xiyu Liang, Jiaojiao Zhao, Enmao Diao
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 attention score with OBCache-V&K | H2O · RULER-4K query-agnostic | avg accuracy: 31.7 -> 40.0 | Improved |
| replace attention score with OBCache-V&K | AdaKV · RULER-4K query-agnostic | avg accuracy: 43.0 -> 55.2 | Improved |
| replace attention score with OBCache-V&K | AdaKV · RULER-4K query-aware | avg accuracy: 75.7 -> 81.9 | Improved |
| replace attention score with OBCache-V&K | H2O · RULER-4K query-aware | avg accuracy: 57.5 -> 67.8 | Improved |
| replace attention score with OBCache-V&K | TOVA · RULER-4K query-aware | avg accuracy: 74.5 -> 76.7 | Improved |
| replace attention score with OBCache-V&K | H2O · RULER-32K query-agnostic | avg accuracy: 31.5 -> 36.7 | Improved |
| replace attention score with OBCache-V&K | AdaKV · RULER-32K query-agnostic | avg accuracy: 45.5 -> 55.1 | Improved |
| replace attention score with OBCache-V | SnapKV · RULER-4K query-aware | avg accuracy: 72.4 -> 71.9 | Hurt |
| replace attention score with OBCache-V | SnapKV · RULER-4K query-agnostic | avg accuracy: 37.9 -> 35.8 | Hurt |
| use key-pruning score instead of value-only score (ablation) | AdaKV + OBCache · RULER-4K query-agnostic | avg accuracy: 51.4 -> 55.0 | Improved |
| add key score to value score (V&K versus K only) (ablation) | AdaKV + OBCache · RULER-4K query-agnostic | avg accuracy: 55.0 -> 55.2 | No clear effect |
| use OBCache-K scores | AdaKV · LongBench 10% budget | score: +1.2 query-aware; +2.6 query-agnostic | Improved |
Closest papers by meaning
- ThinK: Thinner Key Cache by Query-Driven Pruning (ICLR 2025 · similarity 0.92)
- SmallKV: Small Model Assisted Compensation of KV Cache Compression for Efficient LLM Inference (NeurIPS 2025 · similarity 0.91)
- CAKE: Cascading and Adaptive KV Cache Eviction with Layer Preferences (ICLR 2025 · similarity 0.91)
- ArkVale: Efficient Generative LLM Inference with Recallable Key-Value Eviction (NeurIPS 2024 · similarity 0.91)
- SqueezeAttention: 2D Management of KV-Cache in LLM Inference via Layer-wise Optimal Budget (ICLR 2025 · similarity 0.90)
- Ada-KV: Optimizing KV Cache Eviction by Adaptive Budget Allocation for Efficient LLM Inference (NeurIPS 2025 · similarity 0.90)
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.
