Research map / Architectures and efficiency / Attention and transformers
Task-KV: Task-Aware KV Cache Optimization via Semantic Differentiation of Attention Heads
Xiaowei He, Jie Liu, Shaowei 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 |
|---|---|---|---|
| Task-KV vs HeadKV-R2 at 40% budget | Llama-2-7B-Chat · LongBench avg | average score: 35.20 -> 35.32 | Improved |
| Task-KV vs HeadKV-R2 at 40% budget | Mistral-7B-v0.2-Instruct · LooGLE avg | average score: 10.63 -> 10.73 | Improved |
| Task-KV vs HeadKV-R2 at 60% budget | Mistral-7B-v0.2-Instruct · LooGLE avg | average score: 10.89 -> 11.02 | Improved |
| Task-KV vs HeadKV-R2 at 40% budget | Llama-2-7B-Chat · LooGLE avg | average score: 7.80 -> 7.94 | Improved |
| Task-KV vs HeadKV-R2 at 40% budget on Single-Doc QA | Llama-2-7B-Chat · LongBench Single-Doc QA | score: 27.77 -> 26.70 | Hurt |
| Task-KV vs HeadKV-R2 on Synthetic tasks at 40% budget | Llama-2-7B-Chat · LongBench Synthetic | score: 5.25 -> 6.75 | Improved |
| Task-KV vs StreamingLLM at 40% budget | Mistral-7B-v0.2-Instruct · LongBench avg | average score: 39.14 -> 45.98 | Improved |
| Task-KV vs HeadKV-R2 at 50% budget | Llama-2-7B-Chat · Reasoning-in-a-Haystack avg | average score: 36.85 -> 37.15 | Improved |
| Task-KV vs SnapKV at 50% budget | Mistral-7B-v0.2-Instruct · Reasoning-in-a-Haystack avg | average score: 40.83 -> 40.94 | Improved |
| Task-KV vs FullKV at 60% budget | Mistral-7B-v0.2-Instruct · LongBench avg | average score: 46.47 -> 46.42 | No clear effect |
| remove non-heterogeneous heads (ablation) | Mistral-7B-v0.2-Instruct · Passkey Retrieval | retrieval accuracy | No clear effect |
| remove non-heterogeneous heads (ablation) | Mistral-7B-v0.2-Instruct · Reasoning-in-a-Haystack | score | Hurt |
Closest papers by meaning
- Not All Heads Matter: A Head-Level KV Cache Compression Method with Integrated Retrieval and Reasoning (ICLR 2025 · similarity 0.91)
- CateKV: On Sequential Consistency for Long-Context LLM Inference Acceleration (ICML 2025 · similarity 0.91)
- Model Tells You What to Discard: Adaptive KV Cache Compression for LLMs (ICLR 2024 · similarity 0.91)
- SCBench: A KV Cache-Centric Analysis of Long-Context Methods (ICLR 2025 · similarity 0.91)
- AttentionPredictor: Temporal Patterns Matter for KV Cache Compression (NeurIPS 2025 · similarity 0.91)
- SqueezeAttention: 2D Management of KV-Cache in LLM Inference via Layer-wise Optimal Budget (ICLR 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.
