AI Paper Atlas

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

TPAMI 2026 · Architectures and efficiency › Attention and transformers · approach: KV cache eviction and compression · 5 citations (OpenAlex)

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 changedWhereMeasuredEffect
Task-KV vs HeadKV-R2 at 40% budgetLlama-2-7B-Chat · LongBench avgaverage score: 35.20 -> 35.32Improved
Task-KV vs HeadKV-R2 at 40% budgetMistral-7B-v0.2-Instruct · LooGLE avgaverage score: 10.63 -> 10.73Improved
Task-KV vs HeadKV-R2 at 60% budgetMistral-7B-v0.2-Instruct · LooGLE avgaverage score: 10.89 -> 11.02Improved
Task-KV vs HeadKV-R2 at 40% budgetLlama-2-7B-Chat · LooGLE avgaverage score: 7.80 -> 7.94Improved
Task-KV vs HeadKV-R2 at 40% budget on Single-Doc QALlama-2-7B-Chat · LongBench Single-Doc QAscore: 27.77 -> 26.70Hurt
Task-KV vs HeadKV-R2 on Synthetic tasks at 40% budgetLlama-2-7B-Chat · LongBench Syntheticscore: 5.25 -> 6.75Improved
Task-KV vs StreamingLLM at 40% budgetMistral-7B-v0.2-Instruct · LongBench avgaverage score: 39.14 -> 45.98Improved
Task-KV vs HeadKV-R2 at 50% budgetLlama-2-7B-Chat · Reasoning-in-a-Haystack avgaverage score: 36.85 -> 37.15Improved
Task-KV vs SnapKV at 50% budgetMistral-7B-v0.2-Instruct · Reasoning-in-a-Haystack avgaverage score: 40.83 -> 40.94Improved
Task-KV vs FullKV at 60% budgetMistral-7B-v0.2-Instruct · LongBench avgaverage score: 46.47 -> 46.42No clear effect
remove non-heterogeneous heads (ablation)Mistral-7B-v0.2-Instruct · Passkey Retrievalretrieval accuracyNo clear effect
remove non-heterogeneous heads (ablation)Mistral-7B-v0.2-Instruct · Reasoning-in-a-HaystackscoreHurt

Closest papers by meaning

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.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy