Research map / Architectures and efficiency / Attention and transformers
Sparse Attention Across Multiple-Context KV Cache
Ziyi Cao, Qingyi Si, Jingbin Zhang, Bingquan Liu
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 personalized bias to query vector (no recompute) (ablation) | SamKV (Llama 3.1 8B) · Table 4 avg F1 | avg F1: 15.00 -> 15.40 | Improved |
| add personalized bias (with recompute) (ablation) | SamKV (Llama 3.1 8B) · Table 4 avg F1 | avg F1: 22.21 -> 22.75 | Improved |
| add recomputation of sparse tokens (ablation) | SamKV (Llama 3.1 8B) · Table 4 avg F1 | avg F1: 15.40 -> 22.75 | Improved |
| add middle-section KV selection (no recompute) (ablation) | SamKV (Llama 3.1 8B) · Table 4 avg F1 | avg F1: 14.68 -> 15.00 | No clear effect |
| use KV fusion instead of overwrite | SamKV (Mistral 7B) · 2WikiMQA | F1: 27.04 -> 27.88 | Improved |
| use KV fusion instead of overwrite | SamKV (Llama 3.1 8B) · HotpotQA | F1: 35.27 -> 30.52 | Hurt |
| use SamKV (15% sequence) instead of full recomputation | Llama 3.1 8B · HotpotQA | F1: 24.12 -> 35.27 | Improved |
| use SamKV instead of full recomputation | Mistral 7B · MuSiQue | F1: 19.94 -> 18.54 | Hurt |
| use SamKV instead of CacheBlend/EPIC | Mistral 7B · 2WikiMQA | F1: 20.42 / 19.40 -> 27.88 | Improved |
Closest papers by meaning
- SCBench: A KV Cache-Centric Analysis of Long-Context Methods (ICLR 2025 · similarity 0.91)
- Training Free Exponential Context Extension via Cascading KV Cache (ICLR 2025 · similarity 0.91)
- SqueezeAttention: 2D Management of KV-Cache in LLM Inference via Layer-wise Optimal Budget (ICLR 2025 · similarity 0.90)
- SALS: Sparse Attention in Latent Space for KV Cache Compression (NeurIPS 2025 · similarity 0.89)
- QUEST: Query-Aware Sparsity for Efficient Long-Context LLM Inference (ICML 2024 · similarity 0.89)
- RazorAttention: Efficient KV Cache Compression Through Retrieval Heads (ICLR 2025 · similarity 0.89)
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.
