Research map / Architectures and efficiency / Attention and transformers
SparK: Query-Aware Unstructured Sparsity with Recoverable KV Cache Channel Pruning
Huanxuan Liao, Yixing Xu, Shizhu He, Guanchen Li, Xuanwu Yin, Dong Li, Emad Barsoum, Jun Zhao, Kang 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 |
|---|---|---|---|
| use SparK instead of THINK for key channel pruning at lambda 0.8 | SnapKV (KV size 128) · LongBench | average score: 16.97 -> 31.16 | Improved |
| use SparK instead of THINK for key channel pruning at lambda 0.8 | SnapKV (16K, 20% budget) · RULER | average score: 3.03 -> 77.51 | Improved |
| add SparK (0.5) on top of SnapKV | SnapKV (KV size 128) · LongBench | average score: 32.38 -> 32.04 | No clear effect |
| add SparK (0.5) on top of PyramidKV | PyramidKV (KV size 128) · LongBench | average score: 43.80 -> 43.81 | No clear effect |
| add SparK (0.8) on top of PyramidKV | PyramidKV (KV size 128) · LongBench | average score: 43.80 -> 41.87 | Hurt |
| add recovery of pruned channels (SparK vs without recovery) (ablation) | SnapKV · LongBench | average score | Improved |
| replace pruned entries with constant 0.01 instead of removal (ablation) | key channel pruning at 80% · LongBench | performance degradation: 55.7% -> 12.2% degradation on SAMSum | Improved |
| use unstructured instead of structured pruning at 80% (ablation) | channel pruning · LongBench | performance loss: 55.7% vs 28.3% drop | Improved |
| Gaussian or Exponential instead of degenerate recovery distribution (ablation) | SparK (lambda 0.5, KV 128) · LongBench | average score: 32.04 vs 32.71 / 32.56 | No clear effect |
| ratio-free variants SparK-p / SparK-g instead of fixed ratio (ablation) | SparK · LongBench | average score: 32.04 vs 32.11 | No clear effect |
Closest papers by meaning
- ThinK: Thinner Key Cache by Query-Driven Pruning (ICLR 2025 · similarity 0.91)
- QUEST: Query-Aware Sparsity for Efficient Long-Context LLM Inference (ICML 2024 · similarity 0.91)
- RocketKV: Accelerating Long-Context LLM Inference via Two-Stage KV Cache Compression (ICML 2025 · similarity 0.90)
- SCBench: A KV Cache-Centric Analysis of Long-Context Methods (ICLR 2025 · similarity 0.89)
- Efficient Low Rank Attention for Long-Context Inference in Large Language Models (NeurIPS 2025 · similarity 0.89)
- MUSTAFAR: Promoting Unstructured Sparsity for KV Cache Pruning in LLM Inference (NeurIPS 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.
