AI Paper Atlas

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

AAAI 2026 · Architectures and efficiency › Attention and transformers · approach: KV cache eviction and compression · 4 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
use SparK instead of THINK for key channel pruning at lambda 0.8SnapKV (KV size 128) · LongBenchaverage score: 16.97 -> 31.16Improved
use SparK instead of THINK for key channel pruning at lambda 0.8SnapKV (16K, 20% budget) · RULERaverage score: 3.03 -> 77.51Improved
add SparK (0.5) on top of SnapKVSnapKV (KV size 128) · LongBenchaverage score: 32.38 -> 32.04No clear effect
add SparK (0.5) on top of PyramidKVPyramidKV (KV size 128) · LongBenchaverage score: 43.80 -> 43.81No clear effect
add SparK (0.8) on top of PyramidKVPyramidKV (KV size 128) · LongBenchaverage score: 43.80 -> 41.87Hurt
add recovery of pruned channels (SparK vs without recovery) (ablation)SnapKV · LongBenchaverage scoreImproved
replace pruned entries with constant 0.01 instead of removal (ablation)key channel pruning at 80% · LongBenchperformance degradation: 55.7% -> 12.2% degradation on SAMSumImproved
use unstructured instead of structured pruning at 80% (ablation)channel pruning · LongBenchperformance loss: 55.7% vs 28.3% dropImproved
Gaussian or Exponential instead of degenerate recovery distribution (ablation)SparK (lambda 0.5, KV 128) · LongBenchaverage score: 32.04 vs 32.71 / 32.56No clear effect
ratio-free variants SparK-p / SparK-g instead of fixed ratio (ablation)SparK · LongBenchaverage score: 32.04 vs 32.11No clear effect

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