AI Paper Atlas

Research map / Architectures and efficiency / Attention and transformers

Faster Than Flash: Exploiting Attention Sparsity for Efficient Long-Context Decoding

Zhigeng Liu, Zhiyuan Ning, Ruixiao Li, Xiaoran Liu, Yuerong Song, Min Zhang, Ziwei He, Xipeng Qiu

ICML 2026 · Architectures and efficiency › Attention and transformers · approach: Long-context KV and memory attention · 0 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
replace top-k with top-delta selection (FP16 scan, split) (ablation)FFD · Ablation, 16k prefill + 128 decode, RTX 4090 RULERRULER AVG: 86.82 -> 90.39Improved
replace top-k with top-delta selection (FP16 scan, split) (ablation)FFD · Ablation, 16k prefill + 128 decode, RTX 4090throughput (tok/s): 17.43 -> 21.64Improved
add 2-bit scanning under top-k selection (ablation)FFD · Ablation, 16k prefill + 128 decode, RTX 4090throughput (tok/s): 17.43 -> 14.80Hurt
add 2-bit scanning under top-delta selection (ablation)FFD · Ablation, 16k prefill + 128 decode, RTX 4090throughput (tok/s): 21.64 -> 24.65Improved
add fused kernel and CUDA-graph execution under top-k (ablation)FFD · Ablation, 16k prefill + 128 decode, RTX 4090throughput (tok/s): 14.80 -> 21.21Improved
add fused kernel and CUDA-graph execution under top-delta (ablation)FFD · Ablation, 16k prefill + 128 decode, RTX 4090throughput (tok/s): 24.65 -> 53.96Improved
use top-delta instead of top-k at matched keep ratio (ablation)FFD · Per-head selection latencylatency (ms/head): 0.0090 -> 0.0044Improved
use top-delta instead of top-p at matched keep ratio (ablation)FFD · Per-head selection latencylatency (ms/head): 0.1115 -> 0.0044Improved
use top-delta instead of fixed-budget top-k on CWE (split execution) (ablation)FFD · RULER 32k CWEscore: 24.6 -> 68.2Improved
use 2-bit + 8-bit residual instead of 8+8 quantization (ablation)FFD · Decode kernel latency, RTX 4090 16kspeedup vs FlashAttn: 2.82 -> 10.65Improved
raise delta from 5 to 7 (ablation)FFD · RULER 32kaverage score: 87.1 -> 89.4Improved
raise delta from 5 to 7 (ablation)FFD · LongBenchaverage score: 25.86 -> 26.35Improved

Earlier methods this paper compares with

Named baselines from earlier accepted papers in this paper's results tables: better than 3 of 3 on most shared results.

Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.

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