AI Paper Atlas

Research map / Architectures and efficiency / Attention and transformers

FlashBlock: Attention Caching for Efficient Long-Context Block Diffusion

Zhuokun Chen, Jianfei Cai, Bohan Zhuang

ICML 2026 · Architectures and efficiency › Attention and transformers · approach: Sparse attention for long-context LLMs · 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
add block-external attention cachingTrado-8B-Thinking (block size 4) · 800k context, batch 128throughput (TPS): 312 -> 451Improved
add block-external attention cachingTrado-8B-Thinking (block size 8) · 800k context, batch 128throughput (TPS): 532 -> 674Improved
add block-external attention cachingTrado-8B-Thinking · GSM8K/MATH500/AIMEaccuracy: 93.25 -> 93.12 (GSM8K, block 4)No clear effect
add block-external attention cachingTrado-8B-Thinking (block size 8) · GSM8Kaccuracy: 91.74 -> 90.22Hurt
add block-external attention cachingTrado-8B-Thinking · MBPPpass@1: 25.60 -> 33.60 (block 4)Improved
add block-external attention caching on top of SparseD (d=20%)SparseD · GSM8Kaccuracy: 34.72 -> 42.68Improved
add block-external attention caching on top of SparseD (d=20%)SparseD · HumanEvalpass@1: 23.78 -> 33.54Improved
add block-external attention caching on top of SparseD (d=40%)SparseD · GSM8Kaccuracy: 84.61 -> 87.26Improved
add block-external attention caching on top of SparseDSparseD · L1 distance to full attention, density 20%L1 distance: 0.0028 -> 0.0008Improved
add reuse-aware distillation (ablation)FlashBlock (tau=2) · MATH500accuracy: 80.40 -> 85.80Improved
add reuse-aware distillation (ablation)FlashBlock (tau=4) · MATH500accuracy: 48.40 -> 66.20Improved
increase update threshold tau from 2 to 4 (ablation)FlashBlock · MATH500accuracy: 85.80 -> 66.20 (with distillation)Hurt

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