Reproduce “Coarse Semantic Blocks for Efficient Training-Free Sparse Attention” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Sosuke Kobayashi, Shoichiro Yamaguchi, Seiya Tokui

The authors introduce semantic block sparse attention. It groups similar prompt keys, lets queries select blocks, and batches queries choosing the same blocks for GPU execution. On Qwen3.5-35B-A3B at 512K, prefill takes 20.73 seconds versus 49.19 seconds dense, a 2.4× speedup on one H200. At 1M, RULER-5 averages 73.9% versus 73.2% dense, but MRCR-8 falls below dense at 512K.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.