Reproduce “Blurring the Block Boundaries: Overlapping Block Denoising for Efficient Diffusion Language Models” 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
Amir Mann, Jonathan Fhima, Shaked Brody, Or Avitan, Asif Zamir, Shahar Tsiper, Niv Nayman
B3 starts refining the next text block during late denoising passes on the current block, sharing a model forward pass. A self-calibrating readiness gate opens the overlap without a tuned threshold. Across three models and 10 benchmarks, mean tokens per second rise 11.5–14.5% and tokens per forward 15.9–19.4%; average accuracy changes range from −0.40 to +0.76 percentage points. Results cover English text tasks and do not establish reduced total computation.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.