Reproduce “Diffusion ReRoll: Revisable Denoising for Robotic Sequential Prediction” 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
Seonsoo Kim, Seongil Hong, Jun-Gill Kang
ReRoll revises sequence predictions by re-noising selected regions while others continue denoising, then refining them with updated context. It trains on randomized linear-chunk noise patterns aligned with this schedule. In OGBench maze planning, average success rises from 72.78% to 88.07% versus Diffusion Forcing, a 21% relative gain across six variants; this is not a guarantee across tasks.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.