Reproduce “Effective Test-Time Scaling of Discrete Diffusion through Iterative Refinement” 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

Sanghyun Lee, Sunwoo Kim, Seungryong Kim, Jongho Park, Dongmin Park

IterRef adds test-time computation to discrete diffusion by repeatedly remasking and regenerating intermediate states, then accepting candidates with a Multiple-Try Metropolis rule weighted by reward. Across masked language and image generation, it improved reward alignment over Best-of-N, FK steering, SVDD, and related baselines, often matching their 32-NFE performance with 2–4 NFEs and raising MaskGIT CLIPScore from 30.5 to 35.8 over 16 NFEs.

Key insights

  • Reversible intermediate-state correction is more effective than simply drawing more independent samples or guiding only the forward denoising trajectory.
  • Increasing the number of refinement iterations generally helps more than increasing the number of candidates per iteration, although excessive refinement with insufficient candidate diversity can hurt performance.
  • Later denoising stages are often the most useful locations for refinement, but the best timestep allocation depends on the task and reward.

Summary by alphaXiv.

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