Reproduce “Continuous Diffusion Models Can Obey Formal Syntax” 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

Jinwoo Kim, Taylor Berg-Kirkpatrick, Loris D'Antoni

DIFFINITY guides continuous diffusion denoising with a regex-derived probability score computed from the model’s current token distributions, steering whole-sequence revisions without a separately trained constraint classifier. On synthetic natural-language regex benchmarks, it averaged 92.9% constraint satisfaction, versus 20.5% for GPT-2 Large with Guidance constrained decoding. On JSON regexes, satisfaction was 68.4%, below GPT-2 Large’s 77.0%, so results vary by task.

Summary by alphaXiv.

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