Reproduce “D-DOIT: Training-free Adaptation of Discrete Diffusion via Doob's h-Transform” 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

Jieke Wu, Qijie Zhu, Weimin Wu, Zeqi Ye, Minshuo Chen, Han Liu

D-DOIT guides masked discrete diffusion by sampling among candidate next states according to reward-scored completions, using a frozen prediction head. Its exact Doob h-transform reweights transitions by future reward, but the practical proxy sampler lacks that guarantee. In HepG2 enhancer design, it reports 91.7% ATAC-Acc versus 43.5% for equally timed endpoint search, a predictive result, not experimental validation.

Summary by alphaXiv.

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