A new mean-to-score method corrects Bayes realizability violations in Score Entropy Discrete Diffusion (SEDD) by projecting raw scores onto the bridge polytope. This correction reduces generative perplexity from 203.6 to 175.1 in discrete diffusion models.
HOW THIS AFFECTS YOU
●
researcherYou can implement this projection technique to prevent negative pre-normalization weights and improve sampling quality in discrete diffusion.