Uniform Diffusion Models Revisited: Leave-One-Out Denoiser and Absorbing State Reformulation

Researchers revisit uniform diffusion for language generation and show its standard training target is a leave-one-out posterior, not the usual denoising posterior.

They derive exact conversions between the denoiser, leave-one-out posterior, and score, clarifying the ELBO mismatch that has held back uniform diffusion.

The paper also proposes sampling improvements, including temperature and predictor-corrector methods, that can boost performance without extra training.

An absorbing-state reformulation preserves the model’s law while simplifying inference, helping narrow or even reverse the gap with masked diffusion.