Uniform-state diffusion models (USDMs) differ from masked diffusion models in a fundamental way: they can revise any token at any denoising step. This allows the model to correct its own mistakes during generation, a property that masked diffusion lacks. The new arXiv paper, "Know When to Hold 'em," focuses on this self-correction ability.
The authors note that self-correction is not simply about replacing tokens. It also requires knowing which tokens are already correct and should be retained. The abstract argues that effective self-correction depends on both revising incorrect tokens and holding onto correct ones, suggesting the paper proposes a method to manage this balance.
Because the abstract is brief, the specific mechanism and experimental results are not detailed here. The central contribution, however, is clear: a strategy for correct-token retention within USDMs, which could strengthen the practical advantage of this model class over masked alternatives.