Optimizing Decoding Paths in Masked Diffusion Models by Quantifying Uncertainty

Masked Diffusion Models (MDMs) offer flexible non-autoregressive generation, but their output quality is highly sensitive to the decoding order. This paper formalizes this issue by attributing variability to cumulative predictive uncertainty along a generative path. It introduces "Denoising Entropy" as a computable metric to quantify this uncertainty, serving as an internal signal for evaluating the generative process. Two algorithms, a post-hoc selection method and a real-time guidance strategy, are proposed to optimize the decoding path. Experiments show that these entropy-guided methods significantly improve generation quality and accuracy across challenging reasoning, planning, and code benchmarks, effectively turning uncertainty into an advantage for high-quality solutions in generative AI.