Inverse-Confidence Sampling for Continuous Diffusion Language Models

In diffusion language models (DLMs), continuous-space DLMs (CDLMs) are narrowing the quality gap compared to masked diffusion models (MDMs). Yet, CDLMs denoise tokens uniformly, leaving the design of sampling trajectories largely unexplored. We present a counter-intuitive inference-time procedure, termed \namelong~(\name). \name~allocates each step's denoising budget \emph{inversely} proportional to the per-token confidence, which allows informative motifs to emerge early and anchor the remaining tokens. Without retraining, \name~substantially improves sample quality for CDLMs across images (MNIST) and natural language (LM1B, OpenWebText), reaching state-of-the-art metrics amongst comparable DLMs.