A new method called Token-to-Mask (T2M) has been proposed to improve the performance of diffusion language models. T2M identifies low-confidence positions in generated text and remasks them, allowing for more accurate predictions. This approach is training-free and can be applied across different tasks and models without requiring retuning. In experiments, T2M was found to outperform direct editing and autoregressive token-by-token decoding on many tasks. --- Why it matters: This matters because it shows that small adjustments to the inference process of diffusion language models can lead to significant improvements in performance. This could have implications for applications such as text generation, where accurate predictions are crucial. Source: https://arxiv.org/abs/2604.18738