Learning Unmasking Policies for Diffusion Language Models

Researchers have introduced a new decoding method for masked diffusion language models that treats token generation as a continuous flow rather than a series of binary choices. By representing the transition between masked and unmasked states as a gradual prediction process, this approach improves how models handle uncertainty during text generation. This shift allows for more fluid refinement of sequences compared to standard techniques that commit to specific tokens in discrete, rigid steps.