learning-to-unmask

Proposes learning the unmasking order in masked diffusion models using a lightweight policy network, with a weighted loss that outperforms heuristics on combinatorial tasks and protein design.