“The proposed adaptive order policy for masked diffusion models uses a loss function that reweights terms according to policy probabilities, resulting in a policy that favors positions where the denoiser is more likely to be correct.”