Masked diffusion models generate data by starting from masked tokens and unmasking several at a time, which makes generation faster. However, the standard approach samples each token independently, and that ignores correlations among tokens, biasing the output.
A new paper by Anari et al. addresses this problem with rejection sampling. The method is described as query-efficient, meaning it aims to correct the bias without requiring an impractical number of model evaluations.
The abstract does not provide implementation details, but the contribution is a correction procedure for parallel decoding in masked diffusion models. Since this is a single source, there are no conflicting accounts to compare.