Empirically validate the Fisher-Rao-optimal (cosine) schedule for masked discrete diffusion

Establish empirical validation of using the cosine discretization schedule, derived as Fisher-Rao-optimal for masked discrete diffusion models, by demonstrating through experiments that this schedule performs effectively for sampling in masked discrete diffusion generation tasks.

Background

The paper studies discretization schedules for masked discrete diffusion models from an information-geometric perspective. By computing the Fisher-Rao metric along the probability path induced by masked discrete diffusion, the authors derive the optimal schedule as the geodesic under this metric and show it coincides with the popularly used cosine schedule.

The main contribution is theoretical; the authors explicitly state that empirical validation of this schedule choice remains to be done, motivating experiments to test the practical effectiveness of the cosine schedule in masked discrete diffusion sampling.