Schedule optimization for tau-leaping in masked discrete diffusion
The paper shows when tau-leaping schedules can reduce masked diffusion sampling error, and when they only improve the constant.
Secchi and Zanella analyze the factorization error introduced when masked discrete diffusion reveals multiple coordinates in parallel. They express that error through a dependence profile, then derive finite-step stationarity conditions and a limiting optimal smooth schedule. If the dependence profile stays uniformly positive, schedule optimization improves the leading constant but not the `N/K` error scaling. If the profile degenerates, carefully chosen schedules can beat the uniform schedule asymptotically. ArXiv · AI/CL/LG's note
Secchi and Zanella analyze the factorization error introduced when masked discrete diffusion reveals multiple coordinates in parallel. They express that error through a dependence profile, then derive finite-step stationarity conditions and a limiting optimal smooth schedule. If the dependence profile stays uniformly positive, schedule optimization improves the leading constant but not the `N/K` error scaling. If the profile degenerates, carefully chosen schedules can beat the uniform schedule asymptotically. ArXiv · AI/CL/LG's note
score 5