Systematic Hazard Sampling: Minimal-Variance Inference for Discrete Diffusion and Flow Models
Stratified Hazard Sampling (SHS) is a training-free method for discrete diffusion and flow models that estimates jump counts while minimizing variance. It treats token edits as events driven by cumulative hazard, using random phase stratification to place these events. This preserves expected jump counts and achieves the lowest possible conditional variance among unbiased estimators.