While normalizing flows for continuous data have been extensively researched,
flows for discrete data have only recently been explored. These prior models,
however, suffer from limitations that are distinct from those of continuous
flows. Most notably, discrete flow-based models cannot