Cyclical MCMC is a popular method for sampling from complex multimodal distributions. In our #NeurIPS2024 paper, we extend this approach to discrete distributions with novel designs tailored for discrete variables. Applications include graphic models, energy-based models, and language models.
