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.

Excited to be at #NeurIPS2024 all week! If you'd like to chat about: - Probabilistic & Bayesian ML/DL - Sampling & generation in discrete data - LLM & VLM alignment/safety - ML PhD opportunities at Purdue …or anything else, feel free to DM or email me!