Marylou Gabrié's arXiv review merges generative AI with Monte Carlo sampling

Sampling high-dimensional probability distributions is a core challenge in scientific computing, spanning Bayesian inference, statistical physics, and molecular simulation. Classical methods like Markov chain Monte Carlo and tempering often fail when distributions are multimodal or plagued by metastable states. In a new tutorial review on arXiv (arXiv:2608.07648), Marylou Gabrié explores a rapidly emerging paradigm: using generative models—specifically normalizing flows and diffusion models—not as data-driven tools, but as flexible probabilistic samplers for distributions known only up to a normalization constant.

The review catalogs early breakthroughs in this hybrid field, highlighting two major directions: exact samplers built on generative models, and training strategies that work without labeled data. Gabrié positions these methods as a bridge between machine learning and computational statistical physics, offering a potential escape from the intrinsic limitations of classical enhanced sampling techniques. While the paper does not attempt an exhaustive survey, it selects key ideas and evaluates their strengths and weaknesses. Accessible to both physicists and ML researchers, the review serves as a practical entry point for scientists looking to leverage AI-era tools for one of computation's oldest problems—efficiently exploring complex probability landscapes.