FLAIR · project
MDNS: Masked Diffusion Neural Sampler via Stochastic Optimal Control
Sampling discrete configurations becomes difficult when the state space is large and probability concentrates around several separated modes. MDNS studies this setting when the target probability mass function is available only up to a normalizing constant.
A control view of masked diffusion
MDNS trains a neural sampler using masked diffusion and stochastic optimal control for continuous-time Markov chains. Its learning objectives align probability measures over stochastic paths, connecting the generation process to the desired terminal distribution. This supplies a principled framework for learning transitions through a discrete state space.
Evaluating discrete samplers
The paper compares sampling accuracy and computational behavior across target distributions with different statistical properties, including high-dimensional and multimodal settings. Ablations examine the roles of the training choices. The resulting contribution extends the connection between control and generative sampling to discrete variables, relevant to statistical physics, machine learning, and combinatorial problems.
Publication
Masked Diffusion Neural Sampler via Stochastic Optimal Control
Yuchen Zhu, Wei Guo, Jaemoo Choi, Guan-Horng Liu, Yongxin Chen, and Molei Tao
NeurIPS 2025