Slowly Annealed Langevin Dynamics:
Theory and Applications to Training-Free Guided Generation

(8 May 2026) The paper is now available at arXiv 

(27 Apr 2026) The initial version was presented at the 2nd DeLTa 2026 Workshop.

The code is available at the GitHub project.

This work proposes VA-SALD, a new theoretically grounded framework for training-free guided generation with pretrained generative models.

@article{nitanda2026SALD,

  title = {Slowly Annealed Langevin Dynamics: Theory and Applications to Training-Free Guided Generation},

  author={Nitanda, Atsushi and Bu Dake and Lyu, Yueming and Veeravalli, Tanya},

  journal={arXiv:2605.07950},

  year={2026}

}