Peter Potaptchik, Cheuk-Kit Lee, Michael S. Albergo
Tilt Matching is a scalable algorithm for sampling from unnormalized densities and fine-tuning generative models using stochastic interpolants without needing reward gradients.
This paper introduces a new algorithm called Tilt Matching, which is designed to efficiently sample from complex probability distributions and improve generative models. The method leverages stochastic interpolants, mathematical tools that help in understanding how to transition between different states or distributions. Unlike traditional methods, it doesn't require calculating gradients of rewards, making it more scalable and easier to implement. The authors demonstrate that Tilt Matching performs well on various tasks, like sampling in physics simulations and fine-tuning image generation models.
