We introduce a framework for in-context (zero-shot) inference/estimation of drift and diffusion functions underlying SDEs from empirical data of different dimensionalities.

We propose a training-free framework for sampling by exploiting multiscale dynamics and general diffusion paths.