Bayesian Filtering in Physical Systems via Test-time Trained Flow Matching

This arXiv paper introduces a novel approach to Bayesian filtering in physical systems using test-time trained flow matching. The method addresses the challenge of state estimation in high-dimensional systems with complicated posterior distributions. Unlike traditional particle-based representations that lose distributional information, this approach learns continuous posterior distributions through flow matching models. The work targets practitioners developing uncertainty-aware state estimation methods for complex dynamical systems.