Sample Efficient Generative Model for Molecular Dynamics Trajectories via Twisted Sequential Monte Carlo

We study conditional generation of molecular dynamics trajectories beyond unconditional Boltzmann equilibrium sampling from $\propto e^{-U}$. The motivation is inference-time path conditioning: given an initial frame together with constraints such as terminal states, intermediate frames, masks, or general trajectory-level rewards, the goal is to sample a full trajectory with the correct dynamical law. We formulate this problem as path-space inference under a Markov reference process and develop a hierarchy of twisted Sequential Monte Carlo methods built on learned score-based proposals. The resulting algorithms separate learning the reference dynamics from imposing new rewards at inference time, require only sample access to equilibrium frames and reference trajectories, and remain asymptotically exact as the number of particles $K \to \infty$ with controlled variance.