One-Step Generative Surrogate Models via Block-Triangular Joint Drifting

This paper addresses the challenge of applying drifting-based one-step generative models to stochastic transition modeling. Standard trajectory data typically provides only one realized next state per observed current state, making it difficult to empirically approximate the conditional distribution. The authors introduce block-triangular joint drifting as a solution that enables learning from such single-sample trajectory datasets without requiring multiple realizations of next states for each current state.