Block-triangular joint drifting enables one-step generative surrogate models for stochastic trajectories

chaumian · x · 2026-09-23

New arXiv paper (Geissler et al.) introduces block-triangular joint drifting, addressing the data problem that trajectory datasets typically give only one realized next state per observed state. A projected drift field acts on the joint distribution of consecutive states; the block-triangular architecture preserves the current-state marginal while its second component directly samples the conditional distribution of next states.

The surrogate generates stochastic trajectories with one model evaluation per time step, no auxiliary generative steps. Experiments show accurate marginal and trajectory statistics and favorable accuracy-cost tradeoffs versus deterministic, diffusion-, flow-, and distillation-based baselines.

Original post →

More from Research

Research channel →