Can one step predict as well as many steps in physics simulations?

Forecasting weather, turbulence, or other complex dynamics traditionally requires many small computational steps. Recursive Flow Matching trains a generative model to generate accurate trajectories in just 1–4 steps by enforcing self-consistency across different discretization scales, reducing accumulated errors. On scientific benchmarks, it matches multi-step solvers 20× faster while cutting mean squared error by 15%.