Seq-Flow: Efficient Probabilistic Forecasting with Self-Rollout Error Control

Seq-Flow is a conditional flow model that updates a distribution over future trajectories by transporting samples from a previous forecast distribution to an updated one using an ODE. It leverages self‑rollout training, where a moving‑average copy of the model generates forecasts that initialize subsequent flows, thereby controlling error propagation across recursive updates. Experiments on particle‑accelerator beam spill forecasting show a 65% reduction in CRPS under a few‑NFE sampling budget, and the model remains accurate over more than 400 consecutive updates while staying competitive on fluid‑dynamics tasks.

Machine-generated by The Flow from the publisher's headline and feed description — not written or checked by a human. The full article lives at arXiv Machine Learning.

• diffusion