Run it Back: Autoregression vs. Video Diffusion for Final-Conditioned Weather Trajectories
Abstract
Low-likelihood, high-impact extremes are projected to increase in severity and frequency in a warming climate.
Sampling the precursors that lead to these events traditionally requires massive ensembles that scale poorly with lead time and rarity, and typical rare-event sampling and ensemble-boosting methods do not show how a single extreme might arise from diverse synoptic precursors.
Recently, Lin et al.\ (2026) showed this is possible with final-conditioned 66-hour trajectories (trajectories constrained to end at an observed state) from cBottle-video, a pre-trained, off-the-shelf video diffusion climate emulator not explicitly fine-tuned for this task (Brenowitz et al., 2025).
However, its ensembles were underdispersed and geostrophically inconsistent, and it reproduces conditioned frames only approximately, hindering autoregressive rollout.
Here, we train Rewind-AR, a backward-in-time probabilistic autoregressive model with a FourCastNet~3 backbone (Bonev et al., 2025), and generate final-conditioned ensembles for the same three cases: the 2021 Pacific Northwest heatwave, Hurricane Ian, and Superstorm Sandy (in the Supplementary Information).
The leading modes of Rewind-AR's antecedent 500 hPa geopotential height explain far more of the across-member variance than cBottle-video's for both hurricane position and heatwave surface temperature, and Rewind-AR better reproduces the day-over-day heatwave intensification observed in ERA5.
Finally, geostrophic inconsistencies close to the conditioning frame seen in cBottle-video are not reproduced by Rewind-AR.
Nevertheless, up to 22\% of Rewind-AR members become numerically unstable within 66 hours, so improving rollout physical consistency and stability over long lead times is a natural next step.