Run it back: Autoregression vs. Video Diffusion for End-Conditioned Weather Trajectories

The severity of low-likelihood, high-impact (LLHI) extremes is projected to increase in a warming climate. Sampling the range of precursors that lead to these events traditionally requires massive ensembles that scale poorly with lead time and rarity. While specialized rare-event sampling and ensemble-boosting methods are emerging, they do not demonstrate how a single extreme might arise from diverse synoptic precursors. Lin et al. 2026 recently demonstrated this is possible using end-conditioned trajectories from a video diffusion climate emulator (cBottle-video), but the resulting ensembles were structurally underdispersed, geostrophically inconsistent, and over a fixed window length. Notably, cBottle-video was evaluated as an off-the-shelf multi-task checkpoint trained with general masked conditional video diffusion (MCVD). Whether a dedicated reverse-time autoregressive model generates more physically credible backward trajectories remains an open question. Here, we train a backward-in-time probabilistic autoregressive model using a FourCastNet 3 backbone to generate systematically diverse, end-conditioned trajectories for two of these case studies. Although the approach is promising, rollout instability indicates that further refinement may be necessary.