Nouveau Recherche PDF, HTML, DOCX et bien d'autres formats.

Prépublication · 2026

Run it Back: Autoregression vs. Video Diffusion for Final-Conditioned Weather Trajectories

Jerry Lin, Elizabeth A. Barnes — Monde

Source : arXiv

Résumé

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 diffu…

Actions

Citation