TLDR Fixed-feature regression learns from one phase to locate the transition where neural networks fail, and a sample-free bound limits its error to power-law growth.

TLDR Training on clean past frames improves autoregressive video diffusion by compensating for causality violations that let future frames influence the past.