What's this about?
An autoregressive video-generation method that trains the model on its own rollouts (not just ground-truth clips), so it can extrapolate far past its training length without drifting into mush.
What exactly is "exposure bias"?
Models trained only on real data never see their own mistakes; at generation time errors compound — each slightly-off frame conditions the next. Training on self-generated histories teaches recovery. Top HF paper of the day (29↑).
Why is it worth noting?
Long-horizon stability is the wall between video clips and video *worlds* — same battlefront as yesterday's ABot-World-0.