1 paper · 1 filter
Ryan Po, Eric Ryan Chan, Changan Chen +1
Autoregressive video models are promising for world modeling via next-frame prediction, but they suffer from exposure bias: a mismatch between training on clean contexts and infere…