Autoregressive video models generate clips block by block, with each new block conditioned on the ones before it. This sequential dependency means mistakes can accumulate, especially when the model is asked to generate in few steps rather than many.
The new paper, posted on arXiv, introduces Adversarial Gradient Forcing, a training approach that builds on Self Forcing and its successors. Those methods train the model on its own rollouts rather than only ground-truth sequences, which helps close the gap between training and inference.
The abstract does not specify experimental results or architectural details, but the framing suggests the goal is to make few-step autoregressive generation more stable. The title indicates an adversarial component, but the source does not elaborate on how it is applied.