Training a model to operate on its own previous outputs rather than only ground-truth data to improve robustness.