A training technique where a model is exposed to its own previous predictions to reduce mismatch between training and inference behavior.