Model Autophagy

In synthetic-data training, the risk phenomenon in which outputs generated by a model are repeatedly used to train the next generation of models, leading to distribution shift and performance degradation.