The authors demonstrated this by injecting Gaussian noise into pre-trained models, noting that performance degraded more sharply the longer the model was trained.