A phenomenon where a model's test performance suddenly improves long after training loss has plateaued.