Repeatedly applying a function until it converges to a stable value, used here for test-time computation in looped models.