Scalar step size for gradient updates; too high diverges, too low slows training.