Steering a generative model's output during inference using external constraints or observations, rather than only during training.