Steering a generative model toward outputs that maximize a reward function, often used to satisfy constraints or optimize downstream objectives.