The objective function is the mathematical function a training process minimizes or maximizes — loss functions, reward signals, and preference scores are all types of objective. In supervised learning, cross-entropy and mean squared error are common objectives; in reinforcement learning, the objective is cumulative reward; in RLHF, it combines a reward model score with a KL-divergence penalty. The choice of objective deeply shapes what a model learns, and misaligned objectives are a central concern in AI safety research.