Reward Model Overfitting and Generalization

Challenges of reward hacking, proxy objective misalignment, and ensuring the reward model generalizes beyond training data.