Solid grasp of core RL training objectives and loss functions, including temporal-difference and Bellman error losses (Q-learning, DQN), policy gradient objectives (REINFORCE, actor-critic advantage estimation), and clipped surrogate objectives (PPO, TRPO), with an understanding of when each applies and how they behave in training. Developing reward models, feedback models, and objective functions that translate noisy, sparse, delayed, or implicit signals into reliable model training and evaluation targets, and diagnosing and mitigating reward hacking and feedback loops in deployed systems.