“Reinforcement Learning from Human Feedback (RLHF) is often referred to as a bandit problem because it involves choosing one action (a completion) and observing the outcome, rather than a sequential decision-making process.”