Future advanced AI systems may learn sophisticated strategies through reinforcement learning (RL) that humans cannot understand well enough to safely evaluate. We propose a training method which avoids agents learning undesired multi-step plans that receive high reward (multi-step "reward hacks") even if humans are not able to detect that the behavior is undesired. The method, Myopic Optimization with Non-myopic Approval (MONA), works by combining short-sighted optimization with far-sighted reward. We demonstrate that MONA can prevent multi-step reward hacking that ordinary RL causes, even without being able to detect the reward hacking and without any extra information that ordinary RL does not get access to. We study MONA empirically in three settings which model different misalignment failure modes including 2-step environments with LLMs representing delegated oversight and encoded reasoning and longer-horizon gridworld environments representing sensor tampering.

Many machine learning applications rely on large datasets that are conveniently collected from existing sources or that are labeled automatically as a by-product of user actions. However, in settings such as content moderation, accurately and reliably labeled data comes at substantial cost. If a learning algorithm has to pay for reward information, for example by asking a human for feedback, how does this change the exploration/exploitation tradeoff? We study this question in the context of bandit learning. Specifically, we investigate Bandits with Costly Reward Observations, where a cost needs to be paid in order to observe the reward of the bandit’s action. We show that the observation cost implies an $\Omega(c^{1/3}T^{2/3})$ lower bound on the regret. Furthermore, we develop a general non-adaptive bandit algorithm which matches this lower bound, and we present several competitive adaptive learning algorithms for both k-armed and contextual bandits.