cooperative-inverse-reinforcement-learning

This paper introduces a contextual-bandit team game with two-sided informational asymmetry for runtime human oversight of AI agents, characterizing gaps between team-optimal and myopic human oversight strategies.