“Reinforcement learning can be used to evaluate training data quality by revealing backdoors or exploits in an environment when a model learns to "game the system."”