Reinforcement learning agents face significant challenges with sparse rewards, which lack intermediate feedback to guide exploration and attribute success rewards to relevant trajectory parts. To address this, researchers have proposed automating potential-based reward shaping using vision language model guidance. This approach aims to provide more effective guidance for exploration and correct attribution of rewards, mitigating the risk of reward hacking. By leveraging vision language models, the method can generate more informative and relevant reward signals, enabling agents to learn more efficiently. The use of potential-based reward shaping guarantees that the shaped reward is consistent with the original goal, preventing agents from exploiting auxiliary signals. This development has significant implications for the field of reinforcement learning, as it can improve the effectiveness of agents in complex environments1. So what matters to practitioners is that this breakthrough can lead to more robust and reliable reinforcement learning systems, with far-reaching consequences for areas like policy, security, and workforce dynamics.