Researchers have proposed a new framework called Delta for evaluating and improving deep reinforcement learning (DRL) agents. Delta focuses on both safety-critical failures and policy optimality, which is often neglected in current testing methods. It uses a two-phase approach: first, it tests the agent's decision-making policy to identify potential issues; then, it trains a new challenger agent using offline reinforcement learning and compares its performance with the original agent. The study demonstrates Delta's effectiveness across five environments, uncovering an average of 2,518 optimality issues and outperforming baseline methods by 50.2%. The findings suggest that safety testing datasets can be valuable for training competent DRL agents. --- Why it matters: This matters to researchers in AI because it addresses a significant gap in comprehensively evaluating DRL systems, which is crucial as they are increasingly deployed in real-world applications. By identifying both safety-critical and optimality issues, Delta provides a more complete picture of an agent's performance, enabling improvements that can lead to increased efficiency and user trust. Source: https://arxiv.org/abs/2608.22284