First Systematic Evaluation of XAI Support from Hierarchical Reinforcement Learning in Overcooked-AI

A recent investigation published on arXiv (2608.06381) marks the initial comprehensive assessment of explainable AI (XAI) support derived from a naturally explainable learned policy within a recognized benchmark. Utilizing the Hierarchical Ad Hoc Agents (HA^2) framework in Overcooked-AI, the team produced real-time explanations based on hierarchical subtask choices, conveyed through text or audio using an innovative trigger-based mechanism. In a study involving 38 participants, no notable performance differences were observed; however, those receiving explanations exhibited tendencies for quicker performance enhancements. Interestingly, audio explanations led to a significant decrease in the working-alliance bond with the agent, a phenomenon not seen with text, indicating that verbal explanations may increase cognitive load or affect social dynamics, potentially impeding collaboration. This research addresses the shortcomings of earlier XAI studies that depended on manually crafted policies in tailored environments, which restricted their applicability to advanced teaming research. The results underscore the significance of modality in XAI communication and the necessity for deeper exploration of how different explanation formats influence human-agent interactions.