This research addresses a critical gap in the development of LLM Multi-Agent systems: the ability to diagnose failures efficiently. The introduction of "Automated Failure Attribution" as a formal research problem is a significant step forward, as it provides a structured approach to a challenge that has long been a bottleneck in system iteration and optimization. The Who&When dataset and the proposed methods offer a foundation for future work, even if current accuracy rates are modest. The study...