<p>Taxi dispatching, as a critical task in urban transportation systems, aims to optimize order matching and reduces empty mileage by reallocating idle vehicles from surplus supply areas to high-demand regions, thereby improving system efficiency and economic benefits. In recent years, Multi-Agent Reinforcement Learning (MARL), with its ability to model regional collaborative behaviors and dynamic optimization, has become a prominent technical approach to solving this problem. However, existing methods still face challenges in practice. On one hand, the heterogeneity and non-stationarity between regions at the urban scale increase the complexity of policy learning, making it difficult to achieve efficient coordination of dispatching behaviors. On the other hand, the dispatching process highly depends on temporal context, and the lack of model interpretability restricts practical applications and policy tuning. To address these issues, we propose a region-level MARL dispatch framework, where regions are treated as agents and the action space is heterogeneous. This framework models regional state perception and cross-region vehicle migration behaviors to achieve joint optimization among multiple agents. Building on this, we design and implement a visual analytics system, <i>DpLens</i>, which combines multi-view strategy analysis and key state identification to support users in exploring the evolution of dispatch strategies and the collaboration mechanisms among agents, from both macro and micro perspectives and across multiple dimensions. Through case studies of typical urban dispatch tasks and user studies, we demonstrate the effectiveness of our approach in enhancing model interpretability, assisting strategy optimization, and improving system reliability.</p> Graphical abstract <p></p>