<p>Task-oriented dialogue systems have found widespread application in various domains such as intelligent human-computer interaction, customer service, educational support, and medical consultation. These systems not only demand precise understanding of user intentions but also require efficient coordination between the user agent and the dialogue agent over multiple rounds of interaction. However, current models often employ “simultaneous updating” or “turn-based updating” strategies, which lack scientifically grounded incentive mechanisms in their decision-making processes. As a result, they frequently encounter conflicts and exhibit suboptimal collaborative efficiency. This paper integrates Multi-Agent Reinforcement Learning (MARL) techniques to address this issue and proposes a multi-agent incentive mechanism based on Stackelberg game theory. Specifically, the proposed method first models the dialog interaction process as a Stackelberg game, positioning the user agent as the leader. Through policy optimization, the behavior of the dialogue agent is directed, enabling them to adapt their strategies to the dynamic nature of the dialog, thereby better meeting user needs. The Actor-Critic algorithm is then used to update the strategy of each agent in real time, gradually guiding the system toward a Stackelberg game equilibrium. The experimental results show that the method proposed in this paper gradually approaches the Stackelberg game equilibrium through multiple rounds of iteration, achieving significant improvements in core evaluation metrics such as task completion rate, demand matching degree, and user satisfaction based on the F1 score. Specifically, compared to the baseline model, our method improves the task completion rate by 8.06%, the demand matching degree by 4.23%, and the F1 score by 7.3%. These results validate the effectiveness of the Stackelberg game theory in task-oriented dialogue systems.</p>