Abstract
This review systematically explores the potential of the active inference framework in illuminating the cognitive mechanisms of decision-making in repeated games. Repeated games, characterized by multi-round interactions and social uncertainty, closely resemble real-world social scenarios in which the decision-making process involves interconnected cognitive components such as inference, policy selection, and learning. Unlike traditional reinforcement learning models, active inference, grounded in the principle of free energy minimization, unifies perception, learning, planning, and action within a single generative model. Belief updating occurs by minimizing variational free energy, while the exploration-exploitation dilemma is balanced by minimizing expected free energy. Based on partially observable Markov decision processes, the framework naturally incorporates social uncertainty, and its hierarchical structure allows for simulating mentalizing processes, providing a unified account of social decision-making. Future research can further validate its effectiveness through model simulations and behavioral fitting.