Unveiling value functions in social cognition with multi-agentinverse reinforcement learning

Social behavior requires individuals to consider not only their own goals but also those of others. Latent value functions that encode such goals can be recovered from behavior using inverse reinforcement learning in single-agent settings. However, extending it to multi-agent interactions is challenging, because value functions are defined over joint state spaces that grow exponentially with the number of agents. Existing approaches often manage this complexity by imposing strong structural assumptions about social interactions, thereby limiting their applicability and interpretability. Here we show that joint value functions governing social interactions can be effectively represented through value decomposition into individual value maps for each agent and low-dimensional interaction terms. We develop a multi-agent inverse reinforcement learning framework (MAIRL) to infer these representations from behavior. In mouse and primate social tasks, MAIRL reveals interpretable value maps that are conditioned on the distinct social roles animals play during group behavior. Together, these results establish MAIRL as an interpretable and scalable framework for identifying latent value representations guiding multi-agent behavior across species.