Abstract: Imitation learning offers an appealing alternative to reward design, but its extension to multi-agent settings is fundamentally complicated by strategic interactions and equilibrium constraints. In this talk, I present recent advances in learning equilibria directly from data across different multi-agent regimes. I begin with the mean-field setting, showing that standard single-agent imitation learning approaches can fail to recover equilibrium behavior. I then turn to general Markov games, where I characterize the limitations of non-interactive imitation learning and present an interactive algorithm achieving near-optimal sample complexity for learning Nash equilibria from demonstrations. Finally, I consider function approximation, showing that structured representations enable provably efficient multi-agent imitation learning beyond the tabular case. Together, these results characterize when and how equilibrium behavior can be identified and learned from data in multi-agent systems.