The paper introduces MA-JEPA, a stochastic world model for multi-agent reinforcement learning that uses self-supervised joint-embedding prediction to improve sample efficiency.