When Agents Imagine Together: Neuro-Symbolic World Models for Multi-Agent RL

Abstract: In multi-agent systems, what happens next often depends on what several agents do at the same time, which makes world models hard to learn from data alone. Yet much of this dynamics is already known. This talk presents NS-MAWM, a framework that injects partial symbolic rules into multi-agent world models. Rules predict only what they can justify and leave the rest to the network. Across four environments and three architectures, joint rules improve prediction by about 20% when agents interact frequently. A surprising lesson emerges: forcing rules into the model helps at first, but teaching them through regularization works better. It gives better final accuracy, more robustness to wrong rules, and stronger policies. So the question is not only what an agent knows, but how that knowledge should enter a learned model.