Bayesian networks appear ideal for modelling social games like mafia/werewolf, since they're centered around updating beliefs.

But it's not enough to model the predicted game state, you also have to model the predicted beliefs of other agents.

And since those agents also have to do this, it explodes in complexity.