CaM-Wolf: Causal-Aware Multimodal Agents for Social Deduction Games

Researchers introduced CaM-Wolf, a multimodal AI agent for the social deduction game Werewolf.

It combines video-based perception, a reinforcement-learning-trained causal reasoner, and an animated avatar.

In tests and a user study, CaM-Wolf outperformed prior agents and improved human-AI interaction quality.

The work moves social deduction agents beyond text-only systems toward richer, more engaging gameplay.