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

ACM International Conference on Multimedia (ACM MM), 2026

CaM-Wolf framework

The Perceiver reads a player’s video, the Reasoner works through the evidence behind possible roles, and the Performer responds through an animated avatar.

Abstract

CaM-Wolf brings video, speech, and animated responses into a social deduction game agent. It combines a multimodal Perceiver with a Reasoner trained to distinguish causal evidence from misleading clues, then uses an avatar to communicate its decisions. The paper evaluates both gameplay and human-agent interaction.