Dynamic Belief Graph Learning Improves Theory-of-Mind Reasoning in LLMs

arXiv·medium signal

Standard LLMs fail Theory of Mind tasks because they have no mechanism to track how characters' beliefs evolve as narrative events unfold. This paper proposes training with dynamic belief graphs that maintain explicit, updateable belief state representations tied to events, improving ToM benchmark performance. Directly relevant to agent developers building systems that need to model other agents' or users' evolving mental states—particularly multi-agent coordination and dialogue systems.