The Theory of Mind Utility: Formal Specification of a Mentalizing Mechanism

Researchers introduce Theory of Mind Utility (ToM-U), a formal computational framework for modeling how agents infer others' beliefs by tracking information access and credibility. The model uses directed graphs called Local Epistemic World Models to represent epistemic relationships and generates falsifiable predictions about mentalizing failures, advancing cognitive science theory beyond existing Bayesian and simulation-based approaches.

This research presents a significant theoretical advancement in computational cognitive science, formalizing how agents reason about others' mental states—a capability central to social cognition and decision-making. Rather than assuming belief states exist, ToM-U derives them through structured inference over Local Epistemic World Models, which track who knows what and in what order. This differs fundamentally from Bayesian Theory of Mind, which presupposes beliefs, and from simulation or theory-theory approaches that lack formal epistemological machinery.

The framework's contribution lies in its domain-agnostic architecture that positions mentalizing as upstream of goal inference and other social cognitive processes. By specifying explicit mechanisms for recursive mentalizing bounded by computational constraints, ToM-U generates directional, falsifiable predictions about when and why mentalizing fails—properties derived from structural properties rather than ad-hoc assumptions. This formalization enables testable hypotheses about cognitive limitations in tracking nested beliefs and information dependencies.

While primarily academic, this work has implications for AI systems requiring robust social reasoning capabilities. Applications extend to multi-agent systems, human-AI interaction design, and cognitive modeling in robotics. The formal apparatus provides a blueprint for engineering AI systems that track epistemic states more transparently than black-box approaches. The emphasis on bounded recursion and structured failure modes also informs how AI systems might scale mentalizing computations without combinatorial explosion.

• →ToM-U formalizes epistemic state inference through directed graphs representing agent knowledge and information access history
• →The framework generates falsifiable predictions about mentalizing failures from structural properties rather than auxiliary assumptions
• →ToM-U positions mentalizing as computationally upstream of goal inference and downstream social cognition
• →The model constrains recursive reasoning through bounded proliferation mechanisms addressing computational tractability
• →Applications extend to multi-agent AI systems requiring transparent social reasoning and human-AI interaction design