BELIEFS: A Hierarchical Theory of Mind Model based on StrategyInference

Abstract

Theory of Mind (ToM) refers to the ability to infer another agents latent mental states, such as intentions, beliefs, and strategies, to predict their behavior. A core feature of ToM is its recursive structure: individuals reason not only about what others think, but also about what others think about them. Existing computational models typically assume that ToM Level-0 (L0) agents rely on a fixed heuristic (e.g., Win-Stay Lose-Shift, WSLS), an assumption that fails to capture the diversity of non-mentalizing strategies humans actually use. Here we introduce BELIEFS, a probabilistic ToM framework that infers latent L0 strategies directly from behavior using a Hidden Markov Model (HMM) that enables flexible tracking of elemental strategies and dynamic switches between them without relying on a single predefined heuristics. A second HMM tracks the beliefs about the Opponents ToM level and dynamic changes therein. We evaluated BELIEFS across four classic dyadic games (Matching Pennies, Prisoners Dilemma, Bach or Stravinsky, Stag Hunt) under varying learning rates and volatility of strategy switches. Predictive performance, quantified via cumulative negative log-likelihood (NLL) of the opponents choices, was compared against chance and a WSLS-based ToM model, with BELIEFS consistently achieving superior accuracy across conditions. Strategy inference was assessed using trial-wise confusion matrices and Cohens {kappa}, revealing robust above-chance classification. Additionally, separability of ToM levels across games indicated that competitive games are particularly informative for distinguishing recursive reasoning from deterministic L0 strategies. The model also successfully tracked the opponents recursive reasoning depth (i.e. ToM level) by distinguishing action sequences generated by L0 versus L1 opponents. Parameter-recovery analyses confirmed reliable estimation of core transition parameters. Together, these results show that BELIEFS provides a flexible, computationally grounded account of human ToM, jointly inferring surface-level strategies and recursive reasoning, with applications to modeling adaptive behavior in dynamic interactive environments.