Beyond Static Snapshots: Dynamic Belief Graphs for Causal Theory of Mind Reasoning

During a wildfire evacuation, human behavior rarely follows simple stimulus-response patterns. Residents may see flames yet delay departure for hours, seeking additional confirmation from neighbors or official channels. These delays emerge not from irrationality, but from complex belief dynamics: risk perceptions evolve, social signals accumulate, and multiple competing hypotheses about the threat interact and reinforce one another until some threshold triggers action. Capturing this process requires more than recognizing patterns in text; it demands modeling how latent mental states causally entangle across time.

In their paper "Learning Dynamic Belief Graphs for Theory-of-mind Reasoning," Chen et al. address this challenge by introducing a structured cognitive trajectory framework that moves beyond static belief representations. Their work exposes a critical limitation in current Large Language Model approaches to Theory of Mind (ToM), namely that treating beliefs as independent snapshots fails to capture the temporal coherence necessary for genuine mental state modeling. By representing mental states as dynamic belief graphs with learned interdependencies, the authors provide a principled method for tracking how beliefs evolve, interact, and ultimately drive decisions in high-stakes, uncertain environments.

The Static Belief Problem and Post-Hoc Rationalization

Most contemporary LLM-based approaches to Theory of Mind operate through inference-time prompting or latent-state models that sample belief distributions from frozen model priors. These methods treat beliefs as static, independent variables, inferring mental states at each timestep without enforcing consistency across time or accounting for causal dependencies between different belief dimensions. The result is a form of sophisticated pattern matching that often collapses into post-hoc rationalization: the model generates beliefs that plausibly explain observed actions, but these beliefs lack the temporal continuity and mutual constraints that characterize genuine human cognition.

This limitation becomes particularly acute in high-stakes domains such as emergency medicine, financial crises, or disaster response, where actions are sparse and delayed relative to belief formation. Psychological models like the Protective Action Decision Model (PADM) demonstrate that risk perceptions do not exist in isolation; they interact, reinforce, or suppress one another to maintain internal coherence. Empirical studies of wildfire evacuations show that belief updates are highly nonlinear, with residents integrating visual cues, official warnings, and social signals in complex ways that resist simple additive models.

When AI systems treat beliefs as independent and static, they produce mental models prone to semantic drift. Without structural constraints forcing beliefs to evolve consistently from previous states, the model can generate explanations that contradict earlier inferences or ignore the causal pathways through which observations modify latent states. The paper identifies this as a fundamental barrier to accurate action prediction: per-timestep belief inference that assumes independence cannot capture the underlying reasoning process, leading to error accumulation in trajectory forecasting.

Energy-Based Factor Graphs and Structured Learning

To address these limitations, Chen et al. propose representing mental states as dynamic belief graphs, where nodes correspond to latent beliefs and edges capture time-varying dependencies between them. The technical innovation lies in three interconnected components: a semantic-to-potential projection, an energy-based factor graph representation, and a variational learning objective.

The semantic-to-potential projection serves as the interface between LLM capabilities and structured probabilistic reasoning. Rather than treating LLM outputs as direct belief estimates, the model maps textualized probabilistic statements and semantic embeddings into unary and pairwise potential functions. This projection grounds language-derived evidence in consistent graphical-model updates, ensuring that beliefs remain probabilistically coherent even as they evolve. The energy-based factor graph then enforces consistency constraints across belief interdependencies, representing the interactions between different risk perceptions or intention states as energy functions that penalize incompatible configurations.

For action prediction, the authors introduce belief-conditioned action representations combined with action-specific self-attention mechanisms. This architecture captures the nonlinear modulation between belief states and decisions, acknowledging that specific combinations of beliefs jointly drive actions in ways that cannot be decomposed into independent contributions. The entire system is trained by optimizing an Evidence Lower Bound (ELBO) on action-trajectory likelihood, jointly learning the projection parameters, belief-graph potentials, and action model without requiring belief-level supervision.

This variational approach enables online belief updating at test time. As new observations arrive, the model updates the belief graph from past beliefs and current evidence, then predicts actions from the updated state. Evaluated on multiple real-world wildfire evacuation datasets, the framework significantly improves action prediction accuracy while recovering interpretable belief trajectories that align with human reasoning patterns.

From Pattern Matching to Causal Accountability

The contribution of this work extends beyond improved predictive accuracy. It highlights a philosophical distinction between models that pattern match on behavior and those that attempt causal accountability for mental states. Current LLM approaches often produce what might be termed narrative coherence: they generate plausible stories about why agents acted that explain the observations. However, without structural constraints enforcing temporal consistency and interdependency, these narratives remain unconstrained by the mechanics of belief formation.

The energy-based factor graph imposes what the authors call structural pressure on the inference process. By treating beliefs as nodes in a graphical model with learned pairwise potentials, the system cannot arbitrarily assign beliefs at each timestep;