Bayesian Theory of Mind (BToM)

• BToM is a framework that models an agent's mental states (beliefs, desires, intentions) by inverting a generative planning process.
• It employs full Bayesian inverse planning alongside specialized satisficing models to balance predictive accuracy with computational efficiency.
• A surprise-driven switching strategy adapts model inference to uncertainty, enhancing performance in varied grid-world navigation tasks.

The Bayesian Theory of Mind (BToM) framework formalizes the process by which an observer infers an agent’s hidden mental states—specifically, beliefs, desires/goals, and intentions—from observed actions. BToM inverts a generative model of the agent’s planning process to infer these mental states probabilistically, enabling systematic reasoning about intentional action in scenarios involving uncertainty. Core BToM models have been evaluated for their predictive performance and computational tractability, particularly in contexts where humans navigate grid-world environments under varying uncertainty conditions. A central challenge addressed in the literature is reconciling inferential precision with bounded computational resources, motivating "satisficing" models that pursue high accuracy at minimal cost (Pöppel et al., 2019).

1. Formalization and Notation

BToM operates on joint inferences over latent variables describing the agent:

• $A_t = \{a_1, \ldots, a_t\}$: sequence of observed actions up to time $t$,
• $G$: set of possible goals (e.g., colored exits in navigation tasks),
• $B_g$: set of possible goal-beliefs $b_g$ (e.g., color–location assignments),
• $B_w$: set of possible world-beliefs $b_w$ (e.g., map layouts, "true-map" or "unknown"),
• $P(g), P(b_g), P(b_w)$: priors over goals, goal-beliefs, world-beliefs (usually uniform),
• $P(A_t \mid g), P(A_t \mid b_g), P(A_t \mid b_w)$: likelihoods of action sequences under each mental state hypothesis,
• $P(g \mid A_t), P(b_g \mid A_t), P(b_w \mid A_t)$: posteriors upon observing $t$0.

Desire/intention is instantiated by $t$1, while two kinds of beliefs are modeled: the goal-belief ($t$2) and world-belief ($t$3).

2. Full Bayesian Inverse Planning

2.1 Forward Model

The agent, given $t$4, selects $t$5 via a Boltzmann-softmax over utilities:

$t$6

where $t$7 is the rationality parameter and $t$8 is (minus) the distance from the post-action state to $t$9, according to $G$0 and $G$1.

2.2 Inference via Bayesian Marginalization

The posterior predictive over next actions given past behavior is:

$G$2

Posteriors are updated at every step:

• Goal:

$G$3

• World-belief:

$G$4

• Goal-belief (with sensor updates):

$G$5

2.3 Computational Complexity

With $G$6, $G$7, $G$8, each time step evaluates 192 mental-state combinations, requiring repeated shortest-path computations and softmax evaluations. This renders exact inference intractable as task complexity increases.

3. Specialized and Satisficing Models

To mitigate intractability, three specialized models are constructed by conditioning on knowledge of one or more latent variables:

Model Label	Latent Variables Conditioned	Complexity (Combinations)
TWG (True World & Goal)	$G$9	4
TW (True World)	$B_g$0 unknown	$B_g$1
TG (True Goal)	$B_g$2 unknown (freespace)	4
Full BToM	$B_g$3 unknown	192

The TWG model assumes the agent knows the true map and exit assignments; only the goal is inferred. TW assumes knowledge of the map but not the exits. TG assumes knowledge of the exit assignment but not the map, defaulting to a freespace world-belief under path uncertainty. Each model reduces inference complexity by removing latent variable sums, but risks accuracy loss when its conditional assumptions do not match the true generative process.

4. Surprise-Driven Switching Strategy

A rule-based meta-controller monitors model adequacy via surprise measures on observed actions. The mechanism operates as follows:

1. Surprise Measures.
   • Negative log-likelihood: $B_g$4,
   • Distance-to-mode: $B_g$5, with $B_g$6.
2. Switching Criterion.
   • Track cumulated surprise $B_g$7. If $B_g$8, evaluate all specialized models on the action history and switch to the one with lowest accumulated surprise. The threshold $B_g$9 is increased to reduce hypersensitivity after each switch.
3. Meta-algorithm.
   • Initialize in TWG. Switch among $b_g$0 as needed, always selecting the minimal-cost model exhibiting sufficient predictive adequacy.

This approach does not constitute explicit hierarchical Bayesian inference; instead, switching operates as a satisficing heuristic closely aligned to bounded rationality.

5. Experimental Evaluation

5.1 Task and Participants

• Six distinct grid-world (13×20) layouts with four uniquely colored exits.
• Three primary uncertainty conditions:
  • No Uncertainty (NU): full map and one exit’s color revealed.
  • Destination Uncertainty (DU): full map and all exit positions, exit colors hidden until line-of-sight.
  • Path Uncertainty (PU): only local neighborhood revealed, with one exit’s color shown.
• 122 participants performed navigation; 110 completed all mazes, yielding 687 valid trajectories.

5.2 Behavioral Analysis

Optimality was measured by path efficiency (excess length over shortest path). Behavior aggregated across trials displayed:

• NU: direct homing,
• DU: nearest-neighbor search,
• PU: greedy but error-prone homing.

5.3 Model Comparison Metrics

• Predictive Accuracy: Mean negative log-likelihood ($b_g$1) of human actions under each model.
• Computational Cost: Wall-clock time (ms per trajectory on a 3.5 GHz Xeon).
• Satisficing Performance: Joint trade-off of accuracy and computation, plus switch event counts.

5.4 Quantitative Findings

• Accuracy: TWG best in NU; TW best in DU; TG best in PU. Full BToM showed the worst predictive accuracy, attributable to higher entropy and less confident predictions. The switching strategy yielded highest overall predictive performance by selecting appropriate sub-models dynamically.
• Pairwise Wins: The switching approach outperformed each specialized model in >90% of trajectories.
• Timing: TWG ~1 ms, TG ~1.5 ms, TW ~4 ms, Switching ~10 ms, Full BToM ~100 ms. Switching incurred only 10× the minimal specialized model cost and 10% the cost of full inference.
• Switch Events: Average model switches per trajectory were NU ≈ 0.5, DU ≈ 3–4, PU ≈ 4–5.

6. Satisficing Trade-Offs and Implications

Satisficing, in the sense of Simon (1955), seeks efficient "good-enough" inference instead of exact Bayesian optimality. In this context, full BToM is overly general, computationally costly, and insufficiently discriminative. Specialized models are fast and accurate when their priors fit the scenario, but brittle otherwise. The switching strategy—by adaptively escalating inference cost only under high surprise—achieves nearly optimal accuracy per trial with substantially lower computational requirements compared to full Bayesian inference.

Key conclusions include:

• The necessity of tractability-flexibility trade-offs in practical BToM inference.
• The value of surprise-triggered model selection as a meta-level mechanism approximating human-like bounded rationality.
• The observed empirical dominance of a satisficing controller over both full and fixed-hypotheses approaches, in diverse uncertainty scenarios (Pöppel et al., 2019).

7. Conclusions

Empirical and computational analyses show that full Bayesian ToM models, while theoretically robust, are often impractical due to excessive computational demand and over-generalization. Specialized "clamped" BToM models achieve high predictive accuracy under matching conditions but fail in mismatched contexts. A lightweight, surprise-sensitive switching strategy across these specialized models provides robust, efficient, and satisficing mentalizing across uncertainty profiles. This aligns well with plausible cognitive mechanisms in human Theory of Mind reasoning and represents a significant advance in both the principled modeling and practical emulation of social inference (Pöppel et al., 2019).