Rational Mentalizing Model

• Rational Mentalizing Model is a framework that computationally infers latent mental states from behavior, integrating goals, beliefs, and plans.
• It employs Bayesian inverse planning, top‐down predictions, and error-feedback mechanisms to model cognitive processes.
• Empirical evaluations show high correlations between model predictions and observed behavior in both human and AI decision tasks.

Searching arXiv for recent and relevant papers on “Rational Mentalizing Model” and closely related Theory-of-Mind formalisms. Using the arXiv search tool to verify the cited papers and related formulations. Rational Mentalizing Model denotes a family of formal accounts in which mentalizing is treated as a rational computational process over latent social variables rather than as an informal capacity label. Across recent work, the term is used for Bayesian inverse planning over goals, beliefs, and plans; control-theoretic architectures with top-down prediction and bottom-up error feedback; utility-based policies for deciding when to observe, intervene, or teach; computational-level mechanisms that construct and evaluate Local Epistemic World Models; and normative tests of whether artificial agents satisfy identity, redundancy, and counterfactual-consistency constraints when reasoning about other agents or about architectural clones (Ying et al., 2024, Freire et al., 2019, Gurney et al., 10 Jun 2026, Mukherjee et al., 2024, Gurney, 15 Jun 2026).

1. Scope and representational commitments

The central representational commitment of Rational Mentalizing Models is that behavior is generated from latent mental variables that can be formalized and inferred. Depending on the formulation, these variables include goals or desires, beliefs, plans, intended actions, rewards or preferences, action costs, and more general epistemic states. In the Bayesian theory-of-mind lineage, the observer infers coherent sets of goals, beliefs, and plans from observed actions, typically under a Boltzmann-rational or soft-optimal planning model (Ying et al., 2024). In bounded-rational extensions, the observer remains rational but mentalizes an agent who may have mistaken goals, mistaken plans, and mistaken actions (Alanqary et al., 2021). In multimodal settings, latent states can include beliefs, goals, preferences or rewards, and action costs synthesized from language and vision into a symbolic planning domain (Ying et al., 20 Jun 2025).

Different formulations also differ in what counts as the target of mentalizing. In game-theoretic and multi-agent reinforcement-learning settings, the target is another strategic agent whose policy, reward structure, or next action must be predicted (Freire et al., 2019, Oguntola et al., 2023). In explanation-selection accounts, the target is a natural-language belief statement whose explanatory strength depends on accuracy, informativity, and causal relevance to observed action (2505.19376). In computational-level epistemic models, the target is a Local Epistemic World Model built from ordered information access histories, observability constraints, and credibility (Gurney et al., 10 Jun 2026). In silico-centric evaluations, the target is another AI agent, including an architectural clone, and the issue is whether the focal system can recognize identity-based redundancy in interventions (Mukherjee et al., 2024).

A compact comparison makes the family resemblance clear.

Formulation	Core representation	Primary computation
Bayesian ToM	goals $g$, beliefs $b$, plans	inverse planning
DAC ToM	beliefs $b_t$, predictor state $W$	top-down prediction and error feedback
ToM-U	LEWM graph $$W=\langle A,O,E,\text{BLS},\text{obs},\text{cred}\rangle$$	candidate generation and filtering
Silico-centric RMM	clone states $s_i=(\theta,x_i)$	redundancy and identity invariance
Causal engagement model	$C, IA, OT, S, AT, RS, ToM, EA$	when to engage ToM

This variety implies that “Rational Mentalizing Model” is not a single standardized architecture. A plausible implication is that the term functions as a higher-level research program: mentalizing is formalized as structured inference or decision-making over latent social states, with explicit assumptions about what the observer knows, what the target agent knows, and what counts as rational use of that information.

2. Bayesian inverse planning and semantic grounding

A dominant formulation treats rational mentalizing as Bayesian inverse planning. In its compact form, the observer infers a latent goal from an observed trajectory by computing

$$P(g_0 \mid \tau, s_0) \propto P(\tau \mid g_0, s_0) P(g_0).$$

The full likelihood can then marginalize over temporary goals, plans, and action slips, as in

$$P(\tau \mid g_0, s_0) = \sum_{g_{1:T}} P(g_{1:T} \mid g_0) \sum_{p_{1:T}} P(p_{1:T} \mid g_0, g_{1:T}, s_0) \sum_{a_{1:T}} P(a_{1:T} \mid p_{1:T}, s_0) P(s_{1:T} \mid s_0, a_{1:T}),$$

which is the form used to model mistaken goals, resource-bounded planning, and execution errors (Alanqary et al., 2021). In related work on grounded belief language, action likelihoods are generated by a Boltzmann-rational planner over estimated plan costs,

$$\pi(a \mid s, g) \propto \exp(-\beta \cdot \hat Q_g(s,a)),$$

and exact Bayesian filtering is used to maintain posteriors over goals and world states from action sequences (Ying et al., 2024).

This inverse-planning machinery is also used to ground the semantics of belief attributions. In the belief-language framework, natural-language statements are translated into a fragment of first-order epistemic logic and evaluated against the posterior over mental states. The graded truth of a statement is defined by posterior expectation:

$b$0

Under the veridicality assumption, this reduces to an expectation over final belief states, which is how the framework explains gradedness and compositionality in human belief attributions (Ying et al., 2024).

Empirically, this formulation gives a much better fit to human goal and belief attributions than pure logical deduction, non-mentalizing baselines, and mentalizing that ignores the role of instrumental plans. In the doors-and-keys gridworld study, “Full BToM with $b$1” reached Pearson $b$2 for belief statements and $b$3 for goal attributions, whereas the “Heuristic Mentalizer” reached $b$4 under $b$5 (Ying et al., 2024).

The same inverse-planning backbone appears in pedagogical and social-learning settings. In the teaching model, the Bayes-Optimal teacher infers the learner’s latent knowledge state $b$6 from an observed trajectory and chooses the edge whose revelation maximizes expected improvement in learner value after replanning:

$b$7

In the social-learning model, the observer infers the other agent’s goal from past actions and compares the expected utility of observing with the expected utility of acting immediately (Harootonian et al., 2 Apr 2026, Ying et al., 12 Jul 2025). Across these settings, rational mentalizing is not merely attribution; it is attribution embedded in downstream decision-making.

3. Bounded rationality, predictability bias, and satisficing inference

A major development in the literature is the relaxation of classical near-optimality assumptions. One line of work extends Bayesian Theory of Mind to model boundedly rational agents who may make mistakes at three levels: “mistaken goals via temporary goal confusion, mistaken plans via resource-bounded planning with noisy heuristics, and mistaken actions via execution errors” (Alanqary et al., 2021). The agent-side generative program samples temporary goal corruption with probability $b$8, planning budgets $b$9, partial plans from stochastic A* guided by a heuristic $b_t$0, and slips with probability $b_t$1. Inference is then performed with Sequential Inverse Plan Search. In the reported human studies, the “Full bounded model (plan+action, ‘PA’)” achieved $b_t$2 in Doors, Keys & Gems, and the “Full bounded model (goal+plan+action, ‘GPA’)” achieved $b_t$3 in Block Words (Alanqary et al., 2021).

A second bounded-rationality modification concerns plan-space complexity. The “plan predictability oriented model” replaces the standard plan likelihood with a predictability term $b_t$4 and thereby biases inference toward plans that are easy to predict from observed partial paths. The model preserves Bayes’ rule,

$b_t$5

but redefines

$b_t$6

with

$b_t$7

This explicitly models the bias that people prefer predictable plans. In the reported behavioral experiment, the full inverse-planning model had overall correlation $b_t$8, whereas the plan predictability oriented model reached $b_t$9; performance of the full inverse-planning model declined with $W$0 with $W$1, while the predictability-oriented model showed $W$2 (Nakahashi et al., 2018).

A third line formalizes satisficing mentalizing as model selection among specialized Bayesian models. The Switching approach starts with the simplest model, monitors predictive surprise, and re-evaluates when surprise exceeds a threshold. The rule is

$W$3

with cumulative surprisal updated online and the re-evaluation threshold multiplied by $W$4 after each switch. The result is an explicit accuracy–efficiency trade-off: Switching is “about an order of magnitude faster than the Full model while achieving better accuracy,” and in the reported evaluation it outperformed all other models across conditions in average predictive log-likelihood (Pöppel et al., 2019).

These models shift the meaning of rationality. Rational mentalizing is no longer identified with assuming an optimal target. Instead, the observer is rational precisely by representing bounded planning, predictable plans, simplifying presumptions, or satisficing meta-reasoning when those assumptions better explain the target’s behavior.

4. Control-theoretic, utility-based, and causal formulations

Not all Rational Mentalizing Models are Bayesian inverse planners in the narrow sense. In Distributed Adaptive Control, Theory of Mind is implemented as a layered control architecture with a “Reactive Layer” and an “Adaptive Layer.” The adaptive layer maintains beliefs $W$5 or a prediction $W$6, issues top-down predictions, and receives bottom-up error:

$W$7

or, in probabilistic form,

$W$8

Adaptive updates are written as

$W$9

while action selection follows a best-response rule under beliefs. In five game-theoretic tasks, “probabilistic learning agents modeled on rational, predictive and other’s-model phenotypes perform better in game-theoretic metrics across tasks,” with important caveats in ballistic settings against Tit-for-Tat (Freire et al., 2019).

A more explicitly decision-theoretic use of mentalizing appears in social learning. There, the observer compares the utility of watching another agent with the utility of acting immediately. The social utility is

$$W=\langle A,O,E,\text{BLS},\text{obs},\text{cred}\rangle$$0

with

$$W=\langle A,O,E,\text{BLS},\text{obs},\text{cred}\rangle$$1

and the non-social utility is

$$W=\langle A,O,E,\text{BLS},\text{obs},\text{cred}\rangle$$2

The model predicts observation when $$W=\langle A,O,E,\text{BLS},\text{obs},\text{cred}\rangle$$3. In the treasure-hunt experiment, the Rational Mentalizing model correlated with human observation behavior at $$W=\langle A,O,E,\text{BLS},\text{obs},\text{cred}\rangle$$4, with human split-half reliability $$W=\langle A,O,E,\text{BLS},\text{obs},\text{cred}\rangle$$5, and its mean steps and cost per trial, $$W=\langle A,O,E,\text{BLS},\text{obs},\text{cred}\rangle$$6 and $$W=\langle A,O,E,\text{BLS},\text{obs},\text{cred}\rangle$$7, closely matched human performance, $$W=\langle A,O,E,\text{BLS},\text{obs},\text{cred}\rangle$$8 and $$W=\langle A,O,E,\text{BLS},\text{obs},\text{cred}\rangle$$9 (Ying et al., 12 Jul 2025).

A complementary line asks not how to mentalize but when. The causal engagement model treats ToM as a mechanism variable $s_i=(\theta,x_i)$0 driven by “Conflict Complexity,” “Information Asymmetry,” “Objective Tractability,” and “Sophistication,” mediated by variables such as “Accessible Tractability” and “Relative Sophistication.” The trigger is

$s_i=(\theta,x_i)$1

and acceptance is

$s_i=(\theta,x_i)$2

The primary outcome is “Epistemic Accuracy,” not direct task policy, and the model frames mentalizing as a resource-rational decision procedure for engagement under structural conditions (Gurney, 15 Jun 2026).

These formulations treat mentalizing as control, metareasoning, or value-of-information computation. A plausible implication is that Rational Mentalizing Models occupy the boundary between social inference and decision theory: the inferred mind matters because it changes action selection, intervention choice, or epistemic-resource allocation.

5. AI-centered, silico-centric, and multimodal formulations

Recent work has extended Rational Mentalizing Models from human targets to artificial agents and multimodal AI systems. The clearest normative formulation is the silico-centric testbed for clone reasoning. Here the relevant state is $s_i=(\theta,x_i)$3, with shared architecture and weights $s_i=(\theta,x_i)$4 across focal AI and clone, and performance summarized by $s_i=(\theta,x_i)$5. The key normative principle is redundancy under clone identity:

$s_i=(\theta,x_i)$6

together with knowledge invariance,

$s_i=(\theta,x_i)$7

and counterfactual invariance,

$s_i=(\theta,x_i)$8

The empirical result is a paradox. GPT‑4‑Turbo achieved near-perfect Strange Stories scores, with combined means across 250 trials of $s_i=(\theta,x_i)$9, $C, IA, OT, S, AT, RS, ToM, EA$0, and $C, IA, OT, S, AT, RS, ToM, EA$1, and regression showed “No significant effect on score differences.” Yet both focal instances produced instructions in all 250 trials, with mean lengths $C, IA, OT, S, AT, RS, ToM, EA$2 and $C, IA, OT, S, AT, RS, ToM, EA$3 characters, and the referee judged “Useful” in all 250 trials while preference was driven by $C, IA, OT, S, AT, RS, ToM, EA$4Entropy with coefficient $C, IA, OT, S, AT, RS, ToM, EA$5 (Mukherjee et al., 2024). The model thus defines rational mentalizing partly through non-intervention when identity and competence imply zero benefit.

LLM teaching studies use a different target but a closely related rational core. In the graph-teaching task, most models are best fit by the “Bayes-Optimal Teacher,” which infers the learner’s missing transitions and selects the edge maximizing expected value gain after replanning. The study reports that “most LLMs perform well, show little change in strategy over trials, and their graph-by-graph performance is similar to that of humans,” and that “prompt compliance does not guarantee better teaching decisions” because scaffolds “do not reliably improve later teaching on heuristic-incongruent test graphs and can sometimes reduce performance” (Harootonian et al., 2 Apr 2026). This is a teacher-side Rational Mentalizing Model: inference over another agent’s latent knowledge state is coupled to pedagogical action selection.

Multimodal versions push the framework beyond text-only settings. Language-Informed Rational Agent Synthesis uses a vision-capable LLM to synthesize a PDDL domain $C, IA, OT, S, AT, RS, ToM, EA$6, an agent model $C, IA, OT, S, AT, RS, ToM, EA$7, symbolic states from video, and then performs Bayesian inverse planning via Sequential Inverse Agent Modeling:

$C, IA, OT, S, AT, RS, ToM, EA$8

with

$C, IA, OT, S, AT, RS, ToM, EA$9

Across Food Trucks, Astronaut, and Doors-Keys-Gems variants, the model yields “human-like inferences and outperforms larger multimodal LLMs prompted end-to-end,” with correlations such as Food Trucks Belief $$P(g_0 \mid \tau, s_0) \propto P(\tau \mid g_0, s_0) P(g_0).$$0, Astronaut Rewards $$P(g_0 \mid \tau, s_0) \propto P(\tau \mid g_0, s_0) P(g_0).$$1, and DKG-Simple $$P(g_0 \mid \tau, s_0) \propto P(\tau \mid g_0, s_0) P(g_0).$$2 (Ying et al., 20 Jun 2025).

Together, these AI-centered formulations clarify that rational mentalizing in artificial systems is not exhausted by high benchmark accuracy. It also concerns identity-sensitive intervention, pedagogical inference over learner knowledge, and structured synthesis of task-specific world-and-agent models from language and vision.

6. Empirical landscape, explanatory selection, and open problems

Across the literature, Rational Mentalizing Models are evaluated not only by whether they recover hidden goals or beliefs, but also by whether they select the right belief statement, observation policy, or intervention. In the explanation-selection framework, a candidate belief attribution is scored by “accuracy, informativity, and causal relevance,” with causal relevance defined from necessity, sufficiency, and normality terms under interventions on belief states. The reported result is that “Causal alone: $$P(g_0 \mid \tau, s_0) \propto P(\tau \mid g_0, s_0) P(g_0).$$3–$$P(g_0 \mid \tau, s_0) \propto P(\tau \mid g_0, s_0) P(g_0).$$4 — best single factor,” whereas “Accuracy alone” and “Informativity alone” reached $$P(g_0 \mid \tau, s_0) \propto P(\tau \mid g_0, s_0) P(g_0).$$5 and $$P(g_0 \mid \tau, s_0) \propto P(\tau \mid g_0, s_0) P(g_0).$$6, and “Accuracy + Informativity” reached $$P(g_0 \mid \tau, s_0) \propto P(\tau \mid g_0, s_0) P(g_0).$$7 (2505.19376). This shifts attention from whether a belief is merely inferable to why a particular belief is selected as the explanation of behavior.

A related computational-level generalization appears in the Theory of Mind Utility. There, mentalizing is framed as construction and evaluation of candidate Local Epistemic World Models with cumulative confidence

$$P(g_0 \mid \tau, s_0) \propto P(\tau \mid g_0, s_0) P(g_0).$$8

early acceptance when $$P(g_0 \mid \tau, s_0) \propto P(\tau \mid g_0, s_0) P(g_0).$$9, and bounded termination at sophistication ceiling $$P(\tau \mid g_0, s_0) = \sum_{g_{1:T}} P(g_{1:T} \mid g_0) \sum_{p_{1:T}} P(p_{1:T} \mid g_0, g_{1:T}, s_0) \sum_{a_{1:T}} P(a_{1:T} \mid p_{1:T}, s_0) P(s_{1:T} \mid s_0, a_{1:T}),$$0 with low-confidence output otherwise. Recursive mentalizing is limited by

$$P(\tau \mid g_0, s_0) = \sum_{g_{1:T}} P(g_{1:T} \mid g_0) \sum_{p_{1:T}} P(p_{1:T} \mid g_0, g_{1:T}, s_0) \sum_{a_{1:T}} P(a_{1:T} \mid p_{1:T}, s_0) P(s_{1:T} \mid s_0, a_{1:T}),$$1

and failed attempts leave a structured residue $$P(\tau \mid g_0, s_0) = \sum_{g_{1:T}} P(g_{1:T} \mid g_0) \sum_{p_{1:T}} P(p_{1:T} \mid g_0, g_{1:T}, s_0) \sum_{a_{1:T}} P(a_{1:T} \mid p_{1:T}, s_0) P(s_{1:T} \mid s_0, a_{1:T}),$$2 that reduces future credibility on implicated edges (Gurney et al., 10 Jun 2026). This is a distinct notion of rationality: confidence accumulation, bounded proliferation, and residue-sensitive revision.

A further development is the use of mentalizing as intrinsic motivation in multi-agent reinforcement learning. In that formulation, each agent predicts other agents’ first-order beliefs and receives an intrinsic reward equal to the negative second-order prediction loss,

$$P(\tau \mid g_0, s_0) = \sum_{g_{1:T}} P(g_{1:T} \mid g_0) \sum_{p_{1:T}} P(p_{1:T} \mid g_0, g_{1:T}, s_0) \sum_{a_{1:T}} P(a_{1:T} \mid p_{1:T}, s_0) P(s_{1:T} \mid s_0, a_{1:T}),$$3

which is combined with extrinsic reward in a PPO or MAPPO objective (Oguntola et al., 2023). In the reported physical deception task, “2nd-order beliefs (good), adversary 1st-order” yielded good-agent performance $$P(\tau \mid g_0, s_0) = \sum_{g_{1:T}} P(g_{1:T} \mid g_0) \sum_{p_{1:T}} P(p_{1:T} \mid g_0, g_{1:T}, s_0) \sum_{a_{1:T}} P(a_{1:T} \mid p_{1:T}, s_0) P(s_{1:T} \mid s_0, a_{1:T}),$$4 versus “No beliefs” at $$P(\tau \mid g_0, s_0) = \sum_{g_{1:T}} P(g_{1:T} \mid g_0) \sum_{p_{1:T}} P(p_{1:T} \mid g_0, g_{1:T}, s_0) \sum_{a_{1:T}} P(a_{1:T} \mid p_{1:T}, s_0) P(s_{1:T} \mid s_0, a_{1:T}),$$5, while “Adversary 2nd-order, good 1st-order” improved adversary performance from $$P(\tau \mid g_0, s_0) = \sum_{g_{1:T}} P(g_{1:T} \mid g_0) \sum_{p_{1:T}} P(p_{1:T} \mid g_0, g_{1:T}, s_0) \sum_{a_{1:T}} P(a_{1:T} \mid p_{1:T}, s_0) P(s_{1:T} \mid s_0, a_{1:T}),$$6 to $$P(\tau \mid g_0, s_0) = \sum_{g_{1:T}} P(g_{1:T} \mid g_0) \sum_{p_{1:T}} P(p_{1:T} \mid g_0, g_{1:T}, s_0) \sum_{a_{1:T}} P(a_{1:T} \mid p_{1:T}, s_0) P(s_{1:T} \mid s_0, a_{1:T}),$$7 (Oguntola et al., 2023). This extends Rational Mentalizing Models from inference modules to reward-shaping mechanisms.

Several limitations recur across formulations. Task scope is often narrow: single batteries such as Strange Stories, discrete gridworlds, or specific matrix games (Mukherjee et al., 2024, Ying et al., 2024, Freire et al., 2019). Prompt sensitivity, few-shot framing, or meta-cognitive instructions may alter LLM behavior (Mukherjee et al., 2024, Harootonian et al., 2 Apr 2026). Enumeration or sequential Monte Carlo becomes difficult in richer state spaces (Alanqary et al., 2021, Ying et al., 2024). Many models rely on strong assumptions about veridical beliefs, explicit language rules, or soft-optimal planners (Ying et al., 2024, Ying et al., 20 Jun 2025, Ying et al., 12 Jul 2025). And the silico-centric results indicate that success on human-centric Theory-of-Mind benchmarks does not by itself establish rational mentalizing about other AI agents, especially under identity and redundancy constraints (Mukherjee et al., 2024).

The open problem that unifies the field is not the absence of mental-state variables, but the absence of a single mechanism that is simultaneously calibrated, computationally tractable, intervention-sensitive, and robust across human and non-human targets. Some frameworks emphasize inverse planning, others confidence over epistemic graphs, others resource-rational engagement, and others normative non-intervention under clone identity. The emerging picture is that Rational Mentalizing Model names a research agenda: formal mental-state inference coupled to explicit rational constraints on explanation, control, observation, teaching, and intervention.