AutoToM: Scaling Model-based Mental Inference via Automated Agent Modeling

Thank you for your insightful review and for recognizing AutoToM’s clear motivation and explanations, strong ablation studies, and the wide set of benchmarks and baselines that support its quality and significance. We appreciate your constructive feedback and will address your concerns as follows.

W1. The benchmark performance results in Table 1 and Figure 4 do not contain any error bars. The improved performance of AutoToM could be better established using multiple runs for each test case.

Due to budget constraints, we were unable to run every benchmark multiple times. However, for the most challenging benchmark in terms of baseline performance, MMToM-QA, we evaluated AutoToM using three different random seeds. The average performance had a mean accuracy of 82.56% with a standard error of 0.45%. This is consistent with the accuracy of 83.00% reported in Table 1 and indicates relatively stable performance across different runs. Similarly, we also ran o3-mini-high with three different random seeds. It had a mean accuracy of 65.94% with a standard error of 0.59%. This result shows that the evaluation is quite stable and our findings still hold after multiple runs.

Q1. Given the definition of the model utility in Eqn. (5), how does AutoToM avoid getting into local optima of models that are both overly simple and certain about the wrong answers to the queries?

This is indeed an insightful question.

1. We do not begin with an overly simple model—we start with an initial model proposal. For example, when addressing a question about belief, we start with models that incorporate beliefs. Utterances and actions are also explicitly considered during the initial model proposal. In practice, this initial model is usually rich enough and requires only minor adjustments over 1 or 2 iterations.

2. For AutoToM to be confident, the model needs to produce high action/utterance likelihood, P(a | relevant variables) or P(u | relevant variables), to fully account for the observed behavior. It’s difficult for the model to produce high action/utterance likelihood if crucial variables are missing. E.g., if the agent has only partial observations but belief is missing in the agent model, the model would not explain its behavior caused by perishable observability.

3. The only exception we observed in the experiments is that there were very few instances (2 out of 600 questions) in MMToM-QA, where a simpler model (MDP) is used when a more complex model (POMDP) would be more appropriate. This is caused by errors in the GPT-4o’s estimation of action likelihood. As a potential solution, AutoToM supports improving the model using human feedback (see Appendix B). We can also use MCMC for model adjustment to better escape local optima.

Q2. How does the performance of AutoToM vary when using different base LLMs?

To test AutoToM’s performance sensitivity to LLM backends, we conducted additional experiments using alternative models. Note that we used the same prompt for each backend LLM. Specifically, we replace the GPT-4o backend with Qwen3-235B (open-sourced), DeepSeek-V3 (open-sourced), and Gemini-2.5-flash (thinking disabled) on the most challenging MMToM-QA benchmark. Notably, AutoToM with any LLM as the backend outperforms the corresponding LLM performance by a large margin (Rebuttal Table 1). Crucially, we achieve this without extra prompt engineering.

Rebuttal Table 1: Model performance comparison on MMToM-QA. LLM indicates the model itself; AutoToM represents our method with corresponding model as the backend.

Thank you for your positive feedback — we're glad our clarifications were helpful. We truly appreciate the time and effort you've devoted to reviewing our work. We will incorporate the additional results into a future revision.

Thank you for your insightful review and for recognizing that AutoToM removes manual model specification with automated model discovery, generalizes to open‑domain and higher‑order reasoning, and is backed by thorough evaluations and case studies showing practical utility and interpretability. We appreciate your constructive feedback and will address your concerns as follows.

W1. Since AutoToM performs agent model discovery per query, it may incur higher inference latency compared to methods with static models. A comparison of inference time across methods would help contextualize this trade-off.

While AutoToM outperforms large reasoning models in terms of performance, it is also faster or comparable in real-world scenarios. For embodied tasks, AutoToM takes 30 seconds per timestep, which is only slightly slower than its backend LLM, GPT-4o (20 seconds per timestep). o3-mini, by contrast, requires 100 seconds per timestep, making it impractical for embodied decision-making. Additionally, we analyzed the inference speed and computational cost of AutoToM and reasoning models on the most computationally demanding ToM benchmark, MMToM-QA (due to long context). The comparisons, as shown in Rebuttal Table 1, show that AutoToM requires fewer or comparable tokens and inference time.

Rebuttal Table 1: Token cost and inference time comparison on MMToM-QA.

W2. A more detailed explanation or visualization of how the system alternates between variable and timestep adjustments would improve transparency and reproducibility.

Thank you for the constructive feedback. In the paper, Figure 1 overviews the AutoToM procedure; Figure 3 illustrates example hypotheses and calculations for Bayesian inverse planning; and Figure 5 shows a real example of AutoToM’s model adjustment and inference in a BigToM false-belief scenario. Appendix C.3 presents two real examples demonstrating variable adjustment (Example 1) and timestep adjustment (Example 2) in AutoToM.

Following your suggestion, we will add an additional example figure that clearly presents every element (each hypothesis, each calculation) and every step of the process (initial model → inference results → variable adjustment → inference results → timestep adjustment → …), with easily digestible annotations.

Q1. Cost Function Clarification

Thanks for this question. No, our complexity cost is C(M)=α∣M∣, where ∣M∣ counts only the number of latent mental variable types; it does not include timesteps. We start from the last timestep and only add earlier steps if the utility remains low, so parsimony is enforced by this conservative timestep expansion, and with the cost term.

Q2. Hi-ToM Performance

Challenges with higher-order tasks. For higher-order tasks, AutoToM samples one possible state based on b(s) at level l to approximate the state at level l - 1 as imagined by the agent at level l. This process is applied recursively down to level 0, where a standard automated BIP is performed. However, as recursion deepens, errors may accumulate: inaccuracies in belief estimation at any level or state can propagate downward, potentially affecting reasoning at level 0. To further improve the accuracy of recursive reasoning, we need to reduce errors introduced during BIP when sampling beliefs—specifically, by enhancing the accuracy of non-recursive belief inferences.

Other errors occur when AutoToM fails to recognize the relevance of certain mental variables, resulting in an insufficient model. This may be improved with better LLM backends, and we will also explore broader improvements at the framework level, such as learning-based approximations to replace the computationally expensive nested BIPs.

Deciding on higher-order inference. The initial model proposal assesses the required level of recursive reasoning for higher-order ToM inference. This level is more fixed and easier to determine early on compared to the necessary variables and timesteps, which we adjust dynamically based on inference uncertainty.

Q3. LLM Backend Sensitivity

To test AutoToM’s performance sensitivity to LLM backends, we conducted additional experiments using alternative models. Note that we used the same prompt for each backend LLM. Specifically, we replace the GPT-4o backend with Qwen3-235B (open-sourced), DeepSeek-V3 (open-sourced), and Gemini-2.5-flash (thinking disabled) on the most challenging MMToM-QA benchmark. Notably, AutoToM with any LLM as the backend outperforms the corresponding LLM performance by a large margin (Rebuttal Table 2). Crucially, we achieve this without extra prompt engineering.

Rebuttal Table 2: Model performance comparison on MMToM-QA. LLM indicates the model itself; AutoToM represents our method with corresponding model as the backend.

Thank you for your insightful review and for recognizing AutoToM’s novelty, state‑of‑the‑art results, and its generalization with human‑like confidence calibration. We appreciate your constructive feedback and will address your concerns as follows.

W1 & Q2. While AutoToM claims to automate model discovery, the underlying causal structure is fixed, limiting the scope of structural generalization. Can the system be extended to jointly learn both variable structure and causal dependencies, rather than fixing the structure a priori?

AutoToM jointly learns both variable structure and causal dependencies, but only within a space of plausible causal relationships. It automates agent model discovery by identifying the necessary mental variables and timesteps while preserving causal dependencies within a family of structures grounded in standard decision-theoretic primitives (state, observation, belief, action, goal). For example, in partially observable scenarios, belief depends on observation, which in turn depends on state; in fully observable scenarios, AutoToM simplifies this by having belief depend directly on state.

These causal dependencies are constrained to align with causal structures established in prior agent modeling works (e.g., Kaelbling et al., 1998; Gmytrasiewicz and Doshi, 2005) and cognitive studies (e.g., Baker et al., 2009; Baker et al., 2017; Ullman et al., 2009). For example, actions may depend on observations or beliefs, and optionally on goals, but observations cannot depend on the agent’s own actions, since observations reflect what the agent perceives, regardless of its own actions. Rather than introducing implausible causal dependencies, we selectively include and dynamically adjust dependencies within the model space. While extending beyond classical MDP/POMDP/I‑POMDP formulations, AutoToM's model space supports 30 configurations per timestep and allows for arbitrary levels of recursive reasoning, as discussed in Appendix A.5.

Empirically, AutoToM’s model structure generalizes across domains and modalities: we evaluated it on five ToM benchmarks, two cognitive studies, and an embodied assistance task, demonstrating strong overall results and stable behavior across diverse contexts, number of agents, modalities, and recursion level.

To further enhance generalization, domain-specific hypothesis libraries or other inductive biases for proposing variables can be incorporated, without changing the core algorithm. For instance, to enable affective reasoning—specifically, attitude inference as in the OpenToM benchmark—we extended the set of latent variables to include preference and attitude, and constructed new causal dependencies accordingly. This modification allowed AutoToM to perform affective inference effectively, achieving a significantly higher Macro-F1 score than GPT-4o (0.56 vs. 0.48).

W2. The framework does not learn the variables or their dependencies; instead, it selects and scores hypotheses using LLM outputs without probabilistic grounding.

The hypotheses are scored and selected using probabilistic inference rather than by LLM outputs alone. This selection process is illustrated in the key steps in Figure 3: Step 1 is formulated by our model and estimated with the help of an LLM, while Steps 2 and 3 do not involve the LLM.

Step 1: Estimating local conditionals: $P(b_1^t | b_1^{t-1}, o_2^t)=0.4$

In this step, the local conditional is estimated by an LLM. Given the clear dependencies, it's significantly more accurate for an LLM to estimate this than to tackle the entire problem directly.

Step 2: Calculating posterior probabilities: $P(v_i^t| X^{t_{s}:t}) \propto \sum_{V_{-i}^{t_s:t}} P(v_i^t, V_{-i}^{t_{s}:t}, X^{t_{s}:t})$

This step relies purely on probabilistic inference to explicitly compute the final probability distribution, without using the LLM.

Step 3: Model adjustment based on utility: $U(M, q) = R(M, q) - C(M)$

This step involves model adjustment based on the computed utility, which accounts for both the model’s confidence and cost. The LLM is not used in this step.

W3. Some of the evaluation gains may stem from prompt-level tuning or LLM scoring behavior, rather than principled modeling.

The success of AutoToM mainly comes from constructing agent models (Bayesian networks) and conducting Bayesian inference based on these constructed agent models. It substantially outperforms all previous prompting-based methods, such as SymbolicToM and SimToM, particularly in long-context and complex scenarios. These gains result from a principled agent modeling approach, decomposing the problem and clearly modeling mental dependencies.

Concretely, AutoToM extracts observable variables, constructs explicit mental models, and performs probabilistic inference using Bayesian inverse planning, grounded in cognitive studies. The agent models it constructs align with those commonly used by cognitive scientists. AutoToM iteratively refines these models to maximize confidence while minimizing complexity. The model adjustment is guided by a formal objective defined by our model utility (negative entropy of posterior distribution as reward and model complexity as cost), rather than relying on LLM outputs alone.

Q1. To what extent does AutoToM's performance depend on the specific LLM used? Could smaller models offer similar benefits with prompt engineering?

To test AutoToM’s performance sensitivity to LLM backends, we conducted additional experiments using alternative models. Note that we used the same prompt for each backend LLM. Specifically, we replace the GPT-4o backend with Qwen3-235B (open-sourced), DeepSeek-V3 (open-sourced), and Gemini-2.5-flash (thinking disabled) on the most challenging MMToM-QA benchmark. Notably, AutoToM with any LLM as the backend outperforms the corresponding LLM performance by a large margin (Rebuttal Table 1). Crucially, we achieve this without extra prompt engineering.

Rebuttal Table 1: Model performance comparison on MMToM-QA. LLM indicates the model itself; AutoToM represents our method with corresponding model as the backend.

Thank you for your insightful review and for recognizing that our paper addresses an important challenge and presents strong experiments. We will address your concerns as follows.

W1. The paper could be written more clearly, with a more digestible and representative example figure.

Thank you for the constructive feedback. In the paper, Figure 1 overviews the AutoToM procedure; Figure 3 illustrates example hypotheses and calculations for Bayesian inverse planning; and Figure 5 shows a real example of AutoToM’s model adjustment and inference in a BigToM false-belief scenario. Appendix C.3 presents two real examples demonstrating variable adjustment (Example 1) and timestep adjustment (Example 2) in AutoToM.

Following your suggestion, we will add an additional example figure that clearly presents every element (each hypothesis, each calculation) and every step of the process (initial model → inference results → variable adjustment → inference results → timestep adjustment → …), with easily digestible annotations.

W2. Some literature should be discussed more deeply.

Thank you for the suggestion. We will include more discussion on thought-tracing and other related works. We clarify how AutoToM differs below.

Key difference 1: explicit agent modeling. AutoToM constructs agent models (Bayesian networks) and conducts Bayesian inference based on them, specifying relevant mental variables and their causal relations. They describe the decision-making processes that lead to the observed agent behaviors. In contrast, thought-tracing maintains hypotheses only about the mental variable asked in the question, without the structural representations of other related mental variables. This explicit modeling yields higher accuracy on hard benchmarks and robustness to wording or superficial story changes (e.g., thought-tracing needs wording changes for MMToM-QA, while AutoToM does not). Moreover, error sources are easier to identify in AutoToM by inspecting its model structure and variables.

Key difference 2: inference complexity minimization. AutoToM discovers a minimally sufficient model per query: it proposes an initial model and adds variables and/or timesteps only when they improve utility. This avoids under-/over-modeling and scales with longer contexts, more agents, and deeper recursion. As shown in Exp. 3, this complexity control makes AutoToM much more efficient than reasoning models on downstream embodied assistance tasks. By contrast, thought-tracing only reweights hypotheses and does not adapt hypothesis complexity or timesteps.

W3. Would be good to add a token cost vs performance discussion.

Following your suggestion, we added a cost analysis of AutoToM vs. other methods on MMToM-QA, the most challenging dataset due to long context. The average number of tokens per question for AutoToM, o3-mini-high, and Gemini 2.0 Flash Thinking are 8.0k, 10.9k, and 8.8k, respectively. This shows that AutoToM outperforms large reasoning models while using fewer tokens.

Q1. Difference in reported baseline accuracies; adding FANToM.

The discrepancy in baseline performance on MMToM-QA is caused by:

1. Dataset modifications: The authors replaced object and room names with symbols like A, B, etc.

2. Modalities: The authors used a text-only version, while we use translated multimodal fusion. The text-only version is significantly easier—with the same model, o3-mini achieves 71.5% accuracy on text-only versus 64.7% on the multimodal version.

3. Number of questions: The authors evaluated on a subset of 200 questions, whereas we tested on the full benchmark of 600 questions.

Similarly, thought-tracing also used subsets for other benchmarks. Additionally, they relabeled the answers in BigToM.

FANToM: To further demonstrate AutoToM’s ability to solve false-belief tasks, we tested AutoToM on FANToM. We randomly selected a subset of 200 false-belief first-order questions with short contexts due to time and budget constraints.

Results. AutoToM, with a GPT-4o backend, achieved 72.7%, outperforming the GPT-4o baseline, which achieved 57.5%. AutoToM, with a Gemini 2.5 Flash backend, achieved 77.9%, outperforming the Gemini 2.5 Flash baseline, which achieved 38%. With either model as the backend LLM, AutoToM improves upon the original baselines.

Analysis. AutoToM is able to solve false belief questions by extracting the essential variables. In FANToM, AutoToM extracts the state of the conversation (the agents in the conversation, if the main agent is currently in the conversation, and the topics discussed), utterances, and observation of the main agent (depending on whether they are in the conversation or not) to infer belief. In contrast, the two baselines struggle to accurately extract and track the agent’s observation throughout the conversation.

Q2. In Figure 4, what cases does AutoToM excel in, and what challenges remain to improve "Level 2+ Belief" from 70% to 100%?

Where AutoToM excels. AutoToM is notably stable and scales effectively as contexts grow longer and more agents are introduced, outperforming large reasoning models in robustness. This is due to its explicit construction and adjustment of a causal agent model, followed by Bayesian inverse planning over latent mental variables, rather than relying solely on prompt heuristics.

AutoToM also performs better at goal inference, which all LLMs and large reasoning models struggle with. This is because AutoToM conducts explicit inverse planning, which reasons about how agents act conditioned on certain goal hypotheses and other relevant mental variables (e.g., beliefs). In contrast, LLMs and large reasoning models often produce incorrect goal inferences due to spurious action patterns, particularly when there is partial observability, where action plans are often not obvious due to incomplete information or false beliefs.

How to push toward 100%. Errors mainly arise from noisy multimodal extraction, imprecise probability estimates, and occasionally missing mental variables. For Level-2+ Belief, recursive belief sampling (from level l down to 0) accumulates errors as recursion deepens. Improving base (non-recursive) belief inference and BIP estimation will reduce errors. Other errors occur when AutoToM misses relevant mental variables, producing an insufficient model. We will mitigate this with stronger LLM backends and framework-level advances, including learning-based approximations to replace costly nested BIPs.

Q3. Would AutoToM also improve performance for benchmarks that sometimes require affective ToM such as OpenToM?

Thanks for the insightful question. We evaluated AutoToM’s affective ToM by extending the causal structure to include attitude and preference (all other components unchanged) and testing on all 596 OpenToM attitude questions.

Results. Following the OpenToM paper (Xu et al., 2024), we used Macro-F1 as the evaluation metric. The random baseline is 0.33. GPT-4o achieved 0.48, while AutoToM with GPT-4o backend outperformed it with a score of 0.56. AutoToM also approached the performance of the large reasoning model o3-mini-high (0.60), indicating its strong affective reasoning capability.

Analysis. Answering the attitude questions does not require inverse planning, since the model can just directly perform forward estimation of attitude based on observed events and preference. This explains why AutoToM performed similarly compared to o3-mini-high. This is consistent with results for other question types that do not require inverse planning, such as level 0 (no ToM) and level 1 action questions shown in Figure 4a. However, even in the case where inverse planning is not required, AutoToM still scores higher than its backend LLM (GPT-4o). We attribute this to AutoToM’s ability to extract and focus on variables that are causally relevant to the task, while filtering out spurious cues by design (see Xu et al., 2024, Section 2.5) that may mislead GPT-4o.

Q4. Why describe AutoToM as more robust? It would be important to specify exactly robust to what behavior.

Thanks for the suggestion. We define robustness as:

1. Consistent performance under input/task variations. AutoToM shows markedly lower variance than large reasoning models across different input/task variants, including (i) context length, (ii) number of agents, (iii) recursion level, and (iv) question type, as shown in Fig. 4. AutoToM is also more resilient to wording styles. E.g., thought-tracing requires rewording for MMToM-QA, whereas AutoToM needs no such modifications.

2. Human-like uncertainty estimation. AutoToM expresses human-aligned uncertainty in classic cognitive studies as shown in Experiment 2, unlike baselines.

3. Mental state inference useful for downstream tasks. AutoToM can be applied to downstream tasks such as embodied assistance. Specifically, while LLM baselines tend to propose uniform goal distributions due to distraction of long context, AutoToM concentrates probability on likely goals that are consistent with agent behavior, enabling helpful assistance.

4. Human-in-the-loop correction. AutoToM’s explicit agent models make reasoning interpretable and editable, enabling humans to provide feedback to correct errors when needed (Appendix B).

Why this robustness emerges:

1. Explicit modeling + BIP. AutoToM constructs explicit mental models and performs probabilistic inference using Bayesian inverse planning, grounded in cognitive studies. This reduces sensitivity to superficial story edits that can disrupt direct LLM prompting (as observed in ExploreToM).

2. Adaptive model selection. Rather than relying on fixed agent models, AutoToM adjusts its model based on a utility score that balances confidence and complexity. This enables AutoToM to adapt to different problems.

We will include this discussion in the revised Section 4.

Dear reviewer,

We sincerely appreciate your thoughtful engagement and for raising your score. We truly appreciate the time and effort you've devoted to reviewing our work. Your detailed suggestions have greatly helped us clarify our writing, strengthen our comparisons, and expand our discussion of related work, enhancing the quality of the paper :)