acceptodds
Under review as a conference paper at ICLR 2027

OBToM: Ordinal Bayesian Theory of Mind from Likelihood Judgments of Language Models

Abstract

Language models often answer theory-of-mind questions inconsistently when asked to reason directly about mental states. Methods that embed a language model within Bayesian inverse planning instead require verbalized probabilities, which are poorly calibrated, or token log-probabilities, which many deployed models do not expose. We propose Ordinal Bayesian Theory of Mind (OBToM), which separates likelihood judgment from probability estimation. A story is parsed into discrete variables on a fixed mental-state graph linking prior belief and perception to belief, and belief and goal to action. The language model is asked only ordinal questions: given two hypothetical conditions and outcomes, it judges whether one conditional likelihood is somewhat or much larger, whether the two are approximately equal, or whether the story provides no basis for judgment. An ordered-logit model converts these judgments into conditional probability tables with a shared low-rank parameterization, so a few fixed comparisons per story also constrain table rows that were never queried. Belief revision is gated by perceptual access: an agent who missed a change keeps the earlier belief. Mental states are then inferred through exact Bayesian inference over the graph, using a maximum a posteriori estimate of the tables or Hamiltonian Monte Carlo samples. We compare OBToM with direct prompting, perspective-taking, timeline, thought-tracing, and verbalized- and token-probability inverse-planning methods on the same backbone. With either estimator, OBToM is the most accurate on BigToM, FANToM, ExploreToM, and first- and second-order ToMi, and significantly so against every baseline on BigToM, ToMi, and ExploreToM. It uses at most five language-model requests per question, against 24.5 to 177.3 for thought-tracing and automated inverse planning.

Primary area:
neurosymbolic & hybrid AI systems (physics-informed, logic & formal reasoning, etc.)
Keywords:
theory of mind, Bayesian inverse planning, large language models, ordinal elicitation, probabilistic graphical models, false-belief reasoning, uncertainty quantification

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 4 reviews, above 27% of the others.
Reviewer_MFmhOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper focuses on the problem that LLMs answer ToM questions inconsistently, while Bayesian ToM methods require reliable probabilities. The authors identify that existing methods depend on unavailable token probabilities or unreliable verbalized numerical probabilities. To address this gap, the authors propose OBToM, which asks an LLM for ordinal likelihood comparisons and convert them into Bayesian-graph probability tables, and uses exact inference to select the answer. They found that this framework achieves better accuracy on multiple ToM benchmarks.

Critical strengths

The design of the framework is very coherent. It turns simple ordinal LLM judgements into structured Bayesian ToM inference without relying on the model's internal statistics. The experimental results are really promising.

Critical weaknesses
  1. I generally appreciate the framework for its coherent combination of multiple components. The authors showed promising results and compared model requests. However, the authors did not conduct any ablation study for the components, nor did they conduct a probability calibration evaluation. Therefore, it is unclear whether the gains come from ordinal estimation, low-rank sharing, the hard-coded perception gate, or their combination. This is very important to confirm the validity of the desired framework.

  2. The experiments, while their results seem promising, are conducted only on one LLM backbone. I would be more convinced if the framework demonstrates more general performance improvement on other LLM backbones. Additionally, I would like to see some other tests beyond answer accuracy. For example, among answers assigned probability p, what percentage of them are actually correct? For a ToM question with known Bayesian graph CPTs, how far could the framework recover them? A case study might also be helpful.

Questions for authors

see weaknesses

Reviewer_SmtgOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Ordinal Bayesian Theory of Mind, a framework that uses conditional likelihood comparisons from large language models for theory of mind reasoning. A story is parsed into a fixed probabilistic graph containing initial beliefs, perception, updated beliefs, goals, and actions. The LLM makes graded comparisons of outcome likelihoods under hypothetical conditions rather than reporting numerical probabilities or answering the final question. An ordered logit observation model and a shared low-rank parameterization are used to estimate conditional probability tables. Perception gating ensures that events an agent does not perceive leave its earlier belief unchanged. The system then performs Bayesian inference over the graph using MAP estimates or HMC samples. Experiments show strong performance on several false-belief reasoning tasks and fewer LLM requests on average than some methods that search over hypotheses, although the gains vary across tasks.

Critical strengths
  1. The method does not simply assign fixed probabilities to verbal judgments. Instead, it treats graded comparisons as noisy observations of conditional log-likelihood ratios and distinguishes approximate equality from insufficient information. For comparisons across contexts, it retains the normalization terms for each CPT row rather than comparing unnormalized scores. These choices give the local LLM judgments a clear statistical interpretation and allow them to enter probabilistic inference without requiring token probabilities. This is a promising alternative to asking LLMs to report absolute probabilities.

  2. The fixed mental-state graph links perception to belief updates and uses beliefs and goals to predict actions, keeping the actual state of the world separate from the agent’s beliefs. Comparison prompts do not include the final question, and the information available to each factor is restricted. Perception gating keeps earlier beliefs unchanged when an agent does not observe an event. Related questions can therefore be answered using the same joint probability model, and predictions can be traced back to the parsing results, local judgments, and probability factors.

  3. The experiments compare several prompting and probabilistic inference methods using the same backbone. OBToM shows clear gains on BigToM, ToMi, and ExploreToM. The authors use a paired bootstrap that resamples stories, accounting for correlations among questions from the same story, and apply Holm correction for multiple comparisons. These results provide strong evidence for the effectiveness of the complete framework.

Critical weaknesses
  1. Apart from extracting participant arrivals and departures in FANToM, the system relies mainly on deterministic parsers written for the benchmark templates. The ToMi parser does more than extract entities. It also recovers object movements and who observed them, and treats the initial object location as an observed belief. For forward belief questions with a known initial belief and a perception gate set to zero, Equation (4) already determines the updated belief without ordinal judgments, low-rank parameterization, or posterior sampling. The gains of the full system over existing baselines therefore do not establish ordinal probability modeling as the main source of improvement.

  2. The method uses fixed cutpoints, a noise scale, and parameter priors to translate verbal strength into a probability scale. The model is internally consistent, but this does not show that the estimated probabilities are more reliable. The experiments mainly report final answer accuracy. They do not directly compare ordinal judgments, numerical probability reports, and fixed mappings from verbal categories under the same graph structure and inference procedure, nor do they evaluate the calibration of the final probabilities.

  3. The separate contributions of the shared low-rank parameterization and posterior sampling remain unclear. Section 4.3 defines an independent-row ablation using the same judgments and inference procedure, but no results are reported in the main paper or appendix. The model fixes the rank at D=1 while retaining local residuals for each row. Its benefits need to be demonstrated through controlled comparisons, not inferred from the shared low-rank form alone.

  4. The experiments use only one backbone, and most parsers rely on the benchmark generation templates. The gains over the strongest baseline on FANToM are not statistically significant, and accuracy on MMToM-QA is lower than that of ThoughtTracing. These findings do not mean the method lacks value, but its advantages cannot yet be generalized to theory of mind reasoning more broadly.

Questions for authors
  1. Could you hold the parser, candidate states, perception gates, and inference graph fixed, and compare deterministic belief tracking, simple CPTs, an independent-row ordinal model, and the full OBToM model? Could you break down the results by question type and by whether the initial belief is observed, to distinguish cases already determined by the rules from those that require probabilistic inference?

  2. Under the same inference framework, does ordinal elicitation produce more stable estimates, or better probabilistic scores than numerical probability elicitation? How sensitive are these results to the cutpoints, prior scales, rank, and backbone? Does HMC improve the quality of the predicted probabilities over MAP, and what do the convergence diagnostics show?

  3. When parameters are not shared across factors, can the same exact posterior predictive distribution be obtained by first computing the posterior mean of each CPT and then running graph inference once? What would this imply for the current implementation, which performs inference for each posterior sample, and for its computational efficiency?

Reviewer_Mu7eOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

OBToM is a pipeline for solving textual ToM benchmarks using ordinal LLM judgments and Bayesian ToM inference over a constructed mental-state graph. It models conditional probabilities using CPTs with a shared low-rank component, allowing information to be shared across rows to support probability estimation. In the binary, single-change setup, the method obtains five comparative likelihood assessments from the LLM. CPT parameters are then estimated using MAP or sampled using HMC, and answers are obtained through exact graph inference. The graph is constructed using fixed templates and benchmark-specific deterministic code, except for FANToM, where an additional LLM query extracts presence information. The parsing extracts perceptual access and any specified goals. Perception information is used to gate belief updates, so agents retain their previous beliefs when they do not perceive a change.

Across six benchmark settings, the OBToM pipeline scores competitively, yet the lack of controlled ablations and limited analysis of the individual components make it hard to determine where the improvements come from. Although Section 4.3 mentions an independent-row CPT ablation, its quantitative results are not reported in the paper or appendix. The paper also reports fewer LLM queries, but the benchmark-specific parsing already performs some of the processing that would otherwise be assigned to the LLM. These query savings therefore cannot be attributed to ordinal elicitation alone. Given the differences in preprocessing across methods, the hand-coded parsing requirements, and the missing controlled comparisons, it is hard to know how much the parsing, ordinal likelihood judgments, or shared low-rank modeling individually contribute to the gains over baselines. In its current form, my assessment falls between a clear rejection and a weak rejection, as major revisions are needed to substantiate the paper’s central claims.

Critical strengths

S1. End-to-end solution to textual ToM benchmarks. The paper presents a complete pipeline tailored to existing textual ToM benchmarks and demonstrates competitive performance across multiple benchmark families, including first- and second-order ToM questions. The complete pipeline achieves competitive accuracy while requiring substantially fewer LLM requests than hypothesis-based approaches.

Critical weaknesses

W1. Missing controls and incomplete ablation reporting. The experiments do not adequately isolate the contributions of ordinal elicitation, deterministic preprocessing, graph structure, and shared low-rank modeling. A matched ablation should replace ordinal judgments with numerical probability elicitation while retaining the same parsing, perception gates, graph, CPT parameterization, and inference procedure.

Although ordinal elicitation avoids explicit numerical probability reports, it still requires the LLM to distinguish likelihood strength through judgments such as “somewhat” versus “much more likely.” A direction-only comparison baseline and sensitivity analysis over the ordered-logit thresholds would help establish whether this additional strength information and its assumed numerical interpretation are beneficial. A symbolic baseline using the same parsed information would also help establish what the LLM-based stages contribute. Although Section 4.3 mentions an independent-row CPT ablation, its results are not reported in the paper or appendix. Consequently, the reported accuracy gains and query savings cannot be attributed specifically to the proposed ordinal interface.

W2. Incremental novelty and limited generalizability. The method combines established ideas from Bayesian ToM, comparative LLM elicitation, and tensor-factorized conditional probability models. The contribution appears to lie mainly in combining existing techniques. However, stronger controlled comparisons are needed to establish the benefits of its specific design choices. The reliance on benchmark-specific parsing rules also limits the demonstrated scope of the approach. Applicability to open-ended narratives or embodied ToM remains unestablished, and extending the system to these settings would likely require substantial changes to the semantic front end.

W3. Clarity of the pipeline and preprocessing assumptions. I found the interactions between deterministic parsing, perceptual gating, LLM judgments, CPT estimation, and final inference difficult to follow. Although Appendix B describes the parsers, the main presentation would benefit from a more explicit walkthrough showing which information is extracted, which belief transitions are fixed by rules, and which quantities are estimated from LLM judgments. This distinction is particularly important for understanding how much reasoning is performed by the benchmark-specific code and interpreting the comparison of LLM query counts.

Questions for authors

Q1. Could you report the quantitative results of the independent-row CPT ablation described in Section 4.3 and compare them with the shared low-rank model? Have you evaluated whether sharing improves estimation when all comparisons involving selected learned CPT rows are intentionally withheld?

Q2. How does ordinal elicitation compare with numerical probability elicitation when the parser, perception gates, graph, CPT parameterization, and inference procedure are held fixed? A comparison of accuracy and LLM request counts under comparable elicitation budgets would help isolate the ordinal interface’s contribution.

Q3. Could you report the performance of a symbolic baseline using the same parsed story representation and perceptual-access information, without LLM likelihood judgments or CPT estimation? This would clarify the additional contribution of the LLM and probabilistic stages beyond deterministic preprocessing.

Q4. How sensitive is the method to parsing errors or incomplete semantic extraction? Have you evaluated its robustness on narratives whose wording or events depart from the supported templates?

Q5. Does distinguishing “somewhat” from “much more likely” improve performance over direction-only comparisons? How sensitive are the results to the ordered-logit thresholds, the priors on CPT parameters, and the bias toward retaining the previous belief?

Reviewer_WCGfOpenReview ↗
Rating
3
Confidence
2
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a new framework for LLMs to answer theory of mind questions by separating (comparative) likelihood judgment from probability estimation. In principle, it requires less information from LLMs to rank two options by their relative likelihoods than to assign exact probabilities (when you have exact probabilities of both options, you can then rank them; but when you can rank two options, you still do not have exact probabilities of those two options). This paper uses this statistical insight for LLMs to make likelihood judgments. The posterior probabilities are then separately estimated by using Bayesian inference with a gated belief revision process (e.g., LLMs are forbidden to update their beliefs if they are not exposed to more information). Empirically, this paper achieves superior performance on ToM tasks over conventional methods, including direct prompting, perspective-taking, etc.

Critical strengths

It identifies an important unsolved problem. LLMs do not have a consistent ToM of humans. This is likely because they do not have human cognitive states covered in their pretraining dataset ("cognitive dark matter"); hence, their assistance can be limited when users revise their beliefs upon receiving new information (which is often not available to LLMs). This paper identifies this set of problems, and I believe such problems, once better solved, have significant implications for human-AI interaction.

It provides a novel, principled ToM method based on statistical insights. As shared, in principle, it requires less information from LLMs to rank two options by their relative likelihoods than to assign exact probabilities (when you have exact probabilities of both options, you can then rank them; but when you can rank two options, you still do not have exact probabilities of those two options. Such a principled method bears strong positive empirical results once implemented well, as demonstrated by this paper.

It outperforms current empirical baselines for ToM tasks. Results look strong.

Critical weaknesses

The paper does not provide ablation experiments that separate the effects of different components. The lack of controlled experiments raises questions about its rigour. The experimental results seem very strong, but it's unclear whether the strong results are attributable to the stated contributions of this paper: (i) the separation of likelihood judgment from posterior estimation; and (ii) the comparative judgment for likelihood. Other explanations may account for the strong results: the parser's format, etc., in OBToM may be tailored for such ToM tasks, for instance.

Questions for authors
  • Would you like to provide any controlled experiment results that attribute the strong results to the stated mechanisms of this paper?

  • Would you like to share the limitations of this paper in your own words? In particular, I am curious to learn about any uncertainties you've had with experimental setups and any mixed results, or any factors in the experiments that may make your results hard to replicate.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.