TOMVALLEY: EVALUATING THE THEORY OF MIND REASONING OF LLMS IN REALISTIC SOCIAL CONTEXT

1. There's a lack of comprehensive discussion of existing benchmarks. While the authors point out limitations in the previous theory of mind (ToM) benchmarks—such as focusing on static or independent mental states and lacking social settings—they do not adequately compare their work to existing studies. For example, benchmarks like MMToM-QA [1] and MuMA-ToM [2] explore interconnected mental states by asking about one mental state conditioned on another. Benchmarks like OpenToM [3] include narrative stories and character profiles, addressing the social context aspect. There are other benchmarks like EmoBench [4] and Infant Cognition Benchmark [5] that are missed in Table 2.

2. In Section 2.2, the paper introduces the concept of "process-level evaluation" but fails to provide a clear definition for it. This makes it difficult for readers to grasp what this term means and how it relates to the papers' objectives. The coherence of the paragraph can also be improved.

3. There's insufficient explanation of question types and categories. The benchmark includes items related to belief, emotion, intention, and action, and categorizes questions into types like understanding, influence, and transformation. However, these categories are only briefly mentioned without clear definitions or justification for their inclusion. Furthermore, a deeper analysis is needed to explain why LLMs' performance varies significantly across different question types, which could offer valuable insights into the models' strengths and weaknesses.

4. The authors mention that the benchmark includes compositional problems that require rigorous multi-hop reasoning to arrive at the correct answer. However, they do not clearly define what these problems entail or how they were constructed to necessitate such reasoning. Providing a detailed explanation and concrete examples would help readers understand their importance and how they challenge LLMs differently than simpler questions.

5. While the benchmark incorporates social backgrounds, character profiles, and relationships, and claims that humans effortlessly use this information in ToM reasoning, the authors don't clearly explain how this information affects the answers to the questions. It remains unclear whether these details are essential for arriving at the correct answers or if they simply add diversity to the questions. A better explanation is needed to illustrate how social context influences the reasoning process and why it is crucial for evaluating ToM capabilities in LLMs.

1. While ToM is an interesting and promising field, TOMVALLEY largely builds on prior work, offering incremental improvements rather than a completely new approach. The overall structure and evaluation strategy closely resemble past studies, making this contribution feel less innovative. As I mentioned, the dataset combines features such as plot, character profiles, relationships (like OpenToM), and elements of social location and dynamic mental states (like BigToM). Similar to these works, TOMVALLEY also uses LLMs to generate richer scenes from provided structured information, using these scenes as labeled synthetic data for testing (where the generated scene serves as the input, and the structured information as the label). The evaluation method is also very similar to BigToM. While following the choices of previous works is reasonable, the authors overstate their contribution by claiming: “previous works (1) most evaluations focus on a static mental state after several social scenarios, ignoring the changes in mental states across different scenarios; (2) they mainly consider independent mental states, whereas different kinds of mental states (beliefs, intentions, and emotions) and actions can influence one another in real life; (3) social settings and character profiles are absent in their evaluations, even though humans can effortlessly obtain and utilize this information in ToM reasoning processes."

2. Though TOMVALLEY has created a large dataset, its reliance on synthetic data introduces inevitable noise. The authors evaluated answers across five criteria, with around 90% of responses meeting each criterion individually. However, it would be helpful to know how many answers meet all five criteria simultaneously, as this would provide a more holistic view of response quality. (See questions below)

3. Additionally, the design of certain questions can be ambiguous. For instance, in a dialogue context, a question like “What is Angela Hwang's belief?” could be unclear without explicit contextual information, making it difficult for users to discern the exact belief being asked about. (See questions below)

4. The analysis of model limitations also feels somewhat surface-level. It suggests that models struggle due to issues like “lost in the middle” phenomena and a lack of necessary ToM reasoning abilities. While these are valid points, they echo discussions already present in prior ToM research referenced by TOMVALLEY. Adding fresh insights or diving deeper into unique findings from this specific dataset could make the analysis more impactful and distinctive.

• Abstract:
  • Can be more specific about what exactly the authors are testing with the benchmark.
• In the introduction, it is not clear what exactly the questions are testing, this could be made much clearer. The Introduction and the abstract are very vague. They could be much more specific: 1) What are you testing 2) How exactly is the data generated 3) How do you validate the data 4) How do llms perform? Maybe using figure 1 as a reference would be useful
• Section 3: I am not convinced that the authors address the problem of circularity. The evaluation of a conversation being coherent to the mental states present and the dynamics of emotions in a social scenario is quite involved. The model generating these scenarios should then have a rich model of human mental states (this has not been validated in previous work).
  • Step 1: The way the profiles and names are collected could be moved from the appendix to the main text. Why are occupations pooled by gender? These could reinforce stereotypes.
  • Step 2: How are initial and final mental states decided upon?
  • Step 3: 1) How do you ensure that the dialogues generated by an LLM are faithful to the mental states? 2) How do you ensure that the conversation obeys the dynamics of the mental states (decided in step 2) and that these correspond to true dynamics of human mental states? While the authors collect human judgements for these, are coherence, authenticity and dynamism mostly dependent on an LLM being good at being good enough at understanding social contexts?
  • How are the answer options for the questions generated? How is the correct answer picked? How do you ensure that all the other options are incorrect? For example, “how does … change across all social scenarios?” seems like a broad open-ended question.
  • In general, how do you ensure that there is no problem of circularity? That the capability being evaluated is not being used while generating the data?
• Confidence intervals are missing throughout the paper
• Why are myers-briggs personality traits used? These lack scientific validity and soundness. Why aren't alternatives like the big 5 / OCEAN used?
• An actual example like in fig16 would make the paper much clearer.
• It is not clear, what precise capabilities the benchmark is testing and why the construct is valid.
• The limitations are buried in the appendix, and should be addressed in the main draft.
• How many annotations per question do you collect? For both human experiments?

Minor:

It would be nice to see other closed source models like gemini, claude! But this is not necessary.

In the title “realistic social context” → “realistic social contexts”

Line 250: toM → ToM

Line 301: five 5-Likert → the five questions on a 5 point Likert scale

Line 789 qomplexity → complexity

The design of character profiles has been mentioned in the OpenToM literature [1]. The five constructed scenarios, when connected, are quite similar to a long conversation, as in FanToM [2], where ToM capabilities of LLMs are also assessed in a dialog format, and the characters' belief states evolve as the conversation progresses. (I acknowledge that this paper's settings are more grounded in real social contexts, distinguishing it from FanToM.)

For influence-type questions, such as Scenario1->Scenario2, would the model's performance significantly improve if only these two scenarios were provided? Does context length greatly impact the model's performance?

In scenarios resembling real social contexts, interactive evaluation may be a better approach than directly feeding conversations into LLMs.

[1] Xu H, Zhao R, Zhu L, et al. OpenToM: A Comprehensive Benchmark for Evaluating Theory-of-Mind Reasoning Capabilities of Large Language Models[J]. arXiv preprint arXiv:2402.06044, 2024.

[2] Kim H, Sclar M, Zhou X, et al. FANToM: A benchmark for stress-testing machine theory of mind in interactions[J]. EMNLP 2023.

• The primary weakness of this paper is that all evaluations are conducted by presenting 71 questions within a single prompt (Figure 20). This setup, which is also used for CoT evaluation with claims that CoT does not enhance performance, introduces numerous confounding variables, making it unsuitable for benchmarking performance accurately. For example, results from the questions in the beginning will definitely impact later responses. Moreover, CoT is not suited for coming up with batch of answers. This is probably why the CoT does not show performance improvement. Since all of the results and analyses are based on this setup, I find it difficult to correctly interpret them. I strongly encourage the authors to rerun their experiments on each question independently.
• The benchmark is entirely generated by an LLM, but human validation was conducted on only 5% of the dataset—a very limited proportion. Since LLMs are known to struggle with theory of mind tasks, human validation is strongly recommended, especially given the complexity of the benchmark design, which involves multiple steps in the generation pipeline. I would suggest running at least 33% of the data. Ideally, I would encourage 100% of the data.