Entering Real Social World! Benchmarking the Theory of Mind and Socialization Capabilities of LLMs from a First-person Perspective

My biggest concern concerns the lack of to some important related work. Most critically, the authors should discuss “Hypothetical Minds: Scaffolding Theory of Mind for Multi-Agent Tasks with Large Language Models.” A clearer articulation of how EgoSocialArena advances or complements this prior work is needed. Further, work on normative decision making in LLMs requires first person thinking, and so I highly recommend the authors fully develop the related works. There is research without the term Theory of Mind in it that implicitly requires theory of mind (terms like cooperation, norms, competition).The shift from third-person (e.g., “Does John believe?”) to first-person (e.g., “Do you believe?”) may not represent a fundamentally new cognitive process for LLMs. As LLMs lack self-awareness, this change might only be a syntactic transformation rather than a meaningful distinction in reasoning. The authors need to justify why this framing shift contributes to deeper ToM understanding beyond improved prompt engineering.

Even though the design here avoids the babysitting problem, I worry that it introduces a new one. The agents here are performing Theory of Mind on rule based and RL agents that by definition do not have minds. This creates an epistemological tension: ToM presupposes the existence of mental states, which are absent in the models being simulated. The framework risks becoming a performance of ToM without grounding in real cognitive processes, raising philosophical concerns similar to those found in Searle’s Chinese Room Argument or Dennett’s intentional stance.

The use of “socialization” to describe how a model’s mental state changes may lead to confusion, as the term traditionally refers to learning normative behavior within a social context. This distinction could be critical for readers, especially those working in psychology or sociology, where the term has a different connotation. A more precise term, like adaptive inference or situational adaptation, would clarify the intended meaning.

My major concern with the work is the lack of a clear problem definition, which makes its position in the literature vague. Given the huge number of benchmarks available, it is essential to define which aspects of LLM capabilities are measured in specific scenarios. Although the authors generally describe the topics as ToM and socialization, the connections between these two concepts and the rationale for the chosen evaluation scenarios are not discussed in depth. Specifically, I do not fully agree with the definition of socialization as first-order ToM from a first-person perspective (i.e., introspection). The ability to infer one's own and others' mental states in social interactions is just one aspect of socialization; other important aspects include collaboration, communication, deception, normative reasoning, and more.

The lack of a clear problem definition relates to my second concern: the relationship between the three sections of EgoSocialArena is unclear. The proposed benchmark appears to be a combination of three different measurements with diverse targets and criteria. In first-person ToM questions, LLMs read first-person descriptions of social interaction scenarios and answer tests similar to the Sally-Ann task. In intriguing and distinctive social situations, LLMs read stories with novel social norms and answer social reasoning questions. In interactive environments, LLMs play games with other agents through embodied interaction to maximize task performance. These scenarios differ significantly in terms of complexity, information asymmetry, and measurement criteria. While this diversity makes the benchmark more comprehensive, it also increases ambiguity, which might hinder researchers from effectively utilizing it.

Additionally, details about the proposed evaluation scenarios and human baselines are missing from the paper and the appendix. For example, how many players are involved in the interactive environments? What is the average length of the stories used in EgoSocialArena? How were these scenarios designed, and what sources were used? Are these scenarios available online? What are the individual differences among human participants' performance?

Overall, focusing on a first-person perspective and embodied interactive scenarios is a promising direction for developing ToM benchmarks with greater validity. However, the process of assembling measurement scenarios does not align with the overarching goal of this work. The community would benefit more from a comprehensive analysis of existing ToM benchmarks regarding their criteria, scenarios, and the reasons why LLMs yield different results, rather than introducing new benchmarks with ambiguous definitions.

• The definition of socialization still needs some clarification. From the example in Fig 2, socialization seems like transforming the level-1 ToM reasoning into 1st person. Then how is this different from the typical policy or reaction of an agent for a given state? Similarly, the example for counterfactual in Fig. 3 also just infers the result given a state and action pair. These examples are not necessarily social. The LLMs need commonsense or rule-matching capabilities. In these cases, how can we define socialization?
• The rule-based agent for number guessing doesn’t seem to match the idea of different reasoning levels in ToM. For example, in level 2, the rule-based agent does not consider anything about the other agent. In this case, it might be easy for a model to track the level-2 rule-based agent.
• While the paper mentioned the proposed benchmark can be used for embodied AI research, none of the evaluations involve embodied agents.

I have various areas of concerns that I will separate into major and minor:

Major Weaknesses:

1. The authors, while clear in writing, are not nearly specific enough when making contribution claims in the paper, which means it's somewhat difficult to interpret the significance or novelty of the results. For example, in the introduction, the authors write "The powerful capabilities of the o1-preview model are truly surprising" as a core result from their investigation. What does this mean? As a reviewer or reader of the paper, there is no information of value to be derived from a statement like that, especially not when in a summarization of the paper's contributions. Please be more clear, e.g. "The powerful capabilities of the o1-preview model are truly surprising because they underperform xyz model by x% due to this reason".
2. Section 3 describing the conversation of existing benchmarks into EgosocialArena is quite unclear. For example, in 3.1 he authors immediately launch into a description of the conversion process without describing what the original benchmarks are in great detail. Also, in 3.2, the authors claim they chose 3 interesting scenarios -- why are they interesting?

Minor Weaknesses:

1. The authors convey very little analysis of their results. In Section 5, when overviewing the surprising or unsurprising results (e.g. The powerful capabilities of the o1-preview model are truly surprising), please give more intuition behind WHY this result holds, or in other words, please give some analysis for why the benchmark produced these types of differences in models.

Most importantly, It is not clear for all the evaluatons what the LLM is expected to do, and how this is scored. For example, while there is much description fo teh number guessing game, the actual ToM TASK is not described - what information is the LLM supposed to be providing and based on what? Similar for blackjack - "For the Blackjack scenario, we conducted a manual evaluation." - of what? There needs to be much much more detail on teh tasks themselves and the methods generally. Without knowing what these different tasks require, how do we interpret that the same models pass some with flying colours while catastrophically failing others? There is also no review of these inconsistencies and what they might mean in the dicussion - beyond who is better at what task. The critiques as to experimental details are somewhat alleviated by the contents of teh appendix (although these are never referenced in the text) however only partially. First, only a subset of tasks are decribed, second the ones described still do not provide enought information: For e.g. giving the rules of the game blackjack does not explain why this is considered to be a ToM task. Much of this information should be in the main text, and the "related work" section must include literature on the use of these tasks as ToM tests (including why they are considered to assess ToM). Other than the tasks that were adapted from existing ToM third-person vigniettes, it is not clear how the specific task instances were generated (for the static tasks) and what (if any) measures were taken to ensure robustness (e.g. systematic variation)

The fact that one version of each model was run once means that it is impossible to tell whether any chance / difference in performance is robust. Even when temperature is set at zero there are minor variations.