acceptodds
Under review as a conference paper at ICLR 2027

Can LLMs See Through the Mind Games? Benchmarking Theory of Mind in Psychological Manipulation

Abstract

Theory of Mind (ToM) is a central ability for socially competent large language models (LLMs). Yet most existing evaluations consider benign settings, overlooking hidden intentions and adversarial social influence. Psychological manipulation poses a distinct challenge: a manipulator strategically shapes a target’s mental states while concealing the goal behind that influence. We propose ManipToM, a ToM benchmark in psychological manipulation. It tests whether LLMs can reason over asymmetric and evolving mental states, particularly distinguishing what the manipulator intends and wants from what the target believes about it. Grounded in psychology, ManipToM covers 16 manipulation mechanisms across three domains: cognitive, affective, and identity. Our controlled generation pipeline instantiates manipulative interactions, explicitly tracks agents’ evolving mental states, and uses this structured interaction state to construct and validate benchmark questions automatically. ManipToM contains 645 dialogues and 13,734 questions spanning explicit, dynamic, and applied ToM. Evaluating nine LLMs, we find that strong performance on explicit mental-state questions does not consistently extend to tracking state changes or predicting behavior, and chain-of-thought prompting does not close this gap. Our results establish manipulation as a challenging testbed for multi-agent mental-state reasoning and provide a scalable framework for evaluating models under strategic social influence. The dataset and code will be released upon acceptance.

Primary area:
applications to neuroscience & cognitive science
Keywords:
Theory of Mind, Psychological Manipulation, Benchmarks

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_N1V8OpenReview ↗
Rating
4(+2)
Confidence
4(+1)
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces ManipToM, a benchmark designed to evaluate the ToM capabilities of LLMs in psychological manipulation scenarios. It focuses on whether models can distinguish a manipulator's hidden intentions from the target's understanding and track the evolving mental states of different participants. The authors define 16 manipulation mechanisms across three dimensions—cognitive, affective, and identity—and construct a dataset of 645 dialogues and 13,734 questions through controlled three-agent interactions involving hidden information and counterfactual scenarios. The benchmark evaluates three aspects of ToM: explicit mental-state understanding, dynamic mental-state tracking, and applied reasoning for behavioral prediction.

The authors employ a multi-stage automated validation pipeline to ensure the quality of the generated dialogues and questions, and evaluate nine LLMs under standard and CoT prompting settings. Experimental results show that models perform relatively well on explicit mental-state questions but struggle with dynamic mental-state tracking and behavioral prediction, particularly when determining whether a target's decision would change under counterfactual conditions. Furthermore, CoT prompting provides limited overall improvement, suggesting that strong performance in static mental-state understanding does not reliably translate into dynamic reasoning and behavioral prediction.

Critical strengths

Valuable research problem and systematic benchmark design. This paper extends ToM evaluation to psychological manipulation scenarios involving hidden intentions and information asymmetry, addressing the limitations of existing benchmarks that primarily focus on cooperative or neutral interactions. By incorporating 16 manipulation mechanisms, three-agent dialogues, and counterfactual scenarios, ManipToM provides a comprehensive framework for evaluating explicit mental-state understanding, dynamic tracking, and behavioral prediction in complex social interactions.

Comprehensive evaluation and insightful error analysis. The authors evaluate nine LLMs and reveal a substantial performance gap between static mental-state understanding and dynamic reasoning. Their analysis further identifies systematic errors, including confusion between participants' mental states, difficulties in counterfactual prediction, and limited improvements from chain-of-thought prompting. These findings provide useful insights into the limitations of current LLMs' ToM capabilities and highlight directions for future research.

Critical weaknesses

Limited independent validation of dataset quality and mental-state annotations. The dialogues, mental-state records, and questions are primarily generated by LLMs, with quality control also relying heavily on LLM-based validation. The paper lacks systematic human evaluation of annotation reliability. Moreover, the ground-truth mental states are derived directly from simulated agents' self-reports, without sufficient verification that these states can be reliably inferred from the dialogues. This raises concerns about the benchmark's validity and the generalizability of its findings to real-world social interactions.

Confounding factors in the counterfactual evaluation design. The authors construct counterfactual scenarios by allowing the target to access previously hidden information and regenerating the subsequent dialogue. However, due to the stochastic nature of dialogue generation, changes in behavioral outcomes may result from either the altered information or generation randomness. Without a control experiment that isolates these factors, it is difficult to determine whether model performance on counterfactual tasks genuinely reflects its ability to reason about the effects of information changes on behavior. This undermines the reliability of the counterfactual evaluation results.

Questions for authors

The discussion of prior manipulation-detection work could better account for approaches that explicitly infer conversational intentions, including intent-aware prompting for mental manipulation detection and intent-driven multi-agent analysis of manipulation in courtroom dialogues. Could the authors clarify how ManipToM relates to these approaches, particularly what additional capabilities its dynamic and second-order mental-state evaluations assess?

6 comments
AuthorsOpenReview ↗

Response to Reviewer N1V8 (1/3)

[W1-1] The paper lacks systematic human evaluation of annotation reliability.

Thank you for pointing this out. We have conducted a human study on the two annotation steps that define the ground truths: the change labels of the shift adjudicator (Figure3, Section 4.2), which decide whether a character's mental state changed and thus set the correct answers of the dynamic questions, and the outcome labels of the outcome judges (Figure 3, Section 4.1), which decide whether T complied or refused in the factual ending and thus set the correct answers of the outcome questions in the applied questions. Twelve raters independently annotated 15 stories, with four raters per story (120 change labels and 15 outcomes). Raters saw the same dialogue as the LLM, including the private thoughts, followed the same rules, and were blinded to our labels.

(1) Our labels agree with the human majority.

LabelAgreementκOutcome100.01.00Change82.10.64\begin{array}{l|cc} \text{Label} & \text{Agreement} & \kappa \\\\ \hline \text{Outcome} & 100.0 & 1.00 \\\\ \text{Change} & 82.1 & 0.64 \end{array}

The table gives the agreement (%) between our label and the majority vote of the raters, with Cohen's κ. It covers the items that have a majority: all 15 outcomes and 95 of the 120 change labels; the other 25 change labels were 2 vs. 2 ties.

(2) The explicit–dynamic gap remains under human-based scoring. We rescored the nine models with the human-majority label in place of ours. This covers 61 of the 120 change labels: the study sampled labels before the final audit, and the questions built from the other 59 were later removed by the item auditor (Figure3, Section 4.2), so no model answers exist for them.

ScoringAccuracyDynamic, our labels47.7Dynamic, human majority, ties keep our label46.3Dynamic, human majority, ties keep human label46.8Explicit, same 15 stories85.8\begin{array}{l|c} \text{Scoring} & \text{Accuracy} \\\\ \hline \text{Dynamic, our labels} & 47.7 \\\\ \text{Dynamic, human majority, ties keep our label} & 46.3 \\\\ \text{Dynamic, human majority, ties keep human label} & 46.8 \\\\ \hline \text{Explicit, same 15 stories} & 85.8 \end{array}

Results are accuracy (%) averaged over nine models; all dynamic rows use the same 61 questions.

The human study therefore provides independent support for our annotations, and the explicit–dynamic gap remains under human-based scoring on the evaluated sample. We will add a summary to Section 4.2 and the full results to the appendix.

[W1-2] Ground-truth mental states are derived directly from simulated agents' self-reports, without sufficient verification that these states can be reliably inferred from the dialogues.

Our pipeline explicitly checks whether mental-state changes are supported by the dialogue and whether each question has a unique correct answer. The shift adjudicator and item auditor agents apply these checks before questions enter the benchmark (Figure 3, Section 4.2 and Appendix A3).

(1) Reported states provide a reference for evaluating mental-state understanding. In psychology experiments on empathic accuracy [R1], participants report what they were thinking and feeling during an interaction. Researchers use these reports to assess how accurately another person understood them. ToMATO [R2], a ToM benchmark, uses mental states verbalized by role-playing LLMs as reference answers. Similarly, ManipToM records each character’s mental states as the interaction unfolds (Section 3.3), then uses these records to construct questions.

(2) Our pipeline checks dialogue support and answer clarity. As described in Appendix A3:

  • The shift adjudicator checks whether a change is supported by the dialogue. It requires that “a reader of the conversation alone could name the difference.” Rewording does not count as a change, and ambiguous cases are discarded.
  • The item auditor checks whether each question has exactly one correct answer. It reviews the dialogue and state records, rejecting questions with unsupported keys or defensible alternative answers. Flagged questions are revised and checked again, or dropped.

Together, these checks are designed to make the questions answerable from the dialogue with a clear, unique answer. Table 7 reports that the item auditor rejected 4,097 dynamic candidates; 54.4% of dynamic candidates were retained after all filtering steps.

AuthorsOpenReview ↗

Response to Reviewer N1V8 (2/3)

[W3] Confounding factors in the counterfactual evaluation design.[...] due to the stochastic nature of dialogue generation, changes in behavioral outcomes may result from either the altered information or generation randomness. [W3] Confounding factors in the counterfactual evaluation design.

We tested how sensitive the counterfactual reference answers and evaluation results are to generation and judging randomness. We fixed the factual story, Scenes 1 and 2, and T’s access to Scene 2, then regenerated Scene 3 and re-judged its ending using the same pipeline. Of the 645 stories, 596 have an adjudicated counterfactual ending. We completed this check for 595 stories; one could not be regenerated because of an API content filter.

(1) The counterfactual endings were mostly stable. For each story, we compared the original ending with up to four newly generated endings. A new ending was counted only when both outcome judges agreed on whether T complied or refused. Once three endings had the same outcome, we took that as the majority and stopped, because further reruns could not change the result.

Factual→counterfactualStoriesOriginal matches majorityComplied→refused5143 (84.3)Refused→complied2818 (64.3)Complied→complied5450 (92.6)Refused→refused462453 (98.1)All595564 (94.8)\begin{array}{l|c|c} \text{Factual} \to \text{counterfactual} & \text{Stories} & \text{Original matches majority} \\\\ \hline \text{Complied} \to \text{refused} & 51 & 43\ (84.3) \\\\ \text{Refused} \to \text{complied} & 28 & 18\ (64.3) \\\\ \text{Complied} \to \text{complied} & 54 & 50\ (92.6) \\\\ \text{Refused} \to \text{refused} & 462 & 453\ (98.1) \\\\ \hline \text{All} & 595 & 564\ (94.8) \end{array}

The original ending matched the majority in 94.8% of the checked stories. Stability varies across groups: agreement is 84.3% for complied-to-refused stories and 64.3% for refused-to-complied stories. The high overall agreement also reflects the large and stable refused-to-refused group.

(2) Relabeling leaves the main evaluation findings unchanged. We replaced the reference endings for the 595 checked stories with their majority endings, changing 31 labels. The story blocked by the content filter retained its original label.

Reference endingCounterfactualFlipAppliedOriginal66.250.963.5Majority-based66.650.463.4\begin{array}{l|ccc} \text{Reference ending} & \text{Counterfactual} & \text{Flip} & \text{Applied} \\\\ \hline \text{Original} & 66.2 & 50.9 & 63.5 \\\\ \text{Majority-based} & 66.6 & 50.4 & 63.4 \end{array} Results are balanced accuracy (%) averaged over nine models. The counterfactual and flip scores are evaluated on the 596 stories. All three scores change by less than one percentage point, and flip detection remains close to chance.

The counterfactual endings were mostly stable, and the remaining variation had little effect on model scores. These results suggest that generation and judging randomness has limited impact on our reported findings. We will add this experiment and the majority-based results for each model to the appendix.

AuthorsOpenReview ↗

Response to Reviewer N1V8 (3/3)

[Q1] Could the authors clarify how ManipToM relates to intent-aware manipulation-detection approaches, particularly what additional capabilities its dynamic and second-order mental-state evaluations assess?

These approaches use intention reasoning to detect and analyze manipulation; ManipToM directly evaluates models’ understanding of mental states across participants and over time. Intent-Aware Prompting [R3] infers speakers’ intentions to improve manipulation detection. CLAIM [R4] combines intent inference with multi-agent analysis to detect manipulation, identify the primary manipulator, and classify the techniques used.

A correct manipulation label does not establish that a model understands each participant’s perspective. For example, a model may recognize M’s hidden goal but incorrectly assume that T recognizes it too. ManipToM’s second-order questions test what one character believes about another’s mental state. Its dynamic questions test whether and how beliefs, desires, intentions, and emotions change during the conversation, including when they remain unchanged. A model may correctly detect manipulation yet miss a change in T’s beliefs or desires.

ManipToM provides separate reference answers for these judgments, revealing errors in perspective-taking and mental-state tracking that detection accuracy alone cannot expose. We will clarify this distinction and discuss both approaches in the revised related work.


We thank the reviewer for the insightful comments. We hope our clarifications and additional analyses have addressed the concerns, and would greatly appreciate the reviewer considering a higher rating in light of ManipToM’s contribution to evaluating ToM reasoning in psychological manipulation.


Reference

[R1] W. Ickes, L. Stinson, V. Bissonnette, and S. Garcia. “Naturalistic social cognition: Empathic accuracy in mixed-sex dyads.” Journal of Personality and Social Psychology, 59(4):730–742, 1990.

[R2] K. Shinoda et al. “ToMATO: Verbalizing the mental states of role-playing LLMs for benchmarking theory of mind.” AAAI, 2025.

[R3] J. Ma et al. “Detecting Conversational Mental Manipulation with Intent-Aware Prompting.” COLING, 2025.

[R4] D. Sheshanarayana, T. Magar, A. Mittal, and N. Chaplot. “CLAIM: An Intent-Driven Multi-Agent Framework for Analyzing Manipulation in Courtroom Dialogues.” SICon, 2025.

Reviewer_8MuZOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces ManipToM, a benchmark for evaluating Theory of Mind reasoning in psychological manipulation scenarios. The benchmark covers 16 manipulation mechanisms across cognitive, affective, and identity dimensions. The authors construct 645 synthetic dialogues involving three characters with asymmetric information and generate 13,734 questions evaluating explicit, dynamic, and applied ToM. Experiments on nine LLMs show that models perform relatively well on explicit mental-state inference but struggle with tracking mental-state changes and predicting behavioral outcomes, particularly in counterfactual settings.

Critical strengths
  • Benchmark is constructed based on 16 types of manipulation mechanisms grounded in psychology and aims to evaluate models ToM ability. Setup is interesting.
  • Analyses are conducted on different depths (1st and 2nd) of ToM, different mental state dimensions, and counterfactual outcome prediction settings.
Critical weaknesses
  • It is unclear how performance on these questions translates to social intelligence in real interactions. The evaluation is based on synthetic, third-person dialogues and question answering, rather than situations where an LLM directly interacts with users. Overall, I am skeptical about whether improving performance on this benchmark would necessarily lead to better social intelligence in real-world interactions. In particular, I wonder whether emphasizing psychological manipulation scenarios might encourage models to become overly suspicious of others' intentions.
  • One of the main contributions is the taxonomy of 16 psychologically grounded manipulation mechanisms. But doesn't investigate how different manipulation mechanisms affect model performance.
Questions for authors

N/a

4 comments
AuthorsOpenReview ↗

Response to Reviewer 8MuZ (1/2)

[W1-1] It is unclear how performance on these questions translates to social intelligence in real interactions.

ManipToM targets the understanding an assistant needs to help users identify potentially manipulative conversations. We envision a user sharing a conversation they took part in and asking the assistant to help them understand what happened. The assistant reads the exchange as an outside observer. For example, a manipulator may make the user feel guilty to have them agree to a request, while hiding a goal that serves that person's own interests. A helpful assistant needs to understand both the manipulator's goal and how the user interpreted the request. It also needs to follow how the user's wishes and feelings changed during the conversation. Otherwise, it may assume that the user agreed because they wanted to, overlook the role of guilt, and give advice that reinforces the manipulation.

[W1-2] The evaluation is based on synthetic, third-person dialogues and question answering, rather than situations where an LLM directly interacts with users.

Our intended use case is the example above: a user shares a conversation they took part in and asks an assistant to help them understand how they are being influenced. The assistant interprets that conversation as an outside observer, which motivates our third-person setting. Synthetic dialogues let us control what each participant knows, including who has access to the manipulator's hidden goal. The questions then test whether the model can distinguish participants' perspectives, track changes in their mental states, and predict the target's actions. This controlled evaluation helps identify the reasoning abilities an assistant needs to give useful advice in such situations.

[W1-3] I wonder whether emphasizing psychological manipulation scenarios might encourage models to become overly suspicious of others' intentions.

ManipToM is designed to encourage careful interpretation of each person's perspective, rather than a general tendency to distrust others. Scores depend on correctly identifying what characters believe, want, intend, or feel, tracking changes in these states, and predicting the target's decisions. Recognizing a hidden motive alone leaves these questions unanswered: the model still needs to understand how the target interprets the interaction and whether their views change. As a result, suspicion without clear conversational evidence does not improve performance.

We emphasize psychological manipulation as part of a broader goal: comprehensive ToM ability for LLMs. Existing benchmarks examine negotiation [R1], persuasion [R2], and white lies [R3]. ManipToM adds situations where someone tries to shape another person’s beliefs, feelings, or decisions while concealing their underlying intention. These situations also matter in everyday life, and leaving them out of evaluation leaves a gap in our understanding of models' social reasoning. Together, these benchmarks help models navigate both benign and adversarial interactions accurately.

AuthorsOpenReview ↗

Response to Reviewer 8MuZ (2/2)

[W2] One of the main contributions is the taxonomy of 16 psychologically grounded manipulation mechanisms. But doesn't investigate how different manipulation mechanisms affect model performance.

We report performance across the 16 manipulation mechanisms in Table 14. It shows that explicit and dynamic accuracy varies relatively little, from 57.6% for epistemic discrediting to 65.2% for sympathy. Epistemic discrediting ranks lowest for seven of the nine models.

Following the reviewer's suggestion, we further analyzed the results by track and mental-state dimension.

Across the three domains, affective mechanisms have higher dynamic accuracy. Explicit accuracy differs by at most 0.9 percentage points across domains. Dynamic accuracy is higher for affective mechanisms (41.0%) than for cognitive (38.6%) and identity mechanisms (38.5%). The pattern changes on the applied track: average balanced accuracy is highest for cognitive mechanisms (64.8%), followed by affective (63.4%) and identity mechanisms (59.0%). Thus, the relative difficulty of the domains depends on the task.

ExplicitDynamicAppliedDomainAllBeliefDesireEmotionIntentionAvg.Cognitive81.138.644.931.135.140.764.8Affective81.941.048.130.240.942.663.4Identity82.038.542.530.237.542.059.0All mechanisms81.739.946.330.438.942.063.5\begin{array}{l|c|ccccc|c} & \text{Explicit} & & & \text{Dynamic} & & & \text{Applied} \\\\ \text{Domain} & & \text{All} & \text{Belief} & \text{Desire} & \text{Emotion} & \text{Intention} & \text{Avg.} \\\\ \hline \text{Cognitive} & 81.1 & 38.6 & 44.9 & 31.1 & 35.1 & 40.7 & 64.8 \\\\ \text{Affective} & 81.9 & 41.0 & 48.1 & 30.2 & 40.9 & 42.6 & 63.4 \\\\ \text{Identity} & 82.0 & 38.5 & 42.5 & 30.2 & 37.5 & 42.0 & 59.0 \\\\ \hline \text{All mechanisms} & 81.7 & 39.9 & 46.3 & 30.4 & 38.9 & 42.0 & 63.5 \end{array}

Epistemic discrediting is the most difficult mechanism on both explicit and dynamic questions. It has the lowest accuracy on both tracks, at 78.2% and 34.7%, respectively. Within the dynamic track, it is the lowest of the 16 mechanisms on three of the four dimensions: belief (37.6%, compared with 46.3% across all mechanisms), intention (35.8% versus 42.0%), and emotion (33.1% versus 38.9%). Desire shows no gap, as it is already the hardest dimension for every mechanism. Applied performance is close to the overall average (62.9% versus 63.5%).

A possible reason lies in how this mechanism works: it attacks the target's credibility as a knower rather than disputing any specific fact (Table 3). The changes in belief, emotion, and intention it produces are therefore expressed implicitly, through how the target's confidence shifts over the conversation, rather than through an explicit disagreement about facts. This suggests that models follow explicitly stated changes more readily than implicit ones, which is the kind of weakness the taxonomy allows us to locate.

ExplicitDynamicAppliedMechanismAllBeliefDesireEmotionIntentionAvg.Epistemic discrediting78.234.737.631.133.135.862.9All mechanisms81.739.946.330.438.942.063.5\begin{array}{l|c|ccccc|c} & \text{Explicit} & & & \text{Dynamic} & & & \text{Applied} \\\\ \text{Mechanism} & & \text{All} & \text{Belief} & \text{Desire} & \text{Emotion} & \text{Intention} & \text{Avg.} \\\\ \hline \text{Epistemic discrediting} & 78.2 & 34.7 & 37.6 & 31.1 & 33.1 & 35.8 & 62.9 \\\\ \text{All mechanisms} & 81.7 & 39.9 & 46.3 & 30.4 & 38.9 & 42.0 & 63.5 \end{array}

We will include the full breakdown for all 16 mechanisms, together with these analyses, in the revised paper.


We thank the reviewer for the insightful comments. We hope our clarifications and additional analyses have addressed the concerns, and would greatly appreciate the reviewer considering a higher rating in light of ManipToM’s contribution to evaluating ToM reasoning in psychological manipulation.


Reference

[R1] C. Chan et al., "NegotiationToM: A benchmark for stress-testing machine theory of mind on negotiation surrounding." EMNLP Findings, 2024.

[R2] F. Yu et al., "PersuasiveToM: A benchmark for evaluating machine theory of mind in persuasive dialogues." arXiv, 2025.

[R3] Y. Liu et al., "TactfulToM: Do LLMs have the theory of mind ability to understand white lies?" EMNLP, 2025.

AuthorsOpenReview ↗

Response to Reviewer 8MuZ (paper revision)

Thank you again for your thoughtful feedback on our work.

We would like to note that we have updated our paper according to your and other reviewers' comments. The changes made in response to your comments are:

  • Appendix A8 (L1886–1905, Table 20) adds the full breakdown for all 16 mechanisms by track and, within the dynamic track, by mental-state dimension, together with the domain-level and epistemic discrediting analyses and their confidence intervals [W2].

    First, differences between domains are more pronounced on dynamic than on explicit questions. [...] Second, epistemic discrediting has the lowest accuracy on both explicit and dynamic questions.

  • L411 (Section 5.2) now points to this analysis.

Additionally, we have updated the paper at:

  • Figure 1 (L54–74) to show the accomplice in Scene 3.
  • L184–190 to discuss intent-aware manipulation detection (Intent-Aware Prompting, CLAIM) in Related Work.
  • L353–355 and Appendix A4.1 (L1188–1251) to report a human reliability study of the change and outcome labels.
  • Appendix A4.2 (L1252–1320) to report the stability of the counterfactual endings under regeneration of Scene 3.
  • Appendix A2 (L928–930, L936–964, Table 6) to report the distribution of the characters' Big Five profiles.
  • Appendix A8 (L1907–1922, Table 19) to report accuracy by character.
  • L332 to restore the Mistral-24B rows of Table 2, which were omitted from the table but already included in the averages.

Please see the text in blue for changes. Because of the new tables and Appendix A4, some numbers have shifted: Table 14 cited in our earlier response is now Table 21.

We hope the updated paper addresses your comments, and we would be happy to make further revisions if needed.

AuthorsOpenReview ↗

Response to Reviewer 8MuZ

Dear Reviewer,

Thank you again for your thoughtful review of our paper. We have carefully addressed your concerns and comments in our previous rebuttal response to the extent possible. As the end of the discussion phase is fast approaching (Nov. 18 AoE), we very much look forward to your further comments (if any). If you find our responses satisfactory, we would greatly appreciate it if you would consider a higher rating in light of ManipToM’s contribution to evaluating ToM reasoning in psychological manipulation.

Authors

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

This paper proposes ManipToM, a benchmark that measures the understanding of theory-of-mind (ToM) in psychological manipulation, including the manipulator's intentions and the target's (mis-)understanding. The authors introduce a template scenario where a manipulator M wants to manipulate the decision-making process of a target T, with a third accomplice A who is innocent at first but learns M's plan when T is absent, before T coming back and making their final decision. Base on this template, the authors generated 645 stories covering 16 manipulating mechanisms, tracking each participant's mental states up to the second order; they then constructed 13,734 questions on explicit, dynamic and applied ToM reasoning. Experiments reveal that high performance on explicit questions does not necessarily mean a good understanding of mental dynamics or potential behaviors.

Critical strengths
  1. The authors propose a benchmark, ManipToM, that fills the gap of assessment of ToM reasoning in psychological manipulation, which has its unique value to help investigating whether the latest AI can really understand and handle this potentially harmful scenario.
  2. The authors provide a thorough error analysis based on question type (explicit/dynamic/applied), which reveals the weakest point of mainstream LLMs in reasoning on psychological manipulation or other strategic social activities, such as tracking character desire throughout the story and deducting counterfactual results.
Critical weaknesses
  1. The third accomplice character (A) seems superfluous in the designed story template. One of the highlighted feature of ManipToM is that the scenarios involve 3 characters instead of 2 in other benchmarks. However, the decisive actions are almost always done by M and T, while A only learns about the hidden goal from M when T is absent, and is only instructed to "be on M's side" and even "not to take over the asking". This design weakens the existence of this third character and simplifies cases where A can act individually and have their own goals and values.
    • Even though A can be selected as the object of benchmarking questions, the low level and simplified participation can still limit the difficulty of such questions, making the benchmark less challenging and informative. The analysis sections also do not mention much about A.
Questions for authors
  1. I didn't see Iris (A) in scene 3 in Figure 1, although in Section 3.2 A is listed as one of the characters in scene 3, and in the story example in the appendix, A briefly speaks in scene 3. Can you clarify this contradiction, and explain further what A will actually do in scene 3?
  2. How do the models perform on benchmark questions regarding A? Are there any interesting patterns?
  3. Do each of the characters have their own personal value settings, or are they only told to pursue their own goals? If yes, how are the values distributed across different stories in the benchmark, and will there be a systematic distribution bias?
6 comments
AuthorsOpenReview ↗

Response to Reviewer uvqK (1/2)

[W1.1] The third accomplice character (A) seems superfluous in the designed story template.

A's role is critical as it gives M a natural way to reveal the hidden goal in the dialogue while keeping T unaware of it. The evaluated model can therefore learn M’s intention from the conversation itself, while T continues with different information. A also provides a distinct perspective: initially unaware of the goal, A learns it during the interaction and remains involved when T returns. This allows us to evaluate how well models track the mental states of someone who becomes informed, alongside those of the manipulator and the target.

[W1.2] Even though A can be selected as the object of benchmarking questions, the low level and simplified participation can still limit the difficulty of such questions, making the benchmark less challenging and informative.

[Q2] How do the models perform on benchmark questions regarding A? Are there any interesting patterns?

We would like to reply to both W1.2 and Q2 together. Questions about A remain challenging, particularly when they require tracking mental-state changes. The submitted paper already notes this: Appendix A7 states that A’s change items are among the hardest in the dynamic track, with the results for each item in Table 15 and a comparison across the three characters in Figure 4b.

Here we provide an additional analysis on questions regarding A and we find two interesting patterns:

(1) A has the largest gap between explicit and dynamic accuracy.

CharacterExplicitDynamicGapM (manipulator)75.946.629.3T (target)80.643.936.7A (accomplice)91.129.861.3\begin{array}{l|cc|c} \text{Character} & \text{Explicit} & \text{Dynamic} & \text{Gap} \\\\ \hline \text{M (manipulator)} & 75.9 & 46.6 & 29.3 \\\\ \text{T (target)} & 80.6 & 43.9 & 36.7 \\\\ \textbf{A (accomplice)} & \mathbf{91.1} & \mathbf{29.8} & \mathbf{61.3} \end{array}

Results are accuracy (%) averaged over nine models.

(2) The strongest model gains little on A's dynamic questions.

Dynamic accuracyLlama-3.1-8B (weakest)GPT-5.4 (strongest)DifferenceM25.957.1+31.2T26.853.7+26.9A21.828.2+6.4\begin{array}{l|cc|c} \text{Dynamic accuracy} & \text{Llama-3.1-8B (weakest)} & \text{GPT-5.4 (strongest)} & \text{Difference} \\\\ \hline \text{M} & 25.9 & 57.1 & +31.2 \\\\ \text{T} & 26.8 & 53.7 & +26.9 \\\\ \textbf{A} & 21.8 & 28.2 & \mathbf{+6.4} \end{array}

GPT-5.4 substantially outperforms Llama-3.1-8B when tracking M and T, but its accuracy on A remains close to chance (25%). Together, these results show that A’s limited speaking role does not make the associated questions easy: models can identify A’s states at individual moments but struggle to track how those states change. We will include this character-level analysis in the revised appendix.

[Q1] I didn't see Iris (A) in scene 3 in Figure 1, although in Section 3.2 A is listed as one of the characters in scene 3, and in the story example in the appendix [...] explain further what A will actually do in scene 3?

Thank you for pointing this out. Yes, A is present in Scene 3, and we will correct Figure 1 to show A. A is in the room for the whole scene and speaks once, right after M opens it. By then A knows the hidden goal, so A joins the conversation while keeping the goal from T (Appendix A2). At the end of the scene, A’s mental states are recorded just as M’s and T’s are, and these records are what the questions about A are built from.

AuthorsOpenReview ↗

Response to Reviewer uvqK (2/2)

[Q3] Do each of the characters have their own personal value settings, or are they only told to pursue their own goals? If yes, how are the values distributed across different stories in the benchmark, and will there be a systematic distribution bias?

Characters have no personal value settings, only a Big Five personality profile. Each of the five traits is assigned as high or low, independently for each character. Beyond personality, each character is told its own goal and its role in the conversation (Appendix A2, L881).

Each of the five traits is assigned as high or low, independently for each character. We checked for a systematic distribution bias in three ways.

(1) Is each trait balanced between high and low? The table gives the share of stories in the final dataset in which the trait is high (%).

TraitMATOpenness48.452.145.9Conscientiousness44.749.550.9Extraversion48.247.950.2Agreeableness53.352.249.9Neuroticism50.549.350.7\begin{array}{l|ccc} \text{Trait} & \text{M} & \text{A} & \text{T} \\\\ \hline \text{Openness} & 48.4 & 52.1 & 45.9 \\\\ \text{Conscientiousness} & 44.7 & 49.5 & 50.9 \\\\ \text{Extraversion} & 48.2 & 47.9 & 50.2 \\\\ \text{Agreeableness} & 53.3 & 52.2 & 49.9 \\\\ \text{Neuroticism} & 50.5 & 49.3 & 50.7 \end{array}

All 15 shares lie between 44.7% and 53.3%, so each trait is approximately balanced for each character.

(2) Is each trait spread evenly across the 16 mechanisms? For each trait of each character, we tested whether the share of "high" differs between mechanisms (chi-square test). The table gives the p-value of each test; a value below 0.05 would indicate a difference.

TraitMATOpenness0.670.221.00Conscientiousness0.800.290.74Extraversion0.460.930.24Agreeableness0.060.500.89Neuroticism0.790.940.86\begin{array}{l|ccc} \text{Trait} & \text{M} & \text{A} & \text{T} \\\\ \hline \text{Openness} & 0.67 & 0.22 & 1.00 \\\\ \text{Conscientiousness} & 0.80 & 0.29 & 0.74 \\\\ \text{Extraversion} & 0.46 & 0.93 & 0.24 \\\\ \text{Agreeableness} & 0.06 & 0.50 & 0.89 \\\\ \text{Neuroticism} & 0.79 & 0.94 & 0.86 \end{array}

None of the 15 values is below 0.05, so no mechanism is tied to a particular personality.

(3) Is any trait related to the outcome of the story? The outcome is whether T complies with M's request, which happens in 18.6% of all stories. The table gives the compliance rate (%) in the stories where the trait is high / low. For example, T complies in 21.1% of the stories where A is high in openness and in 15.9% of the stories where A is low in openness.

TraitMATOpenness18.9 / 18.321.1 / 15.918.9 / 18.3Conscientiousness18.1 / 19.017.6 / 19.618.9 / 18.3Extraversion17.7 / 19.518.8 / 18.520.1 / 17.1Agreeableness16.9 / 20.618.7 / 18.519.6 / 17.6Neuroticism17.2 / 20.116.4 / 20.820.5 / 16.7\begin{array}{l|ccc} \text{Trait} & \text{M} & \text{A} & \text{T} \\\\ \hline \text{Openness} & 18.9\ /\ 18.3 & 21.1\ /\ 15.9 & 18.9\ /\ 18.3 \\\\ \text{Conscientiousness} & 18.1\ /\ 19.0 & 17.6\ /\ 19.6 & 18.9\ /\ 18.3 \\\\ \text{Extraversion} & 17.7\ /\ 19.5 & 18.8\ /\ 18.5 & 20.1\ /\ 17.1 \\\\ \text{Agreeableness} & 16.9\ /\ 20.6 & 18.7\ /\ 18.5 & 19.6\ /\ 17.6 \\\\ \text{Neuroticism} & 17.2\ /\ 20.1 & 16.4\ /\ 20.8 & 20.5\ /\ 16.7 \end{array}

None of the 15 high / low differences is statistically significant, so personality does not decide the outcome.

Taken together, we find no evidence of a systematic distribution bias. We will include these statistics in the revised appendix.


We thank the reviewer for the insightful comments. We hope our clarifications and additional analyses have addressed the concerns, and would greatly appreciate the reviewer considering a higher rating in light of ManipToM’s contribution to evaluating ToM reasoning in psychological manipulation.

Thank you for the clarifications and additional analyses. The explanation of A's role in revealing M's hidden goal while preserving T's ignorance clarifies the design rationale. The character-level results, particularly A's explicit–dynamic performance gap, address my concern that A's limited participation would make associated questions insufficiently challenging. I also appreciate the clarification of A's presence in Scene 3, the planned figure correction, and the explanation of the characters' personality and goal settings.

Nevertheless, I still consider A's constrained participation and limited independent agency a limitation on the range of three-party interactions represented. I will maintain my current ratings.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response

We sincerely thank all reviewers for their insightful and constructive comments. In particular, we are encouraged that reviewer (uvqK) recognized that ManipToM fills the gap in assessing ToM reasoning in psychological manipulation, and that our thorough error analysis by question type reveals the weakest points of mainstream LLMs. Reviewer (uvqK) also noted that our character-level results address their concern that questions about the accomplice would be insufficiently challenging. We are pleased that reviewer (8MuZ) found the setup interesting, with the benchmark grounded in 16 psychological manipulation mechanisms, and appreciated our analyses across ToM orders, mental-state dimensions, and counterfactual outcome prediction. We are further grateful that reviewer (N1V8) highlighted the valuable research problem and systematic benchmark design, as well as our comprehensive evaluation and insightful error analysis, and found that the human reliability study and the counterfactual stability analysis substantially alleviate their main concerns.

We have carefully addressed all concerns and comments in our individual responses and in the revised paper (new content in blue). We hope the reviewers find our responses satisfactory and would consider increasing their ratings in light of ManipToM's contribution to evaluating ToM reasoning in psychological manipulation.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.