Pith. sign in

REVIEW 3 major objections 5 minor 29 references

Evaluating Theory of Mind in Reasoning Models: Robustness over Reasoning

T0 review · 3 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

Pith's one-line read Reasoning-enabled language models are more robust on Theory of Mind tasks, and the paper argues this robustness, not a new mental-state ability, explains recent gains.

desk verdict Useful, honest ToM evaluation with a plausible robustness story, but the central comparison is an uncontrolled vendor config and the abstract overstates the evidence. read the letter →

arxiv 2608.04646 v1 pith:F2CXBDEQ submitted 2026-08-05 cs.CL

classification cs.CL
keywords TheoryofMindreasoningmodelsreinforcementlearningwithverifiablerewardspromptrobustnessfalse-belieftestsLLMevaluationchain-of-thoughtbenchmarkperturbation
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper asks whether the strong recent performance of reasoning-oriented large language models on Theory of Mind (ToM) tests marks a new social-cognitive ability or a narrower improvement. Its answer is that the gains are best read as robustness: models trained to spend inference time on a chain of thought reach a correct answer that the underlying model could already, in principle, reach, and they do so more reliably when the same task is rephrased or perturbed. The evidence comes from adapted psychological tests (first- and second-order false belief, Strange Stories, Imposing Memories), prompt variants designed to disturb reasoning, a thinking-on versus thinking-off comparison, and reanalysis of third-party benchmarks. If the paper is right, progress on ToM benchmarks should be tracked by stability under variation, not just by average accuracy.

What carries the argument

The machinery is paired variation and consistency scoring. Task-preserving prompt modifications, inspired by earlier demonstrations that GPT-3 failed trivial rephrasings, change surface form while leaving the false-belief content intact; comparing Claude with thinking on and off on these variants isolates the effect of the reasoning process as far as the interface allows; and FANToM's all-questions-correct scoring turns answer instability into a hard penalty, making robustness differences visible in benchmark numbers. The paper also scores answers and their reasoning together on a 0-1-2 scale, so a correct inference path counts as evidence of stability rather than only a correct label.

What would settle it

Re-run the modification battery on Claude with thinking on and off across a much larger set of task-preserving paraphrases and multiple seeds; the robustness claim predicts lower per-paraphrase accuracy variance and higher minimum accuracy for thinking-on, so matching or better performance from the thinking-off configuration would contradict it.

Watch

Extended reading notes

Core claim

Reasoning-oriented LLMs consistently show increased robustness to prompt variations and task perturbations on ToM material. On the authors' tests, thinking-enabled Claude outperformed its thinking-off counterpart and was markedly better on the modified prompts designed to derail reasoning; GPT-5 made only a single partial mistake across the full battery; and third-party benchmark results show reasoning models above their non-reasoning peers, with the largest gap on FANToM, a benchmark that only credits a model when it answers every question of a given type correctly. The authors read this pattern as support for a robustness-based account: reasoning training stabilizes the selection of an already-available inference path rather than adding a fundamentally new capacity to represent mental states. They explicitly restrict the claim to behavior, noting that the comparisons are not perfectly controlled and do not causally isolate reinforcement-learning training.

Load-bearing premise

The paper's interpretation assumes that the differences between thinking-on and thinking-off Claude, and between reasoning and non-reasoning models more broadly, come from the reasoning process itself rather than from API filtering, model version, temperature, or other uncontrolled interface factors; the authors concede the comparisons are behavioral, not perfectly controlled.

Editorial extensions

If this is right

  • ToM evaluation should report accuracy across task-preserving prompt variants, since average benchmark scores alone cannot separate robustness from newly added capability.
  • Earlier prompt-sensitivity failures, such as GPT-3's collapse under trivial alterations, can be reinterpreted as failures to reliably reach an inference path the model already had, rather than as proof that ToM competence was absent.
  • Benchmarks that penalize inconsistency, like all-questions-correct scoring, are more sensitive detectors of the reasoning models' advantage than benchmarks that average over independent questions.
  • If the account holds, reasoning-oriented models should be expected to behave more consistently in social and agentic settings under rephrased inputs, while their ceiling remains bounded by the base model.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • A natural extension is to fold robustness into the operational definition of a model skill: if a capability disappears under task-preserving variation, it is arguably not a stable capability, making robustness a component of ability rather than a separate evaluation axis.
  • The account predicts an inverse relationship between reasoning effort and answer variance across paraphrases; this can be tested by sampling many paraphrases at increasing reasoning budgets and measuring variance in the final answer.
  • The same robustness lens could apply outside ToM: RLVR models' gains on math and code benchmarks may likewise reflect a narrowed, more consistent solution distribution rather than new knowledge, and paraphrased versions of those problems would expose this.
  • A decisive version of the comparison would pair each reasoning model with its exact base model on the same perturbation battery; the authors note this is feasible for models like R1 versus its base, and doing so would turn the behavioral pattern into a cleaner causal test.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 5 minor

Summary. The paper evaluates recent reasoning-oriented LLMs (GPT-5, Claude, R1, Grok-3-mini) on Theory of Mind (ToM) tasks from the battery of van Duijn et al., together with prompt perturbations inspired by Ullman, and compares reasoning-enabled configurations against non-reasoning configurations where available. It reports near-ceiling performance on Sally-Anne, Strange Stories, and Imposing Memories tasks, and mixed but generally higher scores for reasoning models on a set of modified simple ToM prompts, with a thinking-off Claude condition scoring lower on several items. The authors interpret the overall pattern as evidence that reasoning models exhibit increased robustness to prompt and task variation, supporting a robustness-based account of recent ToM gains rather than a new ToM-specific ability. The paper explicitly disclaims causal attribution to RLVR training and describes the comparisons as behavioral rather than fully controlled.

Significance. If the central interpretation holds, the paper makes a timely contribution to the debate on whether recent ToM improvements in LLMs reflect genuinely new social-cognitive capabilities or improved stability in reaching solutions that were already within reach. The manuscript is transparent: it provides open code, data, and qualitative reasoning traces; it incorporates external benchmark results from prior work; and it repeatedly hedges its causal claims. For these reasons, the paper is a useful contribution even though the evidence remains suggestive rather than conclusive. The significance would be strengthened if the quantitative support for the main claim matched the level of the abstract's assertion.

major comments (3)
  1. [Section 3.1 and Table 6] This is the load-bearing comparison for the paper's interpretation. The current text openly acknowledges the confound, but then proceeds to draw the robustness conclusion from the very same comparison. A reader cannot verify that the observed differences are due to reasoning rather than to uncontrolled configuration differences. The authors should either run a controlled study (e.g., same base model with and without inference-time scaling, if accessible) or reframe the conclusion as a hypothesis supported only by suggestive evidence, not as an observed regularity.
  2. [Section 4 and Tables 3-6] The absence of inferential statistics is especially problematic because the differences in Table 6 are small and the item counts appear to be low (e.g., 2C.1 and 2C.2 appear twice, suggesting a small item set). The paper's own phrasing in Section 5.2 acknowledges that 'no clear quantification of such an effect can be provided,' which is at odds with the abstract's unqualified claim of consistent increased robustness. This mismatch should be resolved.
  3. [Section 5.1 and Abstract] The paper's own limitations section is admirably candid, but the abstract and conclusion are not fully aligned with those caveats. Since the central theoretical contribution is the robustness-based interpretation, the mismatch between the hedged body and the assertive abstract is a load-bearing presentation issue that should be corrected.
minor comments (5)
  1. [Table 1] The column layout of Table 1 is difficult to read because the benchmark names and the three sub-columns (ParaphrasedToMi, FANToM, MMToM-QA) are not separated clearly; the numbers for the different benchmarks run together. Reformatting the table with explicit sub-headers or separate rows would improve readability.
  2. [Table 4] The column header 'whitelie' should be 'white lie' for consistency with the prose.
  3. [Section 4.4] The phrase 'significantly better' in the second paragraph of Section 4.4 should be replaced with a descriptive term such as 'numerically higher,' given that no statistical test is reported.
  4. [Section 4.5] The qualitative analysis of reasoning traces is not systematic: the criteria for identifying 'perspective-taking steps' or 'meta-knowledge' are not defined, and it is unclear how many responses were inspected and by whom. A brief coding scheme or inter-annotator agreement would improve the credibility of these qualitative claims.
  5. [References] The citation for the 'orders of reasoning' concept [12] is appropriate, but the definition in Section 2.2 is informal; a more precise reference to the recursive structure, or a page/section number in [12], would help readers.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: the robustness conclusion is an empirical attribution, not a derivation, and the paper explicitly limits its causal claims.

full rationale

The paper does not derive any target quantity from fitted parameters or from a self-citation chain. Its central claim, that reasoning models are more robust under prompt and task variation and that this supports a robustness-based interpretation, is an empirical attribution explicitly framed as a hypothesis rather than a causal identification. The benchmark numbers in Table 1 come from independent work [9]; the psychological tests and scoring rubric in [6] are used as measurement instruments, and although two of the present authors co-authored [6], the rubric is a fixed external protocol, not a result that presupposes the paper's conclusion. The only within-family comparison, Claude thinking on versus thinking off, is acknowledged in Section 5.2 as not equivalent to a base-model comparison, and Section 3.1 states that results 'should be interpreted as behavioral, not architectural.' These concessions are validity limitations, not circular reductions. The phrase 'we read this as evidence for a robustness-based account' is an interpretive label applied to the observed robustness, but it does not make a prediction that is forced by construction; the negative claim, that recent gains do not reflect a new ToM-specific ability, has independent content. No equations are derived from inputs, and no fitted parameter is renamed as a prediction. Accordingly, the circularity score is low.

Assumptions & free parameters 1 free parameters · 4 assumptions · 0 invented entities

The paper does not perform mathematical derivation, so the ledger captures measurement and comparability assumptions. No new entities are postulated. The scoring rubric is inherited from [6], and the prompt perturbations are author-designed; both are assumptions about how to map test responses to a robustness claim.

free parameters (1)
  • partial-credit score value = 1 on a 0-2 scale
    The rubric gives one point for reasoning 'close to the right answer' (Section 3.3); this subjective middle score affects the normalized averages in Tables 3-6 and the robustness comparison, with no sensitivity analysis.
assumptions (4)
  • domain assumption Performance on the selected psychological tests is a valid behavioral measure of Theory of Mind in LLMs.
    The paper relies on the test battery from [6] and the Quesque and Rossetti criteria (Section 2.2) without independently validating them for current models.
  • ad hoc to paper Prompt perturbations in Section 4.4 are task-preserving and approximately equal in difficulty to the originals.
    The modifications are designed by the authors based on principles from [24]; if they are not equally difficult, failures would reflect task difficulty rather than lack of robustness. The authors later concede in Section 5 that two modifications may not be trivial.
  • domain assumption Claude with thinking enabled versus disabled provides a valid proxy for the effect of reasoning on performance.
    Section 5.2 states that Claude without thinking is not its base model, and API and interface differences exist; nevertheless, the thinking-on/off comparison is used as key evidence.
  • domain assumption Third-party benchmark results in Table 1 accurately reflect the models' ToM performance.
    The paper reproduces numbers from [9] and uses them as external supporting evidence without re-running or verifying the evaluation protocol.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Evaluating Theory of Mind in Reasoning Models: Robustness over Reasoning." pith.science (2026). https://pith.science/paper/F2CXBDEQ

@misc{pith2026260804646,
  author       = {Pith},
  title        = {Pith review of: Evaluating Theory of Mind in Reasoning Models: Robustness over Reasoning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/F2CXBDEQ}},
  note         = {Machine review of arXiv:2608.04646}
}
read the original abstract

Large language models (LLMs) have recently shown strong performance on Theory of Mind (ToM) tests, prompting debate about the nature and validity of the underlying capabilities. At the same time, reasoning-oriented LLMs trained via reinforcement learning with verifiable rewards have demonstrated notable improvements across a range of benchmarks. In this work, we examine the behavior of such reasoning models in ToM tasks using novel adaptations of machine psychological experiments together with results from established benchmarks. We observe that reasoning models consistently exhibit increased robustness to prompt variations and task perturbations. Our analysis suggests these gains come at least partly from models being more robust at reaching the correct answer under prompt and task variation. We read this as evidence for a robustness-based account rather than for a new ToM-specific ability.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

29 extracted references · 10 canonical work pages

  1. [1]

    theory of mind

    Baron-Cohen, S., Leslie, A.M., Frith, U.: Does the autistic child have a "theory of mind"?Cognition21(1),37–46(Oct1985).https://doi.org/10.1016/0010-0277(85) 90022-8

  2. [2]

    In: International review of research in mental retardation: Autism (vol

    Baron-Cohen, S.: Theory of mind and autism: A review. In: International review of research in mental retardation: Autism (vol. 23), Vol. 23, pp. 169–184. Academic Press, San Diego, CA, US (2001)

  3. [3]

    https://doi.org/10.48550/arXiv.2505.05410, arXiv:2505.05410 [cs] 14 I.B

    Chen, Y., Benton, J., Radhakrishnan, A., Uesato, J., Denison, C., Schulman, J., Somani, A., Hase, P., Wagner, M., Roger, F., Mikulik, V., Bowman, S.R., Leike, J., Kaplan, J., Perez, E.: Reasoning models don’t always say what they think (May 2025). https://doi.org/10.48550/arXiv.2505.05410, arXiv:2505.05410 [cs] 14 I.B. de Haan et al

  4. [4]

    In: Ku, L.W., Martins, A., Srikumar, V

    Chen, Z., Wu, J., Zhou, J., Wen, B., Bi, G., Jiang, G., Cao, Y., Hu, M., Lai, Y., Xiong, Z., Huang, M.: ToMBench: Benchmarking theory of mind in large language models. In: Ku, L.W., Martins, A., Srikumar, V. (eds.) Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). pp. 15959–15983. Association ...

  5. [5]

    In: Bouamor, H., Pino, J., Bali, K

    van Dijk, B., Kouwenhoven, T., Spruit, M., van Duijn, M.J.: Large language models: The need for nuance in current debates and a pragmatic perspective on understanding. In: Bouamor, H., Pino, J., Bali, K. (eds.) Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing. pp. 12641–12654. Association for Computational Linguistics...

  6. [6]

    children aged 7-10 on advanced tests

    van Duijn, M., van Dijk, B., Kouwenhoven, T., de Valk, W., Spruit, M., van der Putten, P.: Theory of mind in large language models: Examining performance of 11 state-of-the-art models vs. children aged 7-10 on advanced tests. In: Jiang, J., Reitter, D., Deng, S. (eds.) Proceedings CoNLL 2023. pp. 389–402. ACL, Singapore (Dec 2023). https://doi.org/10.1865...

  7. [7]

    Nature645(8081), 633–638 (2025)

    Guo, D., Yang, D., Zhang, H., et al.: DeepSeek-R1 incentivizes reasoning in LLMs through reinforcement learning. Nature645(8081), 633–638 (2025). https://doi. org/10.1038/s41586-025-09422-z

  8. [8]

    Philosophical Transactions of the Royal Society B: Biological Sciences380(1932), 20230499 (08 2025)

    Hu, J., Sosa, F., Ullman, T.: Re-evaluating theory of mind evaluation in large language models. Philosophical Transactions of the Royal Society B: Biological Sciences380(1932), 20230499 (08 2025). https://doi.org/10.1098/rstb.2023.0499, https://doi.org/10.1098/rstb.2023.0499

Show all 29 references
  1. [9]

    In: COLM (2025)

    Kim, H., Sclar, M., Zhi-Xuan, T., Ying, L., Levine, S., Liu, Y., Tenenbaum, J.B., Choi, Y.: Hypothesis-driven theory-of-mind reasoning for large language models. In: COLM (2025)

  2. [10]

    https://doi.org/10.48550/arXiv.2310.15421, arXiv:2310.15421 [cs]

    Kim, H., Sclar, M., Zhou, X., Bras, R.L., Kim, G., Choi, Y., Sap, M.: FANToM: A benchmark for stress-testing machine Theory of Mind in interactions (Oct 2023). https://doi.org/10.48550/arXiv.2310.15421, arXiv:2310.15421 [cs]

  3. [11]

    Pro- ceedings of the National Academy of Sciences121(45), e2405460121 (Nov 2024)

    Kosinski, M.: Evaluating Large Language Models in theory of mind tasks. Pro- ceedings of the National Academy of Sciences121(45), e2405460121 (Nov 2024). https://doi.org/10.1073/pnas.2405460121, arXiv:2302.02083 [cs]

  4. [12]

    In: Carl- son,L.A.,Hölscher,C.,Shipley,T.F.(eds.)Proceedingsofthe33thAnnualMeeting of the Cognitive Science Society, CogSci 2011, Boston, Massachusetts, USA, July 20-23, 2011

    Meijering, B., van Rijn, H., Taatgen, N., Verbrugge, R.: I do know what you think I think: Second-order theory of mind in strategic games is not that difficult. In: Carl- son,L.A.,Hölscher,C.,Shipley,T.F.(eds.)Proceedingsofthe33thAnnualMeeting of the Cognitive Science Society,...

  5. [13]

    van der Meulen, R., Verbrugge, R., van Duijn, M.: Towards properly implementing theory of mind in AI systems: An account of four misconceptions (2025), https: //arxiv.org/abs/2503.16468

  6. [14]

    OpenAI: Reasoning models - OpenAI API, https://platform.openai.com

  7. [15]

    https://doi.org/10.48550/arXiv.2502.20339, arXiv:2502.20339 [cs]

    Paliotta, D., Wang, J., Pagliardini, M., Li, K.Y., Bick, A., Kolter, J.Z., Gu, A., Fleuret, F., Dao, T.: Thinking slow, fast: Scaling inference compute with distilled reasoners (Feb 2025). https://doi.org/10.48550/arXiv.2502.20339, arXiv:2502.20339 [cs]

  8. [16]

    Journal of Artificial Intelligence Research84(29) (Dec 2025)

    Plaat, A., van Duijn, M., van Stein, N., Preuss, M., van der Putten, P., Batenburg, K.J.: Agentic large language models, a survey. Journal of Artificial Intelligence Research84(29) (Dec 2025). https://doi.org/10.1613/jair.1.18675 Evaluating Theory of Mind in Reasoning Models: ...

  9. [17]

    ACM Comput

    Plaat, A., Wong, A., Verberne, S., Broekens, J., van Stein, N., Bäck, T.: Multi-step reasoning with large language models, a survey. ACM Comput. Surv.58(6) (Dec 2025). https://doi.org/10.1145/3774896

  10. [18]

    Perspectives on Psychological Science15(2), 384–396 (Mar 2020)

    Quesque, F., Rossetti, Y.: What do Theory-of-Mind tasks actually measure? theory and practice. Perspectives on Psychological Science15(2), 384–396 (Mar 2020). https://doi.org/10.1177/1745691619896607

  11. [19]

    Nature568(7753), 477–486 (Apr 2019)

    Rahwan, I., Cebrian, M., Obradovich, N., Bongard, J., Bonnefon, J.F., Breazeal, C., Crandall, J.W., Christakis, N.A., Couzin, I.D., Jackson, M.O., Jennings, N.R., Kamar, E., Kloumann, I.M., Larochelle, H., Lazer, D., McElreath, R., Mislove, A., Parkes, D.C., Pentland, A.S., Ro...

  12. [20]

    https://doi.org/10.1007/BF03395673

    Schlinger,H.D.:TheoryofMind:Anoverviewandbehavioralperspective.ThePsy- chological Record59(3), 435–448 (Jul 2009). https://doi.org/10.1007/BF03395673

  13. [21]

    In: Graham, Y., Purver, M

    Shapira, N., Levy, M., Alavi, S.H., Zhou, X., Choi, Y., Goldberg, Y., Sap, M., Shwartz, V.: Clever hans or neural Theory of Mind? stress testing social reasoning in Large Language Models. In: Graham, Y., Purver, M. (eds.) Proceedings of EACL 2024.pp.2257–2273.AssociationforCom...

  14. [22]

    Frontiers in Human Neuro- scienceV olume 19 - 2025(2026)

    Street, W., Siy, J.O., Keeling, G., Baranes, A., Barnett, B., McKibben, M., Kanyere, T., Lentz, A., Arcas, B.A.y., Dunbar, R.I.M.: LLMs achieve adult human performance on higher-order theory of mind tasks. Frontiers in Human Neuro- scienceV olume 19 - 2025(2026). https://doi.o...

  15. [23]

    MindLIX(236), 433–460 (Oct 1950)

    Turing, A.M.: Computing machinery and intelligence. MindLIX(236), 433–460 (Oct 1950). https://doi.org/10.1093/mind/LIX.236.433

  16. [24]

    https://doi.org/10.48550/arXiv.2302.08399, arXiv:2302.08399 [cs]

    Ullman, T.: Large Language Models fail on trivial alterations to theory-of-mind tasks (Mar 2023). https://doi.org/10.48550/arXiv.2302.08399, arXiv:2302.08399 [cs]

  17. [25]

    In: Ku, L.W., Martins, A., Srikumar, V

    Wilf, A., Lee, S., Liang, P.P., Morency, L.P.: Think twice: Perspective-taking im- proves large language models’ theory-of-mind capabilities. In: Ku, L.W., Martins, A., Srikumar, V. (eds.) Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics ...

  18. [26]

    https://doi.org/10

    Wu, Y., Xie, J., Zhang, D., Xu, Z.: DEL-ToM: Inference-time scaling for Theory- of-Mind reasoning via dynamic epistemic logic (May 2025). https://doi.org/10. 48550/arXiv.2505.17348, arXiv:2505.17348 [cs]

  19. [27]

    https://doi.org/10

    Xu, F., Hao, Q., Zong, Z., Wang, J., Zhang, Y., Wang, J., Lan, X., Gong, J., Ouyang, T., Meng, F., Shao, C., Yan, Y., Yang, Q., Song, Y., Ren, S., Hu, X., Li, Y., Feng, J., Gao, C., Li, Y.: Towards large reasoning models: A survey of reinforced reasoning with Large Language Mo...

  20. [28]

    https://doi.org/10.48550/arXiv.2504.13837, arXiv:2504.13837 [cs]

    Yue, Y., Chen, Z., Lu, R., Zhao, A., Wang, Z., Yue, Y., Song, S., Huang, G.: Does reinforcement learning really incentivize reasoning capacity in LLMs be- yond the base model? (May 2025). https://doi.org/10.48550/arXiv.2504.13837, arXiv:2504.13837 [cs]

  21. [29]

    https://doi.org/10.48550/arXiv.2303.18223, arXiv:2303.18223 [cs]

    Zhao, W.X., Zhou, K., Li, J., Tang, T., Wang, X., Hou, Y., Min, Y., Zhang, B., Zhang, J., Dong, Z., Du, Y., Yang, C., Chen, Y., Chen, Z., Jiang, J., Ren, R., Li, Y., Tang, X., Liu, Z., Liu, P., Nie, J.Y., Wen, J.R.: A survey of Large Language Models (Mar 2025). https://doi.org...

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.