A Systematic Review on the Evaluation of Large Language Models in Theory of Mind Tasks

Karahan Sar{\i}ta\c{s}; K{\i}van\c{c} Tez\"oren; Yavuz Durmazkeser

REVIEW 2 major objections 5 minor 14 references

A Systematic Review on the Evaluation of Large Language Models in Theory of Mind Tasks

T0 review · 2 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read A systematic review of 58 studies concludes that large language models' apparent Theory of Mind is limited, fragile under small task changes, and likely driven by spurious correlations rather than genuine mental-state understanding.

desk verdict Useful map of LLM ToM evaluation, but the 'systematic' label overpromises and the skeptical conclusion ignores the strongest counterexample sitting in its own benchmark table. read the letter →

arxiv 2502.08796 v1 pith:VYGBFZHV submitted 2025-02-12 cs.CL cs.CYcs.HC

Karahan Sar{i}taş , K{i}vanç Tezören , Yavuz Durmazkeser This is my paper · ORCID

classification cs.CLcs.CYcs.HC

keywords theoryofmindlargelanguagemodelssystematicreviewbenchmarksevaluationmethodologyATOMStaxonomypromptingstrategiesspuriouscorrelations

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 systematic review assembles 58 studies on whether large language models can pass Theory of Mind tasks—tests that require attributing beliefs, desires, intentions, emotions, perceptions, knowledge, and non-literal meaning to others. The authors organize the field with a taxonomy borrowed from developmental psychology (the seven-category ATOMS framework) plus a situatedness classification, and use that structure to compare benchmarks, metrics, prompting strategies, and error types. Their synthesis sides with the skeptical camp: larger, newer models score higher on ToM benchmarks than their predecessors, but they still trail humans on most tasks, and their success is fragile under small task variations, consistent with reliance on shallow heuristics and spurious correlations rather than a genuine model of other minds. The review therefore cautions against reading current benchmark scores as evidence that LLMs possess human-like mental-state reasoning.

What carries the argument

The organizing tool is a cognitive-science taxonomy: ATOMS (Abilities in Theory of Mind Space), the seven mental-state categories—intentions, desires, emotions, knowledge, percepts, beliefs, and non-literal communication—taken from developmental psychology, combined with a situatedness grid (no agent, passive perceiver, active interactor; physical versus social) and a task-format classification (multiple choice, true/false, question answering, text completion, inference, natural language generation, multi-agent collaboration). The paper's extended benchmark table, built on an earlier taxonomized review and benchmark list, uses this grid to place every reviewed benchmark, and the same categories structure the synthesis of findings, prompting techniques, error analyses, and evaluation challenges.

What would settle it

A direct test would be to take a model that scores at or above human level on a published ToM benchmark and run it on a contamination-controlled, procedurally generated battery spanning all seven ATOMS categories, interleaved with Ullman-style perturbations such as transparent containers, uninformative additions, and shuffled options. If accuracy stays high and the model continues to match or beat humans under every variation, the skeptical conclusion is falsified; if accuracy collapses, the 'spurious correlations' interpretation is supported.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central finding is that the literature does not support the claim that LLMs have robust Theory of Mind. Across the 58 reviewed papers, models show emerging competence—higher-parameter and newer models generally outperform earlier ones—but performance is uneven across the ATOMS categories, falls below human baselines in most direct comparisons (often by 20% or more), and degrades sharply when test prompts are trivially altered, such as with uninformative additions, transparent containers, or reordered options. The authors conclude that LLMs' apparent ToM 'may often rely on spurious correlations rather than robust understanding' and that benchmark scores should be interpreted skeptically. They also document recurrent failure modes—conservative refusals, hallucinated story details, insufficient reasoning depth, commonsense errors, temporal ignorance, and spurious causal inference—and catalog evaluation hazards including training contamination, prompt bias, shortcut learning, limited scope, and manual scoring that can inflate apparent competence.

Load-bearing premise

The review's conclusions rest on the assumption that its corpus of 58 papers, drawn from an arXiv keyword search plus one maintained repository and intentionally capped in size, fairly represents the state of theory-of-mind research on large language models; a skewed or incomplete corpus would change the synthesis.

Editorial extensions

If this is right

Current high scores on ToM benchmarks should not be read as evidence that LLMs understand mental states; the review's skeptical conclusion implies benchmark leaderboards overstate machine ToM.
Progress across model generations is real but bounded: newer and larger models outperform predecessors yet still fall below human performance in most studies, so human-level ToM is not established.
Evaluation practice should shift toward broader coverage of ATOMS categories, situated and symmetric multi-agent settings, automatic grading validated against human judgment, and safeguards against training contamination and prompt or option-order bias.
Prompting frameworks such as SimToM, TimeToM, PercepToM, FaR, and DWM can lift scores, but their gains are selective and their comparative effectiveness is largely unmeasured, so they do not resolve the underlying skepticism.
The documented failure modes—hallucinated beliefs, conservative refusals, temporal ignorance, and spurious causal inference—give concrete targets for future training and evaluation design.

Reading between the lines

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

If the skeptical conclusion is right, ToM accuracy should be treated as a stress-testable behavior rather than an attribute: benchmark reports should include perturbation checks such as Ullman-style variations, option shuffling, and added irrelevant context before any claim about mental-state understanding is made.
The review's evidence suggests a common root for two seemingly separate failures: leaking private information when humans would stay silent and lagging in intention inference during games both require tracking others' unobservable mental states. A testable extension is to check whether prompting methods that improve ToM also reduce privacy leakage.
A practical research program follows from the synthesis: generate ToM tasks procedurally so training contamination cannot explain scores, and report accuracy across all seven ATOMS categories rather than false-belief tasks alone.
Because most reviewed studies use GPT-based models and text-only input, the field's conclusions may not transfer to open-source or multimodal models; extending the same taxonomy to those settings is an implicit gap revealed by the review's own statistics.

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.

Desk Editor's Note

read the letter

Worth a look if you work on ToM evaluation. The paper's real contribution is the extended benchmark table, which adds resource links, a more detailed task taxonomy, and explicit rules for classifying situatedness. The descriptive statistics (GPT dominance, text-only bias) are useful, and the discussion of evaluation metrics, prompting, and challenges like contamination and shortcut learning is well organized and mostly accurate. This is the kind of resource a new grad student could pick up and get oriented from.

The soft spots are real but fixable. First, the "systematic review" label overpromises: one database (arXiv), no protocol, and a self-imposed cap of 58 papers. That is a structured narrative review, not a systematic review in the PRISMA sense. Second, and more seriously, the paper's own Table 1 lists MOTOMQA (Street et al., 2024), which reports adult-human-level performance on higher-order ToM tasks, yet Section 4 never discusses Street et al. or even acknowledges the result. The conclusion claims LLMs "still fall short of human performance" and rely on spurious correlations; that is contradicted by an included benchmark. A review that aims to be systematic must weigh all included evidence, and skipping the strongest counterexample leaves the conclusion unsupported within the paper's own corpus. That is a genuine internal flaw, not an external representativeness quibble. Third, the situatedness citation points to Cao (2021), which is a paper on locative alternation in linguistics, not on ToM; that looks like a wrong reference.

These problems are eminently correctable. The skeptical stance itself is defensible, and the authors are transparent about their search limitations and their interpretive synthesis. The paper is not incoherent; it is just incomplete and overlabeled.

Who gets value: practitioners entering this subfield who want a benchmark map and a catalog of evaluation pitfalls. It deserves a serious referee, but the referee should demand that the authors integrate Street et al. into the qualitative synthesis or justify its exclusion, fix the Cao citation, and tone down the systematic claim. I would send it to peer review requiring major revisions, not desk reject.

Referee Report

2 major / 5 minor

Summary. This manuscript is a systematic review of evaluations of Large Language Models on Theory of Mind tasks. It extends a benchmark taxonomy from Ma et al. (2023c), classifies benchmarks by ATOMS mental states and situatedness, surveys evaluation metrics, prompting strategies, fine-tuning approaches, systematic failures, and evaluation challenges, and concludes that LLMs show emerging but limited ToM abilities that often rely on spurious correlations rather than robust understanding. The paper's main contributions are an expanded benchmark table, a narrative synthesis of 58 papers retrieved from the arXiv API and the Ma et al. repository, and a discussion of methodological pitfalls such as training contamination, prompt bias, and shortcut learning.

Significance. If the synthesis were internally consistent, this review would be a useful reference for researchers navigating the rapidly growing Machine Theory of Mind literature. The benchmark table, the taxonomy of situatedness, and the tabulation of prompting frameworks are practical contributions. The paper also correctly emphasizes important evaluation challenges, such as training contamination and prompt-induced biases, and it explicitly adopts a skeptical position that is a live and defensible stance in the field. However, the review's central conclusion is weakened by an internal inconsistency: one of the benchmarks included in its own Table 1 reports adult-human-level performance, yet the qualitative synthesis does not discuss or reconcile this result. The methodological basis for the 'systematic review' label is also thin, as the search is limited to a single repository, a single query, and a self-imposed cap of 58 papers.

major comments (2)

[Section 6 and Table 1] The paper's conclusion states that LLMs 'still fall short of human performance' in most ToM tasks and aligns with the skeptical perspective. However, Table 1 includes MOTOMQA (Street et al., 2024), whose reported result is that LLMs achieve adult human performance on higher-order theory-of-mind tasks. Section 4, the qualitative literature review, never discusses Street et al. or any other study with results at or above human level. For a systematic review, the synthesis must weigh all included evidence; omitting a directly contradicting result leaves the central conclusion unsupported within the authors' own corpus. Please either incorporate Street et al. into the synthesis, explain why its finding does not generalize, or qualify the conclusion to acknowledge the countervailing evidence.
[Section 4] The search methodology described in Section 4 does not support the 'systematic review' label. The authors used only the arXiv API with the single query "theory of mind" AND ("LLM" OR "large language models"), supplemented by the Ma et al. repository, and then 'intentionally limited the number of papers' to 58. No protocol, inclusion/exclusion criteria, or PRISMA-style reporting is provided, and the single-repository search cannot be expected to capture the full literature, which is spread across ACL, EMNLP, NeurIPS, and other venues. This is load-bearing because the review's synthesis and its skeptical conclusion depend on the representativeness of the selected corpus. The paper should either substantially expand the search and document a reproducible protocol, or be reframed as a narrative and taxonomic review rather than a systematic review.

minor comments (5)

[Throughout] The in-text citation 'C. et al. (2020)' is nonstandard and should be replaced by the actual author name (Beaudoin et al., 2020) to match the reference list.
[Section 1 and Section 3.1] The reference to Cao (2021) for 'situatedness' and as a foundation for the task taxonomy is puzzling: the listed reference is a self-paced reading paper on locative alternation, not a Theory of Mind paper. Please verify the citation and use the correct source.
[Table 1] Table 1 is dense and the column headings are difficult to parse in the rendered text, especially the dual 'Per. Int.' columns and the 'Sym.' column. A clear legend with full column names would improve readability.
[Section 4] The text says 'we reviewed a total of 58 papers related to the Machine Theory of Mind,' but not all 58 appear in Table 1 or are individually discussed in Section 4. Please clarify how the 58 papers map to the benchmark table and the qualitative synthesis.
[Section 5.1] There are several typographical issues, such as 'In Table 2, we list the commonly used prompting techniques... In general, few-shot prompting appears to enhance...' where a new paragraph begins mid-sentence, and 'the model is guided to break down complex tasks into a sequence of intermediate reasoning steps. This approach helps...' where the sentence is split awkwardly. A careful proofread is needed.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity found: the review's conclusions are interpretive syntheses of external empirical work, not quantities defined in terms of the paper's own outputs.

full rationale

This is a systematic literature review with no fitted parameters, no mathematical derivation, and no quantity defined in terms of another quantity within the paper. The review's central outputs are a taxonomized benchmark table and a narrative synthesis of 58 external papers. The taxonomy is explicitly attributed to earlier external work (C. et al. 2020; Ma et al. 2023c), and the paper marks its own extensions to the benchmark table rather than presenting them as first-principles results. The skeptical conclusion in Section 6 is an interpretive summary of the cited empirical literature, not a consequence of the paper's own definitions or a parameter fitted to a subset of data. The stated search limitation in Section 4 ('we intentionally limited the number of papers') is a corpus-selection limitation and a potential quality concern, but it does not create a definitional circularity. Likewise, any imbalance in how included papers are weighted in the qualitative synthesis is a correctness concern, not a circular reasoning step. There are no load-bearing self-citations and no imported uniqueness theorems. The derivation chain, such as it is, runs from external benchmarks and results to an interpretive conclusion, with no step that reduces to its own inputs by construction.

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

This is a review, so no free parameters or invented entities appear. The review's claims rest on the validity of the ATOMS taxonomy and on the representativeness and accurate reporting of the 58 selected papers.

assumptions (4)

domain assumption The ATOMS taxonomy of seven Theory of Mind categories is a valid and sufficient framework for classifying ToM benchmarks.
The review organizes all benchmarks and task formulations through ATOMS (Section 2, Section 3.1) without validating or questioning the taxonomy's completeness.
ad hoc to paper The 58 papers retrieved from the arXiv API query plus the Ma et al. repository are representative of the current Machine Theory of Mind literature.
Section 4 states the search was limited to arXiv and the number of papers was intentionally capped; this selection is a premise for the review's synthesis, not a justified sampling frame.
domain assumption Reported results in the reviewed papers are accurately described by the review's summaries.
The review's conclusions aggregate quantitative and qualitative findings from 58 papers; the authors did not re-run experiments or independently verify them.
ad hoc to paper The operationalized situatedness categories (physical and social perceiver or interactor) are meaningful and can be assigned reliably to benchmarks.
Section 3.1 introduces classification rules where Ma et al. (2023c) had not; the review depends on these rules for its extended taxonomy.

how reviews work

0 comments

Cite this review

Pith. "Pith review of A Systematic Review on the Evaluation of Large Language Models in Theory of Mind Tasks." pith.science (2026). https://pith.science/paper/VYGBFZHV

@misc{pith2026250208796,
  author       = {Pith},
  title        = {Pith review of: A Systematic Review on the Evaluation of Large Language Models in Theory of Mind Tasks},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/VYGBFZHV}},
  note         = {Machine review of arXiv:2502.08796}
}

read the original abstract

In recent years, evaluating the Theory of Mind (ToM) capabilities of large language models (LLMs) has received significant attention within the research community. As the field rapidly evolves, navigating the diverse approaches and methodologies has become increasingly complex. This systematic review synthesizes current efforts to assess LLMs' ability to perform ToM tasks, an essential aspect of human cognition involving the attribution of mental states to oneself and others. Despite notable advancements, the proficiency of LLMs in ToM remains a contentious issue. By categorizing benchmarks and tasks through a taxonomy rooted in cognitive science, this review critically examines evaluation techniques, prompting strategies, and the inherent limitations of LLMs in replicating human-like mental state reasoning. A recurring theme in the literature reveals that while LLMs demonstrate emerging competence in ToM tasks, significant gaps persist in their emulation of human cognitive abilities.

Figures

Figures reproduced from arXiv: 2502.08796 by the authors.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

14 extracted references · 12 canonical work pages

[5]

Susanne A

Language models show human-like content effects on reasoning tasks. Susanne A. Denham. 1986. Social cognition, prosocial behavior, and emotion in preschoolers: contextual validation. Child Development, 57:194–201. Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. Bert: Pre-training of deep bidirectional transformers for language unde...

arXiv 1986
[6]

Rabeeh Karimi Mahabadi, Yonatan Belinkov, and James Henderson

Perceptions to beliefs: Exploring precursory inferences for theory of mind in large language mod- els. Rabeeh Karimi Mahabadi, Yonatan Belinkov, and James Henderson. 2020. End-to-end bias mitigation by modelling biases in corpora. In Proceedings of the 58th Annual Meeting of the Association for Compu- tational Linguistics, pages 8706–8716, Online. Asso- c...

work page 2020
[10]

Violation of expectation via metacognitive prompting reduces theory of mind prediction error in large language models. Joel Z. Leibo, Edgar Duéñez-Guzmán, Alexander Sasha Vezhnevets, John P. Agapiou, Peter Sunehag, Raphael Koster, Jayd Matyas, Charles Beattie, Igor Mordatch, and Thore Graepel. 2021. Scalable evaluation of multi-agent reinforcement learnin...

work page 2021
[11]

In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing

Theory of mind for multi-agent collaboration via large language models. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing. Association for Computational Linguistics. Yuan Li, Yue Huang, Hongyi Wang, Xiangliang Zhang, James Zou, and Lichao Sun. 2024. Quantifying ai psychology: A psychometrics benchmark for large lang...

work page 2023
[12]

john thinks that mary thinks that

Boosting theory-of-mind performance in large language models via prompting. Henrike Moll, Cornelia Koring, Malinda Carpenter, and Michael Tomasello. 2006. Infants determine others’ focus of attention by pragmatics and exclusion. Jour- nal of Cognition and Development, 7(3):411–430. Sonia Murthy, Kiera Parece, Sophie Bridgers, Peng Qian, and Tomer Ullman. ...

work page 2006
[13]

I’m sorry to hear that

A trip towards fairness: Bias and de-biasing in large language models. Shauli Ravfogel, Yanai Elazar, Hila Gonen, Michael Twiton, and Yoav Goldberg. 2020. Null it out: Guard- ing protected attributes by iterative nullspace projec- tion. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 7237–7256, Online. Ass...

arXiv 2020
[14]

Benchmarking large language models for news summarization. Tony Z. Zhao, Eric Wallace, Shi Feng, Dan Klein, and Sameer Singh. 2021. Calibrate before use: Improv- ing few-shot performance of language models. Chujie Zheng, Hao Zhou, Fandong Meng, Jie Zhou, and Minlie Huang. 2024. Large language models are not robust multiple choice selectors. Lianmin Zheng,...

work page 2021
[1985]

theory of mind

Does the autistic child have a “theory of mind” ? Cognition, 21(1):37–46. Simon Baron-Cohen, Michelle O’Riordan, Valerie Stone, Rosie Jones, and Kate Plaisted. 1999. Recog- nition of faux pas by normally developing chil- dren and children with asperger syndrome or high- functioning autism. Journal of Autism and Develop- mental Disorders, 29:407–418. Emily...

work page 1999

Show all 14 references

[2009]

Annals of the New York Academy of Sciences, 1167:103–114

Empathy in early childhood: genetic, environ- mental, and affective contributions. Annals of the New York Academy of Sciences, 1167:103–114. Michal Kosinski. 2024. Evaluating large language mod- els in theory of mind tasks. Eliza Kosoy, Emily Rose Reagan, Leslie Lai, Alison Go...

2024
[2019]

Revisiting the evaluation of theory of mind through question answering. In Proceedings of the 2019 Conference on Empirical Methods in Natu- ral Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 5872–5877, Hong K...

2019
[2020]

Beaudoin C., Leblanc É., Gagner C., and Beauchamp M

Language models are few-shot learners. Beaudoin C., Leblanc É., Gagner C., and Beauchamp M. H. 2020. Systematic review and inventory of theory of mind measures for young children. Frontiers in Psychology, 10:2905. Rui Cao. 2021. Holistic interpretation in locative al- ternatio...

2020
[2021]

In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 1112–1125

Mindcraft: Theory of mind modeling for situ- ated dialogue in collaborative tasks. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 1112–1125. Simon Baron-Cohen, Alan M. Leslie, and Uta Frith

2021
[2023]

Jan-Christoph Klie, Bonnie Webber, and Iryna Gurevych

Fantom: A benchmark for stress-testing ma- chine theory of mind in interactions. Jan-Christoph Klie, Bonnie Webber, and Iryna Gurevych. 2022. Annotation error detection: Analyz- ing the past and present for a more coherent future. Ariel Knafo, Carolyn Zahn-Waxler, Maayan David...

2022
[2024]

Tolga Bolukbasi, Kai-Wei Chang, James Zou, Venkatesh Saligrama, and Adam Kalai

How well can llms negotiate? negotiation- arena platform and analysis. Tolga Bolukbasi, Kai-Wei Chang, James Zou, Venkatesh Saligrama, and Adam Kalai. 2016. Man is to computer programmer as woman is to homemaker? debiasing word embeddings. Michael Brambring and Doreen Asbrock....

2016

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.

[1] [5]

Susanne A

Language models show human-like content effects on reasoning tasks. Susanne A. Denham. 1986. Social cognition, prosocial behavior, and emotion in preschoolers: contextual validation. Child Development, 57:194–201. Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. Bert: Pre-training of deep bidirectional transformers for language unde...

arXiv 1986

[2] [6]

Rabeeh Karimi Mahabadi, Yonatan Belinkov, and James Henderson

Perceptions to beliefs: Exploring precursory inferences for theory of mind in large language mod- els. Rabeeh Karimi Mahabadi, Yonatan Belinkov, and James Henderson. 2020. End-to-end bias mitigation by modelling biases in corpora. In Proceedings of the 58th Annual Meeting of the Association for Compu- tational Linguistics, pages 8706–8716, Online. Asso- c...

work page 2020

[3] [10]

Violation of expectation via metacognitive prompting reduces theory of mind prediction error in large language models. Joel Z. Leibo, Edgar Duéñez-Guzmán, Alexander Sasha Vezhnevets, John P. Agapiou, Peter Sunehag, Raphael Koster, Jayd Matyas, Charles Beattie, Igor Mordatch, and Thore Graepel. 2021. Scalable evaluation of multi-agent reinforcement learnin...

work page 2021

[4] [11]

In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing

Theory of mind for multi-agent collaboration via large language models. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing. Association for Computational Linguistics. Yuan Li, Yue Huang, Hongyi Wang, Xiangliang Zhang, James Zou, and Lichao Sun. 2024. Quantifying ai psychology: A psychometrics benchmark for large lang...

work page 2023

[5] [12]

john thinks that mary thinks that

Boosting theory-of-mind performance in large language models via prompting. Henrike Moll, Cornelia Koring, Malinda Carpenter, and Michael Tomasello. 2006. Infants determine others’ focus of attention by pragmatics and exclusion. Jour- nal of Cognition and Development, 7(3):411–430. Sonia Murthy, Kiera Parece, Sophie Bridgers, Peng Qian, and Tomer Ullman. ...

work page 2006

[6] [13]

I’m sorry to hear that

A trip towards fairness: Bias and de-biasing in large language models. Shauli Ravfogel, Yanai Elazar, Hila Gonen, Michael Twiton, and Yoav Goldberg. 2020. Null it out: Guard- ing protected attributes by iterative nullspace projec- tion. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 7237–7256, Online. Ass...

arXiv 2020

[7] [14]

Benchmarking large language models for news summarization. Tony Z. Zhao, Eric Wallace, Shi Feng, Dan Klein, and Sameer Singh. 2021. Calibrate before use: Improv- ing few-shot performance of language models. Chujie Zheng, Hao Zhou, Fandong Meng, Jie Zhou, and Minlie Huang. 2024. Large language models are not robust multiple choice selectors. Lianmin Zheng,...

work page 2021

[8] [1985]

theory of mind

Does the autistic child have a “theory of mind” ? Cognition, 21(1):37–46. Simon Baron-Cohen, Michelle O’Riordan, Valerie Stone, Rosie Jones, and Kate Plaisted. 1999. Recog- nition of faux pas by normally developing chil- dren and children with asperger syndrome or high- functioning autism. Journal of Autism and Develop- mental Disorders, 29:407–418. Emily...

work page 1999