UniToMBench: Integrating Perspective-Taking to Improve Theory of Mind in LLMs

Chris Kim; Kevin Zhu; Nitin Yarlagadda; Prameshwar Thiyagarajan; Shamant Sai; Soumil Garg; Vaishnavi Parimi; Zhangir Meirbek

REVIEW 4 major objections 5 minor 19 references

UniToMBench: Integrating Perspective-Taking to Improve Theory of Mind in LLMs

T0 review · 4 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 UniToMBench integrates perspective-taking to improve and assess Theory of Mind in LLMs.

desk verdict A useful new ToM dataset, but the paper's central improvement claim is contradicted by its own tables. read the letter →

arxiv 2506.09450 v1 pith:FFWBTUFH submitted 2025-06-11 cs.CL cs.AI

Prameshwar Thiyagarajan , Vaishnavi Parimi , Shamant Sai , Soumil Garg , Zhangir Meirbek , Nitin Yarlagadda , Kevin Zhu , Chris Kim This is my paper

classification cs.CLcs.AI

keywords TheoryofMindLLMevaluationperspective-takingSimToMTOMBENCHsocialcognitionmulti-agentreasoningbenchmarkdesign

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

UniToMBench is a Theory of Mind (ToM) benchmark that combines perspective-taking prompts, multi-turn dialogues, and evolving story scenarios. The paper argues that this combination—SimToM's viewpoint-taking prompting, TOMBENCH-style task categories, and 1,025 hand-written multiple-choice items—improves language models' detection of irony, sarcasm, and hidden emotions, and gives a fuller read of their social reasoning than static single-turn tests. On the reported evaluations, GPT-4o and GPT-4o Mini stay above 80% on many emotion and belief tasks, while knowledge-based tasks vary widely, and the smaller open models usually lose accuracy when perspective-taking is added. The value, if the benchmark works as described, is a single testbed for both improving and measuring machine social cognition.

What carries the argument

The load-bearing mechanism is SimToM (Simulation Theory of Mind) prompting: the model is first prompted to take a specific character's perspective, and only then answers the ToM question. UniToMBench embeds that mechanism in two custom scenario families—500 multi-interaction tasks that track mental states across conversational turns, and 525 evolving-story tasks that track motivation and relationships across time—while keeping TOMBENCH's eight task categories for comparability. The evaluation procedure measures task-completion accuracy, error attribution, and run-to-run consistency at temperature 0.7.

What would settle it

Present the same 1,025 multiple-choice scenarios to human adults and to models after every mental-state verb has been replaced with a neutral verb, and after swapping which character knows which fact; if adult scores are at ceiling and model accuracy is unchanged by the swaps, the items measure literal comprehension, not Theory of Mind.

Watch

Extended reading notes

Core claim

The central claim is that adding the SimToM perspective-taking framework to a unified benchmark changes measured Theory-of-Mind performance in a task-dependent way, and that the change is informative. Specifically, the paper claims that multi-interaction tasks and evolving stories—supported by SimToM prompting—improve LLM performance in detecting irony, sarcasm, and hidden emotions, while static or knowledge-heavy tasks like faux-pas recognition and some belief-tracking tests remain weak points. The evidence is organized as accuracy over 1,025 custom scenarios plus TOMBENCH categories, with error attribution and consistency as secondary metrics. The authors present the benchmark as a tool to both stimulate and evaluate social cognition, not as a claim that models have genuine ToM.

Load-bearing premise

The 1,025 hand-written scenarios are valid, unbiased measures of Theory of Mind, with unambiguous right answers and no way to solve them by story recall or surface wording alone.

Editorial extensions

If this is right

SimToM should be applied selectively, not universally: it raised GPT-3.5 Turbo's Hinting Test score from 24.3% to 68.9%, but it lowered accuracy for Llama 3 8B and Gemma 2 27B on evolving-story and multi-interaction tasks.
Static ToM tests understate model ability: models that appear weak on one-shot false-belief questions can show better belief tracking when tested across evolving dialogues, and vice versa.
The 80%-plus emotion and belief scores of GPT-4o and GPT-4o Mini, paired with wide variation on knowledge-based tasks, identify where future training should concentrate.
Error patterns—missing implicit cues, conflating mental states across agents, and failing at abstract reasoning—give concrete targets for new benchmark tasks.

Reading between the lines

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

Editorial inference: without a human baseline on the same 1,025 items, the 80%+ scores cannot be read as human-level ToM; collecting adult norms would calibrate the benchmark.
Editorial inference: the cake example is solvable from stated preference rather than belief tracking, so a rewording that requires false-belief reasoning would test whether these items measure ToM at all.
Editorial inference: swapping which character knows which fact in the custom scenarios would show whether SimToM-enhanced models track distinct mental states or exploit story templates.

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

Quick take: this is a benchmark paper whose central claim is undercut by its own results. The custom dataset is a real contribution, but the evidence that integrating SimToM 'enhances' ToM performance is mixed at best and mostly negative.

What's actually new: 1,025 hand-written scenarios in two formats—multi-interaction dialogues and evolving stories—that address a genuine gap in static one-shot ToM benchmarks. That is a useful piece of work, and the authors correctly identify in Section 3 that SimToM can overgeneralize and hurt performance. The prose is clear, and the public code link helps.

The soft spots are serious. Section 5 reports single runs at temperature 0.7 with no error bars, repeated trials, or significance tests. The stress-test note is right: the direction of the effect is not consistently positive. On the custom set, SimToM essentially doesn't change GPT-4o's evolving-story score (89.9 vs 90.1) and lowers it for every other model except GPT-4o-Mini's multi-interaction score (55.8 to 65.8). On TOMBENCH, GPT-4o's false belief drops from 89.5 to 79.5, GPT-4o-Mini declines on seven of eight tasks, and GPT-3.5's unexpected outcome crashes from 34.7 to 10.0. The conclusion says integration enhances detection of irony, sarcasm, and hidden emotions, but the supporting tables tell a different story. The paper's own Section 3 lists reasons SimToM can hurt; the abstract and conclusion ignore those caveats.

The other soft spot is construct validity. There is no human baseline, no inter-annotator agreement, and the provided cake example is a preference judgment, not a clear belief or emotion inference. That makes it hard to know what the custom accuracy numbers measure. This is a secondary concern relative to the main claim, but it matters if the benchmark is to be trusted.

Who this is for: people building or evaluating ToM benchmarks. The dataset could be a starting point, and the task formats are worth discussing. But as submitted, the improvement claim is not demonstrated. I'd send it to a serious referee, because the resource itself deserves scrutiny and the field needs more diverse ToM tasks, but I'd expect either major revision or rejection on the evidence.

Referee Report

4 major / 5 minor

Summary. The paper introduces UniToMBench, a benchmark that augments TOMBENCH with 1,025 hand-written multi-interaction and evolving-story scenarios, and evaluates five LLMs with and without the SimToM perspective-taking prompting framework. The authors report accuracy percentages on both TOMBENCH tasks and their custom tasks, and conclude that integrating multi-interaction tasks, evolving stories, and SimToM enhances LLM performance in detecting irony, sarcasm, and hidden emotions. The paper includes a public code repository and uses established TOMBENCH items for comparability.

Significance. If the benchmark and the SimToM integration were validated, UniToMBench could be a useful resource for studying Theory of Mind in LLMs: it combines a custom multi-turn dataset with an existing benchmark, evaluates multiple open and proprietary models, and releases code publicly. However, the central empirical claim is not supported by the reported numbers, and the dataset's validity as a Theory-of-Mind instrument is unestablished. As it stands, the paper does not provide reliable evidence that SimToM improves ToM performance or that the custom scenarios measure mental-state reasoning rather than literal comprehension.

major comments (4)

[Section 5, Tables 1 and 6] The conclusion that SimToM 'enhances LLM performance' is contradicted by the paper's own results. GPT-4o's False Belief Task accuracy drops from 89.5 to 79.5, its Multi-Interaction accuracy drops from 71.2 to 63.0, GPT-3.5-Turbo's Evolving Stories accuracy drops from 69.5 to 55.8, and GPT-4o-Mini declines on seven of eight TOMBENCH tasks (e.g., False Belief Task 72.0 to 67.0, Strange Story Task 79.4 to 69.0, Hinting Test 74.8 to 68.9). With single runs at temperature 0.7 (Section 4.2), no confidence intervals, and no significance tests, the few improvements that do appear are within plausible sampling noise. The paper needs repeated runs, error bars, and paired statistical tests before any improvement claim can be evaluated.
[Section 4.1, Dataset Composition] The construct validity of the 1,025 custom scenarios is not established. There is no human baseline, no inter-annotator agreement, and no psychometric validation. The 'cake' example is ambiguous: Tom prefers chocolate but thinks strawberry is easier, and the question 'Which cake would Tom choose?' does not specify whether he decides as a tie-breaker or for his own preference, so answer B is not uniquely determined by the story. The 'Evolving Story' example about Aaron's photography mostly tests literal comprehension of a sequence, not mental-state reasoning. Without evidence that the gold answers are unambiguous and that the tasks require ToM, the accuracy numbers in Table 6 cannot be interpreted as measures of Theory of Mind.
[Section 4.3, Evaluation Metrics] The paper lists 'consistency across runs' as one of the three evaluation metrics, but no consistency data are reported anywhere in the paper. All models are queried once at temperature 0.7 (Section 4.2), which makes it impossible to assess reproducibility. Either the metric should be removed or the corresponding experiments and results should be added.
[Sections 6 and 7, Discussion and Conclusion] The narrative overstates the findings. The Discussion says SimToM-enhanced models 'excel in maintaining consistency across multi-turn interactions,' but GPT-4o's Multi-Interaction accuracy falls from 71.2 to 63.0 and Llama 3 8B's falls from 74.4 to 68.5. The Conclusion claims improvements in detecting irony and sarcasm, yet the TOMBENCH Non-Literal Communication (Irony/Sarcasm) results are mixed or unreported in the main tables. The claims should be scaled back to match the data.

minor comments (5)

[Section 2 headings] There are typos in section headings: 'Evaulating ToM Capabilities' and 'Methods for Enchancing ToM' should be 'Evaluating' and 'Enhancing'; 'SimTom' in the related-work paragraph should be 'SimToM'.
[Table 1 caption] The caption says 'The averages of the performance scores from all tests are presented,' but the table shows individual task rows and no average row; please clarify what averages, if any, are reported.
[Appendix A, Tables 2-5] The categories 'Intentions Explanations' appear twice in each table (e.g., 77%/71%/28% and 83.1%/70.8%/76.4% for GPT models), with no distinction between the two entries; this appears to be a labeling error and should be corrected.
[Section 4.2, Experimental Setup] The text says 'All models were queried at a temperature of 0.7 for consistency,' but a temperature above zero makes outputs stochastic; without seeds or multiple runs, temperature does not by itself ensure consistency.
[Section 3, Goal] The 'Goal' paragraph is set as an unformatted heading; it should be formatted consistently with the rest of the section structure.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: all reported quantities are raw accuracy measurements; the SimToM/TOMBENCH integration is not reduced to a fitted parameter or author-imported uniqueness result.

full rationale

The paper's chain is empirical rather than derivational. It assembles a hand-written dataset of 1,025 scenarios (Section 4.1), evaluates five LLMs with and without the SimToM prompt at temperature 0.7 (Section 4.2), and reports percentages (Tables 1-6). No equation defines the reported accuracy in terms of the benchmark's own labels, and no parameter is fitted to model outputs and then re-presented as a prediction. The load-bearing comparison is between baseline and SimToM-conditioned runs on the same items; because the items and gold answers are fixed before model inference, any improvement or decline is a measured outcome, not a built-in consequence of the construction. The two cited frameworks, SimToM (Wilf et al., 2023) and TOMBENCH (Chen et al., 2024), are prior external works with no author overlap with this paper, so there is no self-citation chain importing an ansatz or uniqueness result. Section 8 itself concedes limited scope, and Section 5 shows mixed results (e.g., GPT-4o drops from 71.2 to 63.0 on Multi-Interaction tasks in Table 6), which is a threat to the empirical enhancement claim but not to circularity. I therefore find no specific reduction of a derived quantity to an input, and score 0.

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

The central claim rests on the validity of the custom dataset and the single-run evaluation. No free parameters are fitted to data beyond the arbitrary temperature setting; the main assumptions are that the hand-written scenarios measure ToM and that the TOMBENCH subset used is representative.

free parameters (1)

temperature = 0.7
Evaluation temperature chosen without sensitivity analysis; results may vary across temperatures. The paper does not show whether conclusions hold for other temperatures.

assumptions (4)

domain assumption Accuracy on four-option multiple-choice questions is a valid measure of Theory of Mind.
The benchmark concludes LLMs' ToM ability from accuracy on MCQs; no human performance or psychometric validation is provided. Section 4.1.
domain assumption The hand-written scenarios (1,025) cover diverse and representative social situations.
The authors claim diversity and iterative refinement but provide no external validation, inter-annotator agreement, or comparison to existing ToM tasks. Section 4.1.
domain assumption SimToM prompting was correctly implemented across all models.
The paper does not provide the exact prompts or code in the appendix, so replication depends on the public repo without a commit hash. Section 4.2.
domain assumption The TOMBENCH subset used is comparable to the full benchmark.
The paper says it incorporates questions from TOMBENCH but does not specify which subset or how many items per category. Section 4.1.

how reviews work

0 comments

Cite this review

Pith. "Pith review of UniToMBench: Integrating Perspective-Taking to Improve Theory of Mind in LLMs." pith.science (2026). https://pith.science/paper/FFWBTUFH

@misc{pith2026250609450,
  author       = {Pith},
  title        = {Pith review of: UniToMBench: Integrating Perspective-Taking to Improve Theory of Mind in LLMs},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/FFWBTUFH}},
  note         = {Machine review of arXiv:2506.09450}
}

read the original abstract

Theory of Mind (ToM), the ability to understand the mental states of oneself and others, remains a challenging area for large language models (LLMs), which often fail to predict human mental states accurately. In this paper, we introduce UniToMBench, a unified benchmark that integrates the strengths of SimToM and TOMBENCH to systematically improve and assess ToM capabilities in LLMs by integrating multi-interaction task designs and evolving story scenarios. Supported by a custom dataset of over 1,000 hand-written scenarios, UniToMBench combines perspective-taking techniques with diverse evaluation metrics to better stimulate social cognition in LLMs. Through evaluation, we observe that while models like GPT-4o and GPT-4o Mini show consistently high accuracy in tasks involving emotional and belief-related scenarios, with results usually above 80%, there is significant variability in their performance across knowledge-based tasks. These results highlight both the strengths and limitations of current LLMs in ToM-related tasks, underscoring the value of UniToMBench as a comprehensive tool for future development. Our code is publicly available here: https://github.com/Shamant/unifiedtombenchmark.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

19 extracted references · 3 canonical work pages

[1]

Sébastien Bubeck, Varun Chandrasekaran, Ronen Eldan, Johannes Gehrke, Eric Horvitz, Ece Kamar, Peter Lee, Yin Tat Lee, Yuanzhi Li, Scott Lundberg, Harsha Nori, Hamid Palangi, Marco Tulio Ribeiro, and Yi Zhang. 2023. https://arxiv.org/abs/2303.12712 Sparks of artificial general intelligence: Early experiments with gpt-4 . Preprint, arXiv:2303.12712

arXiv 2023
[2]

Zhuang Chen, Jincenzi Wu, Jinfeng Zhou, Bosi Wen, Guanqun Bi, Gongyao Jiang, Yaru Cao, Mengting Hu, Yunghwei Lai, Zexuan Xiong, and Minlie Huang. 2024. https://arxiv.org/abs/2402.15052 Tombench: Benchmarking theory of mind in large language models . Preprint, arXiv:2402.15052

arXiv 2024
[3]

Leyang Cui, Sijie Cheng, Yu Wu, and Yue Zhang. 2021. https://doi.org/10.18653/v1/2021.findings-acl.61 On commonsense cues in BERT for solving commonsense tasks . In Findings of the Association for Computational Linguistics: ACL-IJCNLP 2021, pages 683--693, Online. Association for Computational Linguistics

work page doi:10.18653/v1/2021.findings-acl.61 2021
[4]

M. U. Hadi, Q. A. Tashi, A. Shah, R. Qureshi, A. Muneer, M. Irfan, A. Zafar, M. B. Shaikh, N. Akhtar, J. Wu, S. Mirjalili, and M. Shah. 2024. Large language models: a comprehensive survey of its applications, challenges, limitations, and future prospects. 10.36227/techrxiv.23589741.v6

doi:10.36227/techrxiv.23589741.v6 2024
[5]

Michal Kosinski. 2024. https://doi.org/10.1073/pnas.2405460121 Evaluating large language models in theory of mind tasks . Proceedings of the National Academy of Sciences, 121(45)

doi:10.1073/pnas.2405460121 2024
[6]

Nunzio Lore, Sepehr Ilami, and Babak Heydari. 2024. https://arxiv.org/abs/2408.05241 Large model strategic thinking, small model efficiency: Transferring theory of mind in large language models . Preprint, arXiv:2408.05241

arXiv 2024
[7]

Shima Rahimi Moghaddam and Christopher J. Honey. 2023. https://arxiv.org/abs/2304.11490 Boosting theory-of-mind performance in large language models via prompting . Preprint, arXiv:2304.11490

arXiv 2023
[8]

OpenAI, Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, Red Avila, Igor Babuschkin, Suchir Balaji, Valerie Balcom, Paul Baltescu, Haiming Bao, Mohammad Bavarian, Jeff Belgum, Irwan Bello, Jake Berdine, Gabriel Bernadett-Shapiro, Christopher Berner...

arXiv 2024

Show all 19 references

[9]

Hannes Rakoczy. 2022. https://doi.org/10.1038/s44159-022-00037-z Foundations of theory of mind and its development in early childhood . Nature Reviews Psychology, 1(4):211--222

2022 doi
[10]

Sneheel Sarangi, Maha Elgarf, and Hanan Salam. 2025. https://arxiv.org/abs/2501.09056 Decompose-tom: Enhancing theory of mind reasoning in large language models through simulation and task decomposition . Preprint, arXiv:2501.09056

2025 arXiv
[11]

Murray Shanahan. 2023. https://arxiv.org/abs/2212.03551 Talking about large language models . Preprint, arXiv:2212.03551

2023 arXiv
[12]

Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, Dan Bikel, Lukas Blecher, Cristian Canton Ferrer, Moya Chen, Guillem Cucurull, David Esiobu, Jude Fernandes, Jeremy Fu, W...

2023 arXiv
[13]

Sean Trott, Cameron Jones, Tyler Chang, James Michaelov, and Benjamin Bergen. 2023. https://doi.org/10.1111/cogs.13309 Do large language models know what humans know? Cognitive Science, 47(7):e13309

2023 doi
[14]

Tomer Ullman. 2023. https://arxiv.org/abs/2302.08399 Large language models fail on trivial alterations to theory-of-mind tasks . Preprint, arXiv:2302.08399

2023 arXiv
[15]

Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Brian Ichter, Fei Xia, Ed Chi, Quoc Le, and Denny Zhou. 2023. https://arxiv.org/abs/2201.11903 Chain-of-thought prompting elicits reasoning in large language models . Preprint, arXiv:2201.11903

2023 arXiv
[16]

Alex Wilf, Sihyun Shawn Lee, Paul Pu Liang, and Louis-Philippe Morency. 2023. https://arxiv.org/abs/2311.10227 Think twice: Perspective-taking improves large language models' theory-of-mind capabilities . Preprint, arXiv:2311.10227

2023 arXiv
[17]

Wenfeng Zheng, Siyu Lu, Zhuohang Cai, Ruiyang Wang, Lei Wang, and Lirong Yin. 2024. https://doi.org/10.32604/cmes.2023.046692 Pal-bert: An improved question answering model . Computer Modeling in Engineering and Sciences, pages 1--10

2024
[18]

online" 'onlinestring :=

ENTRY address archivePrefix author booktitle chapter edition editor eid eprint eprinttype howpublished institution journal key month note number organization pages publisher school series title type volume year doi pubmed url lastchecked label extra.label sort.label short.list...
[19]

write newline

" write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 gl...

Pith tools

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

[1] [1]

Sébastien Bubeck, Varun Chandrasekaran, Ronen Eldan, Johannes Gehrke, Eric Horvitz, Ece Kamar, Peter Lee, Yin Tat Lee, Yuanzhi Li, Scott Lundberg, Harsha Nori, Hamid Palangi, Marco Tulio Ribeiro, and Yi Zhang. 2023. https://arxiv.org/abs/2303.12712 Sparks of artificial general intelligence: Early experiments with gpt-4 . Preprint, arXiv:2303.12712

arXiv 2023

[2] [2]

Zhuang Chen, Jincenzi Wu, Jinfeng Zhou, Bosi Wen, Guanqun Bi, Gongyao Jiang, Yaru Cao, Mengting Hu, Yunghwei Lai, Zexuan Xiong, and Minlie Huang. 2024. https://arxiv.org/abs/2402.15052 Tombench: Benchmarking theory of mind in large language models . Preprint, arXiv:2402.15052

arXiv 2024

[3] [3]

Leyang Cui, Sijie Cheng, Yu Wu, and Yue Zhang. 2021. https://doi.org/10.18653/v1/2021.findings-acl.61 On commonsense cues in BERT for solving commonsense tasks . In Findings of the Association for Computational Linguistics: ACL-IJCNLP 2021, pages 683--693, Online. Association for Computational Linguistics

work page doi:10.18653/v1/2021.findings-acl.61 2021

[4] [4]

M. U. Hadi, Q. A. Tashi, A. Shah, R. Qureshi, A. Muneer, M. Irfan, A. Zafar, M. B. Shaikh, N. Akhtar, J. Wu, S. Mirjalili, and M. Shah. 2024. Large language models: a comprehensive survey of its applications, challenges, limitations, and future prospects. 10.36227/techrxiv.23589741.v6

doi:10.36227/techrxiv.23589741.v6 2024

[5] [5]

Michal Kosinski. 2024. https://doi.org/10.1073/pnas.2405460121 Evaluating large language models in theory of mind tasks . Proceedings of the National Academy of Sciences, 121(45)

doi:10.1073/pnas.2405460121 2024

[6] [6]

Nunzio Lore, Sepehr Ilami, and Babak Heydari. 2024. https://arxiv.org/abs/2408.05241 Large model strategic thinking, small model efficiency: Transferring theory of mind in large language models . Preprint, arXiv:2408.05241

arXiv 2024

[7] [7]

Shima Rahimi Moghaddam and Christopher J. Honey. 2023. https://arxiv.org/abs/2304.11490 Boosting theory-of-mind performance in large language models via prompting . Preprint, arXiv:2304.11490

arXiv 2023

[8] [8]

OpenAI, Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, Red Avila, Igor Babuschkin, Suchir Balaji, Valerie Balcom, Paul Baltescu, Haiming Bao, Mohammad Bavarian, Jeff Belgum, Irwan Bello, Jake Berdine, Gabriel Bernadett-Shapiro, Christopher Berner...

arXiv 2024