Elamur Research⌕
☰
home / papers / 2025 / Decompose-ToM: Enhancing Theory of Mind Reasoning in Large Language Models through Simulation and Task Decomposition

Decompose-ToM: Enhancing Theory of Mind Reasoning in Large Language Models through Simulation and Task Decomposition

Paper · 2025 · Sneheel Sarangi, Maha Elgarf and Hanan Salam

Summary

Large language models still struggle with Theory of Mind (ToM) reasoning, especially higher-order questions (beliefs about other agents' beliefs) and dialogue scenarios where participants join and leave. The authors propose Decompose-ToM, an inference-time prompting algorithm inspired by two ideas from developmental psychology: pretend play (Simulation Theory) and knowledge access. The method first identifies whose perspective a question depends on and rewrites the question from that character's viewpoint. It then goes through the story one statement at a time, asking the LLM whether the character could know each statement. An LLM-maintained text world state of locations and objects supports these judgments. Only the known statements are kept as a reduced story. This repeats recursively until the question becomes purely factual, which is answered with chain-of-thought.

Tested on Hi-ToM and FANToM with four LLMs, the approach yields large gains on third- and fourth-order questions. It also keeps accuracy stable on long dialogues. It needs no fine-tuning and little task-specific prompt adjustment, but its per-statement LLM calls make it computationally costly.

Key points

  • Setup: Four models (Llama-3-8B, Llama-3-70B, Gemini-1.5-Flash, GPT-4o) were compared under zero-shot multiple-choice prompting, zero-shot CoT, SimToM, and Decompose-ToM.
  • Hi-ToM (600 questions, 15 answer options, up to 4th-order ToM, "Tell"/"no-Tell" categories): on fourth-order questions, Decompose-ToM reaches 87.5% with Llama-3-70B and 83.3% with GPT-4o. These are gains of about 44 and 40 points over the strongest baseline. Gains for the smaller models are modest or mixed.
  • FANToM (1,540 binary-choice belief questions, first and second order, short and long dialogues): results are roughly on par with SimToM for most models. Gemini-1.5-Flash is the exception, beating SimToM by 6.65% on average. GPT-4o scores 88.4% (short) and 86.2% (long).
  • Long contexts: the accuracy gap between long and short FANToM stories shrinks to 0.9%. The gap is about 7.6–7.7% for MC and CoT prompting and 4% for SimToM.
  • Limitations: the method assumes agents' knowledge is never revised backward, so it handles non-chronological stories poorly. It is also expensive because it processes every statement separately. The code is publicly released.

Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.

 Institutions

🎓New York UniversityNew York · United States201 works · 100,180 citations🎓New York University Abu DhabiAbu Dhabi · United Arab Emirates122 works · 4,602 citations

 Fields, topics and concepts

📚 Shelves 7The field of science this work belongs to (OpenAlex fields).Computer Science · 16,667 worksArtificial Intelligence · 16,194 worksArts and Humanities · 16,039 worksLanguage and Linguistics · 15,891 worksComputer Vision and Pattern Recognition · 2,869 worksPsychology · 1,552 worksDevelopmental and Educational Psychology · 499 works

🧭 Topics 2Research subjects and languages the work is about.Language models & pretraining · 3,320 worksReasoning · 1,426 works

Cite

Cite as Sneheel Sarangi, Maha Elgarf and Hanan Salam. 2025. Decompose-ToM: Enhancing Theory of Mind Reasoning in Large Language Models through Simulation and Task Decomposition. Proceedings of the 31st International Conference on Computational Linguistics.

The paper itself, not this page. BibTeX from the library's record of the work.

@inproceedings{sarangi2025decompose,
  title = {{Decompose-ToM: Enhancing Theory of Mind Reasoning in Large Language Models through Simulation and Task Decomposition}},
  author = {Sneheel Sarangi, Maha Elgarf and Hanan Salam},
  booktitle = {Proceedings of the 31st International Conference on Computational Linguistics},
  year = {2025},
  url = {https://aclanthology.org/2025.coling-main.682/}
}

References

38 references, 10 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Simon Baron-Cohen. 1995. Mindblindness: An Essay on Autism and Theory of Mind. The MIT Press. find
  2. Simon Baron-Cohen, Alan Leslie, and Uta Frith. 1985. Does the autistic child have a theory of mind? Cognition, 21:37–46. find
  3. Wayne Christensen and John Michael. 2012. Ian apperly, mindreaders: the cognitive basis of theory of mind: Psychology press, 2011, 232 pages, (isbn 978-1-84169-697-3) 80.00. Phenomenology and the Cognitive Sciences, 12(4):907–914. find
  4. Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. 2024. The llama 3 herd of models. arXiv preprint arXiv:2407.21783.
  5. Michael C Frank and Noah D Goodman. 2012. Predicting pragmatic reasoning in language games. Science, 336(6084):998–1002.
  6. Kanishk Gandhi, Jan-Philipp Fränken, Tobias Gerstenberg, and Noah D. Goodman. 2023. Understanding social reasoning in language models with language models. Preprint, arXiv:2306.15448. find
  7. Jian Jiang, Viswonathan Manoranjan, Hanan Salam, and Oya Celiktutan. 2024. Towards generalised and incremental bias mitigation in personality computing. IEEE Transactions on Affective Computing. find
  8. R D Kavanaugh. 2006. Pretend play and theory of mind. In L S Balter C, editor, Child psychology: A handbook of contemporary issues, pages 153–166. Psychology Press. find
  9. Hyunwoo Kim, Melanie Sclar, Xuhui Zhou, Ronan Bras, Gunhee Kim, Yejin Choi, and Maarten Sap. 2023. FANToM: A benchmark for stress-testing machine theory of mind in interactions. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 14397–14413, Singapore. A… find
  10. Peter Kinderman, Robin Dunbar, and Richard P. Bentall. 1998. Theory-of-mind deficits and causal attributions. British Journal of Psychology, 89(2):191–204. find
  11. Michal Kosinski. 2023. Theory of mind may have spontaneously emerged in large language models. Preprint, arXiv:2302.02083.
  12. Woosuk Kwon, Zhuohan Li, Siyuan Zhuang, Ying Sheng, Lianmin Zheng, Cody Hao Yu, Joseph E. Gonzalez, Hao Zhang, and Ion Stoica. 2023. Efficient memory management for large language model serving with pagedattention. In Proceedings of the ACM SIGOPS 29th Symposium on Operating Systems Principles.
  13. Matthew Le, Y-Lan Boureau, and Maximilian Nickel. 2019. Revisiting the evaluation of theory of mind through question answering. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-I… find
  14. Alan M. Leslie. 1987. Pretense and representation: The origins of “theory of mind.”. Psychological Review, 94(4):412–426. find
  15. Jialin Li, Alia Waleed, and Hanan Salam. 2023. A survey on personalized affective computing in human-machine interaction. arXiv preprint arXiv:2304.00377. — [arXiv:2304.00377](https://arxiv.org/abs/2304.00377) find
  16. Bethany Liddle and Daniel Nettle. 2006. Higher-order theory of mind and social competence in school-age children. Journal of Cultural and Evolutionary Psychology Akadémiai Kiadó, 43:1589–5254. find
  17. Angeline S Lillard. 1993. Pretend play skills and the child’s theory of mind. Child Dev., 64(2):348. find
  18. Ziqiao Ma, Jacob Sansom, Run Peng, and Joyce Chai. 2023. Towards a holistic landscape of situated theory of mind in large language models. In Findings of the Association for Computational Linguistics: EMNLP 2023, pages 1011–1031, Singapore. Association for Computational Linguistics. find
  19. Rachel L C Mitchell and Louise H Phillips. 2015. The overlapping relationship between emotion perception and theory of mind. Neuropsychologia, 70:1–10. Copyright © 2015 Elsevier Ltd. All rights reserved. find
  20. Shima Rahimi Moghaddam and Christopher J. Honey. 2023. Boosting theory-of-mind performance in large language models via prompting. Preprint, arXiv:2304.11490. — [arXiv:2304.11490](https://arxiv.org/abs/2304.11490) find
  21. Aida Nematzadeh, Kaylee Burns, Erin Grant, Alison Gopnik, and Tom Griffiths. 2018. Evaluating theory of mind in question answering. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pages 2392–2400, Brussels, Belgium. Association for Computational Linguistics… find
  22. OpenAI. 2023. Gpt-4 technical report. arXiv preprint arXiv:2303.08774.
  23. David Premack and Guy Woodruff. 1978. Does the chimpanzee have a theory of mind? Behavioral and Brain Sciences, 1(4):515–526. find
  24. Li Qu, Pinxiu Shen, Yu Yan Chee, and Luxi Chen. 2015. Teachers’ theory-of-mind coaching and children’s executive function predict the training effect of socio-dramatic play on children’s theory of mind. Social Development, 24(4):716–733. find
  25. Machel Reid, Nikolay Savinov, Denis Teplyashin, Dmitry Lepikhin, Timothy Lillicrap, Jean-baptiste Alayrac, Radu Soricut, Angeliki Lazaridou, Orhan Firat, Julian Schrittwieser, et al. 2024. Gemini 1.5: Unlocking multimodal understanding across millions of tokens of context. arXiv preprint arXiv:2403.…
  26. Hanan Salam, Oya Celiktutan, Hatice Gunes, and Mohamed Chetouani. 2023. Automatic context-aware inference of engagement in hmi: A survey. IEEE Transactions on Affective Computing. find
  27. Hanan Salam, Viswonathan Manoranjan, Jian Jiang, and Oya Celiktutan. 2022. Learning personalised models for automatic self-reported personality recognition. In Understanding Social Behavior in Dyadic and Small Group Interactions, pages 53–73. PMLR. find
  28. Maarten Sap, Ronan Le Bras, Daniel Fried, and Yejin Choi. 2022. Neural theory-of-mind? on the limits of social intelligence in large LMs. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 3762–3780, Abu Dhabi, United Arab Emirates. Association for Compu… find
  29. Natalie Shapira, Mosh Levy, Seyed Hossein Alavi, Xuhui Zhou, Yejin Choi, Yoav Goldberg, Maarten Sap, and Vered Shwartz. 2023. Clever hans or neural theory of mind? stress testing social reasoning in large language models. Preprint, arXiv:2305.14763. find
  30. Freda Shi, Xinyun Chen, Kanishka Misra, Nathan Scales, David Dohan, Ed Chi, Nathanael Schärli, and Denny Zhou. 2023. Large language models can be easily distracted by irrelevant context. Preprint, arXiv:2302.00093.
  31. Tomer Ullman. 2023. Large language models fail on trivial alterations to theory-of-mind tasks. Preprint, arXiv:2302.08399. find
  32. Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Brian Ichter, Fei Xia, Ed H Chi, Quoc V Le, and Denny Zhou. 2022. Chain of thought prompting elicits reasoning in large language models. Advances in Neural Information Processing Systems, 35:24824– 24837.
  33. Henry M. Wellman and David Liu. 2004. Scaling of theory-of-mind tasks. Child Development, 75(2):523–541. find
  34. Alex Wilf, Sihyun Lee, Paul Pu Liang, and Louis-Philippe Morency. 2024. Think twice: Perspective-taking improves large language models’ theory-of-mind capabilities. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 8292–8308, Ba…
  35. Yufan Wu, Yinghui He, Yilin Jia, Rada Mihalcea, Yulong Chen, and Naihao Deng. 2023. Hi-ToM: A benchmark for evaluating higher-order theory of mind reasoning in large language models. In Findings of the Association for Computational Linguistics: EMNLP 2023, pages 10691–10706, Singapore. Association f… find
  36. Hainiu Xu, Runcong Zhao, Lixing Zhu, Jinhua Du, and Yulan He. 2024. OpenToM: A comprehensive benchmark for evaluating theory-of-mind reasoning capabilities of large language models. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pa… find
  37. Melanie Sclar, Sachin Kumar, Peter West, Alane Suhr, Yejin Choi, and Yulia Tsvetkov. 2023. Minding language models’ (lack of) theory of mind: A plug-and-play multi-character belief tracker. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Pap…
  38. Karen Shanton and Alvin Goldman. 2010. Simulation theory. WIREs Cognitive Science, 1(4):527–538. find

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/sarangi-2025-decompose-tom-enhancing-theory-mind/