Elamur Research⌕
☰
home / papers / 2026 / Let's Put Ourselves in Sally's Shoes: Shoes-of-Others Prefilling Improves Theory of Mind in Large Language Models

Let's Put Ourselves in Sally's Shoes: Shoes-of-Others Prefilling Improves Theory of Mind in Large Language Models

Paper · 2026 · Kazutoshi Shinoda, Nobukatsu Hojo, Kyosuke Nishida et al.

Summary

Large language models still trail humans on Theory of Mind (ToM), and fine-tuning on ToM data tends to hurt out-of-distribution generalization. Existing inference-time fixes mostly target belief tracking in Sally-Anne-style stories where objects change location, so they transfer poorly to other situations. The authors propose Shoes-of-Others (SoO) prefilling. Rather than adding an instruction to the prompt, it seeds the start of the model's answer with a short sentence that asks the model to adopt the target character's viewpoint. The character's name is extracted from the question by rules, and the model then continues the text. The method was tested on ToMATO (conversations) and the English part of ToMBench (narratives), neither of which involves world-state changes. SoO prefilling generally raised accuracy across beliefs, intentions, desires, emotions and knowledge, at both first and second order. It usually beat zero-shot CoT prompting and CoT prefilling. An LLM-as-a-judge analysis links the gains to more faithful reasoning, not simply longer outputs. The approach is training-free and cheap, but it needs an interface that lets users fix the output prefix.

Key points

  • Setup: Both benchmarks are four-option multiple-choice, so the random baseline is 25%. ToMATO has about 5.4k questions and ToMBench about 2.4k. Results are accuracy averaged over three runs.
  • Models: Prefilling was applied to Mistral-7B-Instruct-v0.3, Llama-3-8B-Instruct and Llama-3-70B-Instruct. GPT-3.5 turbo and GPT-4o mini were evaluated with prompting variants only, because OpenAI APIs do not allow output prefixes.
  • Largest gain (Llama3 8B, first-order average): On ToMATO, accuracy rose from 54.9% (Vanilla) to 67.5% with SoO prefilling, versus 64.0% for CoT prefilling. On ToMBench, it rose from 49.2% to 59.2%, versus 54.6% for CoT prefilling.
  • Llama3 70B: SoO prefilling gave the best first-order averages (83.1% on ToMATO, 68.0% on ToMBench). It also led on second-order ToMATO, with a larger benefit on false-belief items.
  • Ablation and analysis: Replacing the character's name with a generic "others" phrasing collapsed ToMATO accuracy to 29.4% or 16.2% (vs. 62.8% with the name). In a GPT-4o mini judge comparison, faithfulness win rates correlated positively with accuracy gains, while thought length did not.

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

🏢NTTTokyo · Japan26 works · 5,083 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 1Research subjects and languages the work is about.Language models & pretraining · 3,320 works

Cite

Cite as Kazutoshi Shinoda, Nobukatsu Hojo, Kyosuke Nishida, Yoshihiro Yamazaki, Keita Suzuki, Hiroaki Sugiyama and Kuniko Saito. 2026. Let's Put Ourselves in Sally's Shoes: Shoes-of-Others Prefilling Improves Theory of Mind in Large Language Models. Findings of the Association for Computational Linguistics: EACL 2026.

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.18653/v1/2026.findings-eacl.6

@inproceedings{shinoda2026let,
  title = {{Let's Put Ourselves in Sally's Shoes: Shoes-of-Others Prefilling Improves Theory of Mind in Large Language Models}},
  author = {Kazutoshi Shinoda, Nobukatsu Hojo, Kyosuke Nishida, Yoshihiro Yamazaki, Keita Suzuki, Hiroaki Sugiyama and Kuniko Saito},
  booktitle = {Findings of the Association for Computational Linguistics: EACL 2026},
  year = {2026},
  doi = {10.18653/v1/2026.findings-eacl.6},
  url = {https://doi.org/10.18653/v1/2026.findings-eacl.6}
}

References

29 references, 9 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, Alan M. Leslie, and Uta Frith. 1985. Does the autistic child have a “theory of mind” ? Cognition, 21(1):37–46. find
  2. Björn Callenmark, Lars Kjellin, Louise Rönnqvist, and Sven Bölte. 2014. Explicit versus implicit social cognition testing in autism spectrum disorder. Autism: The International Journal of Research & Practice, 18(6). find
  3. Mia Xu Chen, Benjamin N. Lee, Gagan Bansal, Yuan Cao, Shuyuan Zhang, Justin Lu, Jackie Tsay, Yinan Wang, Andrew M. Dai, Zhifeng Chen, Timothy Sohn, and Yonghui Wu. 2019. Gmail smart compose: Real-time assisted writing. In KDD, page 2287–2295. find
  4. Zhuang Chen, Jincenzi Wu, Jinfeng Zhou, Bosi Wen, Guanqun Bi, Gongyao Jiang, Yaru Cao, Mengting Hu, Yunghwei Lai, Zexuan Xiong, and Minlie Huang. 2024. ToMBench: Benchmarking theory of mind in large language models. In ACL, pages 15959–15983.
  5. Mark H Davis, Laura Conklin, Amy Smith, and Carol Luce. 1996. Effect of perspective taking on the cognitive representation of persons: a merging of self and other. Journal of personality and social psychology, 70(4):713. find
  6. Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, and 1 others. 2024. The llama 3 herd of models. Preprint, arXiv:2407.21783.
  7. Chris D Frith and Uta Frith. 2006. The neural basis of mentalizing. Neuron, 50(4):531–534. find
  8. Matt Gardner, William Merrill, Jesse Dodge, Matthew Peters, Alexis Ross, Sameer Singh, and Noah A. Smith. 2021. Competency problems: On finding and removing artifacts in language data. In EMNLP, pages 1801–1813. find
  9. Guiyang Hou, Wenqi Zhang, Yongliang Shen, Linjuan Wu, and Weiming Lu. 2024. TimeToM: Temporal space is the key to unlocking the door of large language models’ theory-of-mind. In Findings of ACL, pages 11532–11547. find
  10. Albert Q. Jiang, Alexandre Sablayrolles, Arthur Mensch, Chris Bamford, Devendra Singh Chaplot, Diego de las Casas, Florian Bressand, Gianna Lengyel, Guillaume Lample, Lucile Saulnier, Lélio Renard Lavaud, Marie-Anne Lachaux, Pierre Stock, Teven Le Scao, Thibaut Lavril, Thomas Wang, Timothée Lacroix,… find
  11. 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 EMNLP, pages 14397–14413. find
  12. Takeshi Kojima, Shixiang Shane Gu, Machel Reid, Yutaka Matsuo, and Yusuke Iwasawa. 2022. Large language models are zero-shot reasoners. In NeurIPS.
  13. Tamera Lanham, Anna Chen, Ansh Radhakrishnan, Benoit Steiner, Carson Denison, Danny Hernandez, Dustin Li, Esin Durmus, Evan Hubinger, Jackson Kernion, Kamilė Lukošiūtė, Karina Nguyen, Newton Cheng, Nicholas Joseph, Nicholas Schiefer, Oliver Rausch, Robin Larson, Sam McCandlish, Sandipan Kundu, an… — Measuring Faithfulness in Chain-of-Thought Reasoning
  14. Matthew Le, Y-Lan Boureau, and Maximilian Nickel. 2019. Revisiting the evaluation of theory of mind through question answering. In EMNLP, pages 5872– 5877. find
  15. Qing Lyu, Shreya Havaldar, Adam Stein, Li Zhang, Delip Rao, Eric Wong, Marianna Apidianaki, and Chris Callison-Burch. 2023. Faithful chain-of-thought reasoning. In IJCNLP-AACL, pages 305– 329.
  16. 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 EMNLP, pages 1011–1031. find
  17. OpenAI. 2024. Gpt-4 technical report. arXiv:2303.08774.
  18. Preprint, David Premack and Guy Woodruff. 1978. Does the chimpanzee have a theory of mind? Behavioral and Brain Sciences, 1(4):515–526. find
  19. Xiangyu Qi, Ashwinee Panda, Kaifeng Lyu, Xiao Ma, Subhrajit Roy, Ahmad Beirami, Prateek Mittal, and Peter Henderson. 2025. Safety alignment should be made more than just a few tokens deep. In ICLR. find
  20. Perrine Ruby and Jean Decety. 2004. How would you feel versus how do you think she would feel? a neuroimaging study of perspective-taking with social emotions. Journal of Cognitive Neuroscience, 16(6):988–999. find
  21. 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 EMNLP, pages 3762–3780. find
  22. Sneheel Sarangi, Maha Elgarf, and Hanan Salam. 2025. Decompose-ToM: Enhancing theory of mind reasoning in large language models through simulation and task decomposition. In COLING, pages 10228– 10241.
  23. Tobias Schuwerk, Maria Vuori, and Beate Sodian. 2015. Implicit and explicit theory of mind reasoning in autism spectrum disorders: The impact of experience. Autism: The International Journal of Research and Practice, 19(4):459–468. find
  24. 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 ACL, pages 13960–13980.
  25. Natalie Shapira, Mosh Levy, Seyed Hossein Alavi, Xuhui Zhou, Yejin Choi, Yoav Goldberg, Maarten Sap, and Vered Shwartz. 2024. Clever hans or neural theory of mind? stress testing social reasoning in large language models. In EACL, pages 2257–2273. find
  26. Ashish Sharma, Inna W. Lin, Adam S. Miner, David C. Atkins, and Tim Althoff. 2021. Towards facilitating empathic conversations in online mental health support: A reinforcement learning approach. In WWW, page 194–205. find
  27. Kazutoshi Shinoda, Nobukatsu Hojo, Kyosuke Nishida, Saki Mizuno, Keita Suzuki, Ryo Masumura, Hiroaki Sugiyama, and Kuniko Saito. 2025. Tomato: Verbalizing the mental states of role-playing llms for benchmarking theory of mind. In AAAI, pages 1520– 1528. \\\# Options [A] {{option1}} [B] {{option2}} [C]… find
  28. Charlie Snell, Jaehoon Lee, Kelvin Xu, and Aviral Kumar. 2024. Scaling llm test-time compute optimally can be more effective than scaling model parameters. Preprint, arXiv:2408.03314.
  29. Chat Messages for Evaluation on ToMBench Tomer Ullman. 2023. Large language models fail on trivial alterations to theory-of-mind tasks. Preprint, arXiv:2302.08399. find

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/shinoda-2026-let-s-put-ourselves-sally/