Evaluating Large Language Models in Theory of Mind Tasks

Written for the Research library (claude-opus-5-5) from the paper text, in our own words: the license "CC BY-NC-ND 4.0 (arXiv)" does not allow redistributing the text itself. Original — arxiv.org/pdf/2302.02083

Summary

Theory of mind (ToM) is the ability to attribute hidden mental states, such as beliefs or intentions, to other people. It underpins human social behaviour. Earlier language models, including RoBERTa and early GPT-3, failed simple ToM tests. Kosinski asks whether ToM could appear in LLMs without anyone building it in, simply as a side effect of learning language. To avoid contamination from training data, a research assistant who did not know the hypothesis wrote 40 new false-belief tasks: 20 Unexpected Contents and 20 Unexpected Transfer. Each task pairs one false-belief story with three closely matched true-belief controls, plus reversed versions of all four. A model scores a point only if it answers all eight correctly. Eleven models, from GPT-1 to ChatGPT-4 and Bloom, completed sentence-continuation prompts at temperature 0, and the first word of each completion was scored. Accuracy grew with each model generation. ChatGPT-4 solved three quarters of the tasks, about the level reported for six-year-old children in earlier studies. The author argues that machines which act as if they track minds matter socially, however the mechanism is explained.

Key points

• Each task has two prompt types: one asks for the real state of the world (e.g., what the container holds), the other for the protagonist's belief. Stories use the label and content words equally often, so word frequency alone cannot reveal the answer.
• Models tested: GPT-1, GPT-2, six GPT-3 family models, ChatGPT-3.5-turbo, ChatGPT-4 (June 2023) and the open-access Bloom.
• Results: older models solved 0% of tasks. GPT-3-davinci-003 (Nov 2022) and ChatGPT-3.5-turbo (Mar 2023) each solved 20%, and ChatGPT-4 solved 75%.
• Contamination safeguards: the scenarios were drafted in August 2021 and released in February 2023. Only models trained on data from before that release were tested, through APIs that did not keep inputs for training.
• Tasks, instructions and code are on OSF (osf.io/csdhb). The 2023 arXiv preprint was later revised and published in PNAS in 2024 under a CC BY-NC-ND licence.