[Paper Review] Evaluating Large Language Models in Theory of Mind Tasks

Michał Kosiński|arXiv (Cornell University)|Feb 4, 2023

Topic Modeling132 citations

TL;DR

The study evaluates 11 LLMs on a battery of 640 false-belief prompts across 40 tasks; GPT-4 solves 75% of tasks, approaching six-year-old child performance, while smaller models solve none and earlier models reach ~20%.

ABSTRACT

Eleven Large Language Models (LLMs) were assessed using a custom-made battery of false-belief tasks, considered a gold standard in testing Theory of Mind (ToM) in humans. The battery included 640 prompts spread across 40 diverse tasks, each one including a false-belief scenario, three closely matched true-belief control scenarios, and the reversed versions of all four. To solve a single task, a model needed to correctly answer 16 prompts across all eight scenarios. Smaller and older models solved no tasks; GPT-3-davinci-003 (from November 2022) and ChatGPT-3.5-turbo (from March 2023) solved 20% of the tasks; ChatGPT-4 (from June 2023) solved 75% of the tasks, matching the performance of six-year-old children observed in past studies. We explore the potential interpretation of these findings, including the intriguing possibility that ToM, previously considered exclusive to humans, may have spontaneously emerged as a byproduct of LLMs' improving language skills.

Motivation & Objective

• Assess whether large language models exhibit theory-of-mind (ToM) abilities in a battery of false-belief tasks.
• Compare performance across eleven LLMs from small/older to state-of-the-art models.
• Quantify the emergence of ToM-like skills as models improve language capabilities.

Proposed method

• Use a custom battery of false-belief tasks (640 prompts across 40 tasks).
• Each task includes a false-belief scenario, three matched true-belief controls, and reversed versions of all four scenarios.
• A model must answer 16 prompts correctly across eight scenarios to solve a single task.
• Evaluation spans eleven LLMs with varying sizes and architectures.
• Results discuss performance relative to human benchmarks from prior ToM studies.
• Code and tasks are made available for replication (Colab).

Experimental results

Research questions

• RQ1 How do eleven LLMs perform on a battery of theory-of-mind false-belief tasks?
• RQ2 Does ToM performance improve with newer, more capable language models?
• RQ3 Are ToM-like capabilities emergent as a byproduct of language modeling progress?
• RQ4 How does model performance compare to human ToM benchmarks from prior research?

Key findings

• Smaller and older LLMs solved no tasks in the battery.
• GPT-3-davinci-003 (Nov 2022) and ChatGPT-3.5-turbo (Mar 2023) solved about 20% of the tasks.
• ChatGPT-4 (Jun 2023) solved about 75% of the tasks.
• GPT-4’s performance matched the level observed in past studies of six-year-old children.
• The results raise the possibility that ToM may emerge spontaneously as language skills improve in LLMs.