Theory of Mind in Large Language Models: Examining Performance of 11 State-of-the-Art models vs. Children Aged 7-10 on Advanced Tests

Max J. van Duijn, Bram van Dijk, Tom Kouwenhoven, Werner de Valk, Marco Spruit, Peter van der Putten

Theory of Mind in Large Language Models: Examining Performance of 11 State-of-the-Art models vs. Children Aged 7-10 on Advanced Tests

Introduction

Theory of Mind (ToM)—the capacity to infer the mental states, intentions, and beliefs of others—is a foundational element of human social intelligence. As Large Language Models (LLMs) become increasingly integrated into human-centric applications, determining whether they possess genuine ToM or merely simulate it through pattern recognition is critical. This research evaluates 11 state-of-the-art LLMs against a human baseline of children aged 7–10, focusing on advanced tasks that require reasoning about non-literal language and recursive intentionality.

Methods and Datasets

The study benchmarked 11 models, categorized into Base-LLMs (e.g., BLOOM, LLaMA-30B) and Instruct-LLMs (e.g., GPT-4, PaLM2-chat). The evaluation utilized three standardized psychological batteries:

  1. Sally-Anne Tests: Assessing 1st and 2nd order false beliefs.
  2. Strange Stories: Evaluating the interpretation of white lies, irony, and sarcasm.
  3. Imposing Memory Task: Testing recursive intentionality up to five levels (e.g., "A thinks that B believes...").

To distinguish between memorization and reasoning, the researchers introduced three levels of "Deviations":

  • Deviation 0: Original, well-known test text.
  • Deviation 1: Superficial modifications (e.g., changing names).
  • Deviation 2: Novel scenarios with identical logical structures.

The human baseline consisted of 73 children (ages 7–10) from international schools. Responses were open-ended and scored by expert human raters on a 0–2 scale.

Analysis of Model Performance

A primary finding is the critical role of instruction tuning. Base models, regardless of scale, consistently failed ToM tasks, often performing below the level of 7-year-old children. This suggests that the human-in-the-loop feedback used to train Instruct-LLMs may inadvertently simulate the social pressures that drive ToM development in humans.

Overall performance trends across items and deviations for various LLMs.

Furthermore, the study revealed a significant "fragility" in AI reasoning. While models like GPT-4 demonstrated high proficiency on standard tests (Deviation 0), their performance frequently declined when presented with novel scenarios (Deviation 2). This performance drop suggests that even advanced models rely on sophisticated pattern matching—or "clever Hans" heuristics—rather than deep, robust social reasoning.

Key Results

GPT-4 emerged as the most capable model, frequently outperforming children, particularly in complex recursive intentionality tasks where human performance typically plateaus. However, the "GPT family" (GPT-4 and GPT-3.5-turbo) maintained a substantial lead over competitors like PaLM2 and LLaMA.

Performance of Base and Instruct LLMs on Sally-Anne tasks (SA1 and SA2) across different levels of deviation.

In qualitative assessments, instruct-tuned models excelled at identifying linguistic tropes in "Strange Stories," often matching or exceeding child-level performance in explaining the social motivations behind white lies or sarcasm. Conversely, models found the logical displacement in Sally-Anne tasks more challenging than interpreting social nuances.

Discussion and Limitations

The study highlights that while LLMs can achieve superhuman performance on specific ToM benchmarks, their "intelligence" remains brittle and potentially tied to training data exposure. Data contamination remains a significant hurdle in LLM evaluation. Future research should investigate whether techniques like Chain-of-Thought prompting can enhance robustness and whether LLMs are developing an "alien" form of social intelligence that mimics human outcomes through fundamentally different mechanisms.

Summary updated

BibTeX
@inproceedings{duijn2023theory,
  author = {Max J. van Duijn and Bram van Dijk and Tom Kouwenhoven and Werner de Valk and Marco Spruit and Peter van der Putten},
  title = {{Theory of Mind in Large Language Models: Examining Performance of 11 State-of-the-Art models vs. Children Aged 7-10 on Advanced Tests}},
  booktitle = {CoNLL},
  year = {2023},
  url = {https://openreview.net/forum?id=tBsmgctGwZ},
}

Loading related content...

Figure