Reproduce “Probing the Robustness of Theory of Mind in Large Language Models” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Christian Nickel, Laura Schrewe, Lucie Flek

Researchers at the University of Bonn developed a new, systematically varied dataset to assess the robustness of Theory of Mind (ToM) in large language models (LLMs). The study found that while LLMs achieve moderate accuracy on individual ToM-related questions, they consistently fail to demonstrate a comprehensive understanding of entire scenarios, especially when faced with specific linguistic or conceptual complexities.

Key insights

  • LLMs exhibit a significant disparity between their moderate turn accuracy on individual ToM questions and near-zero goal accuracy, indicating a lack of comprehensive situational understanding.
  • The 'automatic change knowledge' complexity class and tasks involving 'preposition replacement' proved particularly challenging for all evaluated LLMs, suggesting weaknesses in handling unobserved environmental changes and spatial reasoning.
  • Model performance generally increased with the number of parameters, but Yi-34B-Chat performed remarkably close to Llama-2-70B, suggesting that model size is not the sole determinant of ToM capabilities.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.