Reproduce “Proceedings of 1st Workshop on Advancing Artificial Intelligence through Theory of Mind” 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

Mouad Abrini, Omri Abend, Dina Acklin, Henny Admoni, Gregor Aichinger, Nitay Alon, Zahra Ashktorab, Ashish Atreja and 100 more

The volume connects cognitive-science theories of Theory of Mind with AI applications, examining how systems interpret beliefs, intentions, and goals. Its papers compare task performance with tests of coherent mental-state reasoning and propose approaches to human-AI interaction. In one test, GPT-4o's belief estimates showed little correlation with Bayesian-inverted predictions (r = .03), while desire estimates correlated strongly (r = .57), within the tested paradigms.

Key insights

  • Large language models demonstrate emergent, yet often superficial and inconsistent, Theory of Mind (ToM) capabilities, frequently lacking the coherence and robustness of human social reasoning.
  • There's a notable distinction in AI between "theory-based" reasoning for intentions and a lack of "simulation-based" reasoning for tasks like perspective-taking, which often does not improve with model scaling.
  • Advanced ToM in AI introduces significant safety and ethical challenges, including potential for deception, privacy invasion, and multi-agent exploitation, necessitating responsible development and mitigation strategies.

Summary by alphaXiv.

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