This systematic review examines how large language models (LLMs) perform on Theory of Mind (ToM) tasks—the ability to attribute mental states to oneself and others.

• Current research reveals diverse methodologies for evaluating ToM capabilities in AI systems
• LLMs show promising but inconsistent abilities to understand beliefs, intentions, and perspectives
• Standardized evaluation frameworks are emerging but remain fragmented
• Performance on ToM tasks varies significantly across different model architectures and training approaches

Medical Relevance: Understanding ToM in AI systems has significant implications for medical applications, including better modeling of cognitive disorders like autism and schizophrenia, and developing more empathetic healthcare AI assistants.

A Systematic Review on the Evaluation of Large Language Models in Theory of Mind Tasks