Large language models have shown remarkable success in many tasks, but they still face challenges in complex reasoning. One area of specific interest is theory-of-mind (ToM) reasoning, which involves understanding agents beliefs, goals, and mental states. This study measures the ToM performance of GPT-4 and three GPT-3.5 variants, and investigates the effectiveness of in-context learning in improving their ToM comprehension. The study found that appropriate prompting enhances LLM ToM reasoning and underscores the context-dependent nature of LLM cognitive capacities. The capacity of LLMs to reliably perform ToM reasoning is important for several reasons, including social understanding and inferential reasoning.
