Mind Over Machine: Evaluating Theory of Mind Reasoning in LLMs and Humans

As Large Language Models (LLMs) become increasingly integrated into applications demanding human-like understanding-such as mental health support, education, and social robotics-their capacity to exhibit Theory of Mind (ToM) reasoning is essential. Although previous research has evaluated LLMs' capability in ToM tasks, a critical gap remains as few studies have systematically investigated how LLMs' ToM reasoning diverges from human reasoning and the extent of these differences. This study introduces a reinforcement learning-based framework designed to bridge this gap. This approach seeks to enhance LLM alignment with human ToM reasoning, effectively narrowing the differences in their reasoning processes. Finally, future research directions to advance this field will be discussed, including strategies for developing LLMs that can better approximate human social cognition. This work lays a foundation for responsible LLM deployment, offering guidelines for applications in sensitive contexts where accurate ToM understanding is crucial.