DebaToM: An Innovative Debate Framework for Theory of Mind Reasoning Using Large Language Models

Theory of Mind (ToM) is of great significance in human cognition and social reasoning. With the development of Large Language Models (LLMs), exploring the ToM capabilities of these models has become a research hotspot. However, existing methods have limitations such as single perspective reasoning and a lack of iterative interaction when dealing with complex ToM scenarios. In this paper, we propose an innovative debate framework, DebaToM, for ToM reasoning. Our framework enables multiple LLMs to participate in debates and negotiations, achieving comprehensive reasoning through iterative interactions and simulating the multi-round cognitive process of humans. Our DebaToM framework consists of a generator and a discriminator based on LLMs. The two continuously optimize the reasoning process through multiple rounds of interactions. Experimental results on two widely-used ToM reasoning datasets show that our DebaToM framework significantly outperforms multiple baseline models when dealing with ToM reasoning tasks, demonstrating a strong performance advantage. For example, when using LLaMA as the base model, the average improvement across the two widely used ToM reasoning datasets is 6.15%. Further analyses demonstrate the advantages of our DebaToM in multi perspective reasoning and simulating the human reflection and interaction process.