Exploring Theory of Mind in Large Language Models through Multimodal Negotiation

With the advancement of Large Language Models (LLMs), they are increasingly being used as a backend for interactive virtual agents and assistants. Therefore, a critical social skill for these agents is Theory of Mind (ToM): the ability to model and reason about other agents. Research has investigated ToM in LLMs using standard, modified, and extended versions of false-belief tasks. These tests include explicit prompts asking LLMs to answer questions about other agents. However, in real situations, people have to use ToM unprompted to navigate social life. Additionally, oftentimes, people have to rely on nonverbal cues such as facial expressions. This work seeks to address this gap by studying implicit ToM in LLMs in a negotiation task. In negotiation, agents have to implicitly reason about other agents to reach an agreed-upon best possible deal. We conducted the negotiation experiment by prompting different LLMs to roleplay as characters and pitting them against rule-based agents that may respond with different facial expressions. We measure and compare the outcomes of the negotiation across models. Our results show that strong LLMs like GPT-4 turbo and Claude 3 Opus can perform decently and adjust their offers based on access to facial expression information, but weaker models are far behind. Our work contributes to our understanding of LLMs’ capabilities and limitations for serving as intelligent and interactive agents.