MetaMind: Modeling Human Social Thoughts with Metacognitive Multi-Agent Systems
Xuanming Zhang, Yuxuan Chen, Min-Hsuan Yeh, Yixuan Li
2025-05-28
Summary
This paper talks about MetaMind, a new system that helps AI models better understand and predict what people are thinking and feeling in social situations, kind of like how humans do.
What's the problem?
The problem is that AI models usually have a hard time figuring out what others might believe, want, or intend, which is an important skill called Theory of Mind. Without this skill, AI can't interact with people in a truly human-like way.
What's the solution?
The researchers built a multi-agent system that works a bit like the human mind, breaking down social understanding into steps like coming up with ideas about what someone might think, improving those ideas, and then deciding how to respond. This approach helps the AI perform much more like a real person in social reasoning tasks.
Why it matters?
This matters because it brings AI closer to understanding people on a deeper level, which can make technology more helpful and natural in things like virtual assistants, education, and mental health support.
Abstract
MetaMind, a multi-agent framework inspired by metacognition, enhances LLMs' ability to perform Theory of Mind tasks by decomposing social understanding into hypothesis generation, refinement, and response generation, achieving human-like performance.