Harvard and MIT Introduce MatrAIx to Simulate Global Human Behavior with 8.3 Billion AI Agents
Researchers from Harvard, MIT, OpenAI, and Google DeepMind have introduced MatrAIx, a system that simulates global human behavior using 8.3 billion AI agents across 1,290 dimensions. The team has also open-sourced a core dataset of 1 million filtered personas on Hugging Face and GitHub. This project scales agent-based social simulation to the entire global population level, providing a powerful infrastructure for evaluating digital products and AI systems before real-world deployment. It enables researchers to predict complex societal behaviors and test software usability across diverse demographics. The AI agents achieved an overall behavioral consistency rate of 91.5% across 400 test scenarios, with performance peaking at 92%-96% in survey and chat environments but dropping to 83% in app navigation tasks. The system consists of three main components: the Persona 8B dataset, the MatrAIx Playground, and a library of 1,010 versioned tasks.
## BACKGROUND
Agent-based modeling (ABM) is a computational method used to simulate the actions and interactions of autonomous agents to study complex system behaviors and emergent phenomena. In machine learning, a coreset is a highly representative subset of a larger dataset, selected to maintain data quality and training efficiency while reducing computational costs.