A team led by Harvard University and the Massachusetts Institute of Technology, with participation from organizations including OpenAI and Google DeepMind, launched the MatrAIx system this month.
The system can generate 8.3 billion AI agents to simulate human behavior worldwide.
The project uses 1,290 dimensions to model people from different backgrounds and lifestyles. These AI agents can fill out surveys, chat with customer service, browse the web, operate apps, and more like real people, with test consistency reaching 91.5%.
The team validated the system through a controlled study involving 400 tests, covering 10 behavioral attributes across all four environments. In 366 of the 400 tests, the agents displayed the correct traits or correctly suppressed those traits when irrelevant, resulting in correct performance in 91.5% of cases.
This figure varied by environment: consistency rose to between 92% and 96% in survey and chat scenarios, while accuracy fell to 83% in app operation tests.
The deterministic, quality-filtered million-persona core set (coreset) is now publicly available on Hugging Face for research use.
Open-source link:
https://github.com/MatrAIx-ai/MatrAIx-Persona-8B