Follow us on Google News (click on ☆)
This system, trained on over 10 million human decisions, opens major new perspectives in psychology and neuroscience. Unlike traditional approaches, it doesn't just analyze isolated behaviors but generalizes its predictions to new situations.
Centaur: a step toward universal modeling of cognition
The model is based on a database called Psych-101. These tests cover varied domains like risk-taking or moral reasoning. Centaur excels where classical theories fail: it predicts not only choices but also reaction times.
Researchers adapted an existing language model (Llama 3.1) by modifying just 0.15% of its parameters. This efficiency allows Centaur to function as a "virtual laboratory." According to its creators, it could help study mental disorders by simulating decision-making mechanisms specific to each pathology.
One of its strengths lies in its ability to adapt to novel scenarios. For example, it transfers strategies learned in a space treasure game to a flying carpet quest—a flexibility surpassing specialized models developed over decades.
Applications and limits of a psychologist AI
Centaur could transform clinical research or behavioral economics. By reproducing thought patterns, it would allow testing therapies or public policies in silico. However, its data comes mostly from educated Western populations, limiting its scope.
The team emphasizes the need for transparency. The model is open-source to avoid biases and ensure ethical use. Next step: understanding how its internal calculations reflect brain activity and refining its analysis of individual differences.
For scientists, Centaur is just a beginning. Enriched with more diverse data, it could become a "unified theory" of cognition. But its development will need to balance innovation with ethical vigilance.