Science

Andi's Writeup

Researchers have developed Centaur, a computational model that can predict and simulate human behavior across a wide range of psychological experiments[^1]. Built by fine-tuning Meta's Llama 3.1 70B language model on a dataset called Psych-101, Centaur was trained on over 10 million choices made by 60,000 participants across 160 psychology experiments[^1].

The model outperforms existing cognitive models in predicting human behavior, even generalizing to entirely new scenarios it wasn't trained on[^1]. "You can basically run experimental sessions in silico instead of running them on actual human participants," said Marcel Binz, cognitive scientist at the Helmholtz Institute for Human-Centered AI[^2].

Centaur demonstrates unprecedented capabilities in capturing human cognition:

• Predicts behavior with 64% accuracy across varied tasks[^3]
• Generalizes to modified experimental scenarios, like switching from "spaceships" to "magic carpets" in decision-making tasks[^4]
• Shows alignment between its internal representations and human neural activity[^1]
• Performs well on out-of-distribution tasks in moral decision-making, economic games, and logical reasoning[^1]

"It's the first model that can do any kind of task exactly like a human can," said Russ Poldrack, cognitive scientist at Stanford University[^4].

[^1]: Nature - A foundation model to predict and capture human cognition

[^2]: Nature - This AI 'thinks' like a human — after training on 160 psychology studies

[^3]: Live Science - New AI is better at predicting how we behave than ever before, scientists say

[^4]: Gigazine - A basic model 'Centaur' that predicts human responses in psychological experiments has appeared