Overview

• Centaur combines a large language model with Psych-101’s ten million decision data points to simulate human choices across diverse experimental tasks.
• In benchmarks on 32 held-out tasks, it outperformed 14 established cognitive and statistical models and predicted reaction times with surprising precision.
• The peer-reviewed study was published July 2 in Nature under DOI 10.1038/s41586-025-09215-4.
• Researchers are now dissecting Centaur’s internal algorithms to uncover the computational patterns underlying decision strategies.
• Plans are underway to enrich the dataset with demographic and psychological variables and to pilot in silico simulations for clinical decision-making under an ethical framework.