Centaur: a foundation model of human cognition

https://arxiv.org/abs/2410.20268?utm_source=tldrai This article introduces Centaur, a computational model capable of predicting and simulating human behaviour in any experiment expressible in natur

Centaur is a groundbreaking computational model designed to predict and simulate human behavior across a wide range of psychological experiments. It was developed by fine-tuning a state-of-the-art language model on Psych-101, a newly created, large-scale dataset. **Key Highlights:** * **Psych-101 Dataset:** This extensive dataset comprises data from over 60,000 participants, documenting more than 10,000,000 choices across 160 distinct experiments. These experiments cover diverse areas, including bandit tasks, decision-making, memory, and reinforcement learning. The dataset is openly accessible, with plans for continuous integration of more domains and individual differences data. * **Superior Prediction Capability:** Centaur significantly outperforms existing cognitive models in predicting the behavior of unseen participants. This capability extends even to experiments with varied cover stories, altered task structures, or entirely new domains, demonstrating strong generalization. * **Neural Alignment:** The model's internal representations show a closer alignment with human neural activity after its fine-tuning process, suggesting a deeper understanding of cognitive processes. * **Potential for Unified Cognition Model:** The authors propose Centaur as a strong candidate for a unified model of human cognition. This paves the way for substantial advancements in cognitive science. * **Direct Applications:** Centaur enables practical applications such as *in silico* prototyping of experimental studies (simulating experiments computationally before physical execution) and the automation of cognitive science research. We believe Centaur marks a significant step forward in understanding and modeling human behavior, offering both theoretical insights and practical tools for cognitive scientists.

By Romain Peter