A foundation model to predict and capture human cognition

The human mind is remarkably general. Not only do we routinely make mundane decisions, such as choosing a breakfast cereal or selecting an outfit, but we also tackle complex challenges, such as figuring out how to cure cancer or explore outer space. We learn skills from only a few demonstrations, reason causally and fuel our actions through curiosity. Whether we are climbing mountains, playing video games, or creating captivating art, our versatility defines what it means to be human.

By contrast, most contemporary computational models, whether in machine learning or the cognitive sciences, are domain specific. They are designed to excel at one particular problem and only that problem. Consider, for instance, AlphaGo, which is a computer system created by Google DeepMind to master the game of Go. The system can play this particular game at an impressive level, but it cannot do much beyond that. A similar pattern can be observed in the cognitive sciences. For instance, prospect theory, which is one of the most influential accounts of human cognition, offers valuable insights into how people make choices, but it tells us nothing about how we learn, plan or explore.

If we want to understand the human mind in its entirety, we must move from domain-specific theories to an integrated one. The importance of such a unified approach has already been recognized by the pioneers of our field. For example, in 1990, it was stated that “unified theories of cognition are the only way to bring [our] wonderful, increasing fund of knowledge under intellectual control”. How can we make meaningful progress towards such theories?

An important step towards a unified theory of cognition is to build a computational model that can predict and simulate human behaviour in any domain. In this paper, we take up this challenge and introduce Centaur—a foundation model of human cognition. Centaur was designed in a data-driven manner by fine-tuning a state-of-the-art large language model on a large corpus of human behaviour. For this purpose, we curated a large-scale dataset called Psych-101, which covers trial-by-trial data from 160 psychological experiments (see Methods, ‘Data collection’ and Extended Data Fig. 1). We transcribed each of these experiments into natural language, which provides a common format for expressing vastly different experimental paradigms. The resulting dataset has an unprecedented scale, containing more than 10,000,000 human choices and including many canonical studies from domains such as multi-armed bandits, decision-making, memory, supervised learning, Markov decision processes and more (see Fig. 1a for an overview and examples).

We subjected Centaur to a series of rigorous tests and demonstrate that it captures human behaviour at several levels of generalization. First, we show that Centaur predicts behaviour of held-out participants (those who are not part of the training data) better than existing cognitive models in almost every single experiment. We then demonstrate that its ability to capture human behaviour also generalizes to held-out experiments. In this context, we find that Centaur accurately predicts human behaviour under modified cover stories, problem structures and even in entirely new domains. Finally, we show that Centaur’s internal representations become more human aligned, even though it was never explicitly trained to capture human neural activity.

Taken together, our results demonstrate that it is possible to discover computational models that capture human behaviour across a wide range of domains. We think that such a predictive model offers many direct opportunities to obtain a better understanding of the human mind and we present a case study that demonstrates this potential.