Centaur: A Foundation Model of Human Cognition

In this paper we have introduced Centaur, a foundation model of human cognition that was obtained by fine-tuning a state-of-the-art language model on Psych-101, which is a large-scale dataset of human behavior. This approach allowed us to leverage the vast knowledge embedded in large language models and also align them with human behavior. Centaur successfully captured human behavior and passed a wide range of out-of-distribution checks. It generalized not only to unseen participants, but also to different cover stories, structural variations and entirely new domains. In addition to analyzing the model on a behavioral level, we also conducted a series of analyses on its internal representations, in which we found increased alignment with human neural activity.

We also conducted a case study demonstrating how both Psych-101 and Centaur can be used for guiding the development of predictive, yet interpretable, cognitive models. The individual steps of our procedure are generic, so it could serve as a blueprint for model-guided scientific discovery in other experimental paradigms in the future. Looking beyond this example, Centaur finds many more applications in the context of automated cognitive science. It may, for instance, be used for in silico prototyping of experimental studies. In this context, one could use the model to figure out which designs lead to the largest effect sizes, how to design a study to reduce the number of required participants or to estimate the power of an effect.

The present paper takes initial steps in leveraging Centaur to gain deeper insights into human cognition, and it also opens up exciting new avenues for future exploration. First, one could further probe Centaur’s internal representations to understand how it represents knowledge and processes information. The resulting insights could, in turn, be used to generate hypotheses about knowledge representation and information processing in humans that could be validated in future experimental studies. We believe that tools such as sparse auto-encoders and attention map visualization provide promising avenues toward accomplishing this goal, and we hope to explore them in future studies.

Furthermore, it might also be possible to train models with different architectures from scratch using the dataset that we created in the process of this paper. Doing so would enable us to investigate the neural architecture of human cognition at a scale that could not have been done before. We might, for example, ask questions such as whether human information processing is better described by attention-based architectures or by architectures with a vector-based memory, or how much we can improve by incorporating theories from the neuroscience literature. We expect an eventual outcome of such an approach to contain both domain-specific and domain-general modules, thereby allowing us to investigate the interplay between the two.

As far as we know, Psych-101 is already the broadest and largest dataset of human behavior available, and we view its development as an ongoing process and plan to develop it further. The focus in its current state is largely on learning and decision-making, but we intend to eventually include more domains, such as psycholinguistics, social psychology and economic games. Experiments with information about individual differences are another source of neglected data in the current iteration of Psych-101. Ideally, we want to include all types of relevant information about subjects (including age, personality traits or socioeconomic status) in the prompt, such that a model trained on these data can capture individual differences. Experiments from developmental psychology or computational psychiatry provide an ideal source for this purpose. Finally, although we have already included some cross-cultural and meta-studies, the current iteration still has a strong bias toward a Western, educated, industrialized, rich and democratic (WEIRD) population.

Eventually, we hope to provide any psychological data in a standardized format that facilitates benchmarking, thereby complementing existing efforts from the neuroscience community. Although the natural-language format (together with quite a bit of reverse-engineering) used in this work allows us to express a vast range of experimental paradigms, it introduces a selection bias against experiments that cannot be expressed in natural language. The long-term objective should therefore be to move toward a multimodal data format.