A foundation model to predict and capture human cognition

Psych-101 and Centaur both constitute valuable tools for scientific discovery. In the following section, we present an example of how each of them can be used to improve our understanding of human decision-making. The individual steps of this process are illustrated in Fig. 5.

Psych-101 contains human behavioural data in a natural-language format, which means it can be readily processed and analysed by a language-based reasoning model such as DeepSeek-R1. To demonstrate this use case, we asked DeepSeek-R1 to generate an explanation of participants’ behaviour in a multi-attribute decision-making experiment. In this paradigm, participants are given two different options that are each characterized by various features (in our case, four expert ratings for two products) and they must then decide which of the two options they prefer (Fig. 5a). The model produced several explanations, one of which caught our attention: “The participant employed a two-step decision-making strategy. First, they determined which product had the majority of positive ratings (1 s) across all experts. If the products were tied in the number of positive ratings, the participant then considered the rating from the highest validity expert to break the tie.” This strategy combines two well-known heuristic decision-making strategies that, as far as we know, have not been considered in this combination before. We then took this verbal strategy, implemented it as a formal computational model and found that it explained human response behaviour more accurately than the three strategies considered in the original study (a weighted-additive strategy, equal weighting and take-the-best heuristic; Fig. 5b).

However, the DeepSeek-R1-discovered model Akaike information criterion (AIC; 181.7) still fell short of the goodness-of-fit of Centaur (AIC, 72.5), indicating that there is still room for improvement. We therefore used a method known as scientific regret minimization, which uses a black-box predictive model as a reference to identify responses that are in principle predictable but are not captured by a given model. Typically, scientific regret minimization requires the collection of a large-scale experiment-specific dataset to train this predictive model. Centaur, however, can be used out-of-the-box and without the need to collect any domain-specific data, thereby circumventing this step and broadening the scope of scientific regret minimization considerably (indeed, the multi-attribute decision-making data set under consideration contained fewer than 100 participants, placing it far out of reach for conventional scientific regret minimization). When inspecting the responses that were well predicted by Centaur but not by the DeepSeek-R1-discovered model, we observed that they all involved problems in which participants chose the option with fewer positive ratings overall but which was rated positively by a higher-validity expert (see Fig. 5c for an illustration of these problems and Methods, ‘Model-guided scientific discovery’ for further details). This pattern indicates that the switch between the two heuristics is probably not as strict as initially suggested by the DeepSeek-R1-discovered strategy. To capture this, we replaced the either-or rule with a weighted average of both heuristics. We found that the model that resulted from this process matched Centaur in terms of its goodness-of-fit (AIC, 71.7) but was still interpretable. We entered the resulting AIC values of all the models in a group-level model-selection procedure and estimated the protected exceedance probability, which is defined as the probability that a particular model has a higher frequency within a group than all the other candidate models. The protected exceedance probability of the model that resulted from scientific regret minimization was P = 0.83. Notably, the result of this model comparison stands in contrast to the one that was conducted with the original set of models and indicates that people rely on a combination of heuristics when making decisions, as opposed to following a weighted-additive strategy.