A foundation model to predict and capture human cognition

Data collection

We constructed Psych-101 by transcribing data from 160 psychological experiments into natural language. Each prompt was designed to include the entire trial-by-trial history of a complete session from a single participant. The experiments included were selected using the following criteria: publicly available data on a trial-by-trial level; the possibility of transcription into text without a significant loss of information; and coverage of a broad spectrum of domains. The transcription of each experiment was done manually by the authors. Approval from the institutional review board was obtained by the individual studies as required. We designed our natural-language prompts using the following principles: instructions should follow the original study as closely as possible; simplifications were made where appropriate; and a maximum prompt length of roughly 32,768 tokens was used. Full information about all the experiments included is provided in the Supplementary Information, Example prompts.

Fine-tuning procedure

Llama 3.1 70B was the base model for our fine-tuning procedure. We used a parameter-efficient fine-tuning technique known as QLoRA, which adds so-called low-rank adapters to each layer of a four-bit quantized base model. The base model was kept fixed during fine-tuning and only the parameters of the low-rank adapters were adjusted. We added low-rank adapters of rank r = 8 to all linear layers of the self-attention mechanisms and the feedforward networks. Each low-rank adapter modifies the forward pass as follows:

where XW is the (quantized) linear transformation of the base model and XL1L2 is the low-rank adapter component, with X being the input to the layer with dimensionality h and Y being the output of the layer with dimensionality o. The hyperparameter α controls the trade-off between the two. R is the set of real numbers. Low-rank adapter computations were performed in half-precision floating-point format. For further details on this technique, please see the original work.

We fine-tuned the model for one epoch on the entire dataset using a standard cross-entropy loss (we experimented with prolonged training but found that this led to overfitting). We only back-propagated the loss at human responses and masked out the loss for all other tokens. The effective batch size was set to 32, the learning rate to 0.00005 and the weight decay to 0.01. We used an 8-bit AdamW optimizer with a linearly increasing warm up over the first 100 gradient steps. The fine-tuning procedure was implemented using the unsloth library (https://unsloth.ai/).

We have also trained a smaller version of Centaur, called Minitaur, that uses Llama 3.1 8B as the base model following the same recipe. Minitaur captures human behaviour close to its training distribution but generalizes less robustly than the larger model to out-of-distribution experiments (Extended Data Fig. 7). Nevertheless, we believe that Minitaur is useful for prototyping because it does not require access to any specific hardware (it runs, for instance, on the free GPU instances in Google Colab).

Evaluation metric

We used (negative) log-likelihoods averaged over responses as our evaluation metric. For experiments with multi-token responses, we summed log-likelihoods within a response and averaged across responses. We used one-sided t-tests whenever we tested whether Centaur outperformed a competing model in predicting human behaviour, because our hypotheses were directional and based on the prior expectation that Centaur would perform better. Because the number of observations in our analyses is generally large, reported significant effects survive after correcting for multiple comparisons where appropriate.

Domain-specific cognitive models

We selected as our baseline models 14 cognitive and statistical models that together cover most of the experiments in Psych-101. Further details regarding the included models and their specifications are provided in Supplementary Information, Modelling details.

For our main analysis, we were interested in predicting the behaviour of held-out participants. Therefore, we fitted a joint set of parameters for all participants in the training data and evaluated how well a model with these parameters predicts the responses of held-out participants. Mirroring the evaluation metric for the language-based models, we evaluated goodness-of-fit using (negative) log-likelihoods averaged over responses.

For the out-of-distribution evaluations, we fitted model parameters using the most similar experiment in the training set, and then we evaluated how well a model with the resulting parameters predicts human responses in the unseen setting. The most similar experiment for the magic-carpet version of the two-step task was a two-step task experiment with the default spaceship cover story. The most similar experiment for Maggie’s farm was the horizon task. We included no baseline model for the logical reasoning task, because none of the experiments in the training data were similar to it.

Neural alignment

The neural alignment analysis on the two-step task was conducted using data collected in a previous study. We used a regularized linear regression model to predict fMRI data from internal representations of Centaur and Llama (a separate model was used for each participant and region). We fitted each of these models on data from two scanning blocks and evaluated them on data from the third. The regularization strength was selected using a nested cross-validation procedure. For each run, we split the beta maps into cortical and subcortical regions of interest (ROI) using the Schaefer 2018 atlas with 100 ROIs. We averaged the betas within each ROI, reducing the number of betas from the number of voxels to the number of ROIs. All cortical and subcortical ROIs from the atlas were evaluated. Reported Pearson correlation coefficients correspond to the average across all ROIs.

Internal representations were extracted from the models’ residual stream and transformed using a principal component analysis. We set the number of retained components such that they explained 95% of the variance.

The fMRI data were preprocessed using fMRIPrep 24.0. We used the default settings of fMRIPrep, and all the scans were aligned to the MNI152NLin2009cAsym atlas. To extract effect estimates for each subtrial of the task (such as the second step of the fifth trial, or the feedback of the tenth trial), we built separate general linear models (GLMs). Each GLM included the subtrial of interest as a separate regressor, whose z-scored beta estimates were used for the alignment analysis. This part of the data was not modelled using other regressors. Furthermore, we included different regressors capturing all the first steps, all the second steps and all the feedback steps. Finally, we used six rotation and translation estimates as well as framewise displacement as noise regressors. The haemodynamic response was modelled using the spm model. A high-pass filter of 0.01 Hz and a Gaussian kernel with 6 mm full-width at half-maximum was applied. The GLMs were built using nilearn.

The neural alignment analysis on the sentence-reading task was conducted using publicly available code from the original study. No other changes were made apart from replacing GPT2-XL with Centaur and Llama. Please see the original study for further details.

Model-guided scientific discovery

In our model-guided scientific discovery analysis, we focused on participants in the test set to avoid any potential contamination issues. We fitted parameters of all cognitive models individually for each participant using a maximum-likelihood estimation. Models were compared with each other using the AIC. The three models from the original study were implemented by the following equations:

where xA and xB are vectors containing four expert ratings (either 0 or 1) and β is a free parameter controlling the noise level.

We prompted DeepSeek-R1 (in the Distill-Llama-70B variant) to generate explanations of human decision-making; the corresponding prompt is provided in Supplementary Information, Model-guided scientific discovery. We then formalized the explanation shown in Fig. 5b into the following computational model:

For the scientific regret minimization pipeline, we computed the difference in log-likelihoods between Centaur and the DeepSeek-R1-discovered model. We visualized and inspected the ten data points with the greatest difference. This process resulted in the following computational model:

where σ is a free parameter constrained between 0 and 1 that controls the trade-off between the two strategies.