Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Turning large language models into cognitive models

Marcel Binz, Eric Schulz

TL;DR

The study investigates turning large language models into domain-general cognitive representations by finetuning a lightweight readout on LLaMA embeddings derived from psychological data. Using CENTaUR, a linear-readout model, the authors show state-of-the-art fit in two decision-making tasks and demonstrate robustness to individual differences, with cross-task generalization to a hold-out experiential-symbolic task. The embeddings provide a shared representational space enabling multi-task transfer, suggesting a scalable path to data-efficient cognitive modeling. Overall, this work opens new directions for applying pre-trained language models to cognitive psychology and behavioral science by reframing them as generalist cognitive models.

Abstract

Large language models are powerful systems that excel at many tasks, ranging from translation to mathematical reasoning. Yet, at the same time, these models often show unhuman-like characteristics. In the present paper, we address this gap and ask whether large language models can be turned into cognitive models. We find that -- after finetuning them on data from psychological experiments -- these models offer accurate representations of human behavior, even outperforming traditional cognitive models in two decision-making domains. In addition, we show that their representations contain the information necessary to model behavior on the level of individual subjects. Finally, we demonstrate that finetuning on multiple tasks enables large language models to predict human behavior in a previously unseen task. Taken together, these results suggest that large, pre-trained models can be adapted to become generalist cognitive models, thereby opening up new research directions that could transform cognitive psychology and the behavioral sciences as a whole.

Turning large language models into cognitive models

TL;DR

The study investigates turning large language models into domain-general cognitive representations by finetuning a lightweight readout on LLaMA embeddings derived from psychological data. Using CENTaUR, a linear-readout model, the authors show state-of-the-art fit in two decision-making tasks and demonstrate robustness to individual differences, with cross-task generalization to a hold-out experiential-symbolic task. The embeddings provide a shared representational space enabling multi-task transfer, suggesting a scalable path to data-efficient cognitive modeling. Overall, this work opens new directions for applying pre-trained language models to cognitive psychology and behavioral science by reframing them as generalist cognitive models.

Abstract

Large language models are powerful systems that excel at many tasks, ranging from translation to mathematical reasoning. Yet, at the same time, these models often show unhuman-like characteristics. In the present paper, we address this gap and ask whether large language models can be turned into cognitive models. We find that -- after finetuning them on data from psychological experiments -- these models offer accurate representations of human behavior, even outperforming traditional cognitive models in two decision-making domains. In addition, we show that their representations contain the information necessary to model behavior on the level of individual subjects. Finally, we demonstrate that finetuning on multiple tasks enables large language models to predict human behavior in a previously unseen task. Taken together, these results suggest that large, pre-trained models can be adapted to become generalist cognitive models, thereby opening up new research directions that could transform cognitive psychology and the behavioral sciences as a whole.
Paper Structure (12 sections, 4 figures)

This paper contains 12 sections, 4 figures.

Table of Contents

  1. Introduction
  2. Finetuned language models beat domain-specific models
  3. Model simulations reveal human-like behavior
  4. Language model embeddings capture individual differences
  5. Evaluating goodness-of-fit on hold-out tasks
  6. Discussion
  7. Acknowledgements:
  8. Funding:
  9. Data and materials availability:
  10. Fitting procedure:
  11. Model simulations:
  12. Baseline models:

Figures (4)

  • Figure 1: Illustration of our approach and main results. (a) We provided text-based descriptions of psychological experiments to a large language model and extracted the resulting embeddings. We then finetuned a linear layer on top of these embeddings to predict human choices. We refer to the resulting model as CENTaUR. (b) Example prompt for the choices13k data set. (c) Negative log-likelihoods for the choices13k data set. (d) Example prompt for the horizon task. (e) Negative log-likelihoods for the horizon task. Prompts shown in this figure are stylized for readability. Exact prompts can be found in the Supplementary Materials.
  • Figure 2: Model simulations. (a) Performance for different models and human participants on the choices13k data set. (b) Performance for different models and human participants on the horizon task. (c) Human choice curves in the equal information condition of the horizon task. (d) Human choice curves in the unequal information condition of the horizon task. (e) LLaMA choice curves in the equal information condition of the horizon task. (f) LLaMA choice curves in the unequal information condition of the horizon task. (g) CENTaUR choice curves in the equal information condition of the horizon task. (h) CENTaUR choice curves in the unequal information condition of the horizon task.
  • Figure 3: Individual differences. (a) Negative log-likelihood difference to the best-fitting model for each participant. Black highlights the best-fitting model, while white corresponds to a difference larger than ten. (b) Negative log-likelihoods for models that were finetuned using the random-effects structure described in the main text.
  • Figure 4: Hold-out task evaluations. (a) Example prompt for the experiential-symbolic task of garcia2023experiential. (b) Human choice curves as a function of win probabilities for both options. (c) Human indifference points as a function of win probability for the E-option. Indifferent points express the win probabilities at which a decision-maker is equally likely to select both options. (d) LLaMA choice curves as a function of win probabilities for both options. (e) LLaMA indifference points as a function of win probability for the E-option. (f) CENTaUR choice curves as a function of win probabilities for both options. (g) CENTaUR indifference points as a function of win probability for the E-option.