CPSY 1950 — Computational Models of Human Behavior: Demo Notebook

Full project specs

https://thomas-serre.com/cpsy1950/final-project

This repository contains a demo notebook for CPSY 1950 at Brown University, illustrating how large language models can be evaluated as models of human decision making using the two-step task: a classic paradigm from computational psychiatry.

Overview

The notebook covers two experiments:

1. Model Alignment with Human Decisions: Given a real human participant's trial by trial transcript from the two step task, we query a frontier LLM at each decision point and measure how surprised the model is by the human's actual choice using negative log-likelihood (NLL).
2. Model Behavior in an Open Environment: We let the model play the two-step task autonomously against a simulated environment with drifting reward probabilities, and analyze its behavior using the classic stay/switch analysis to probe for model based vs model free reasoning. Both experiments use the magic carpet cover story version of the task, in which the abstract decision structure is embedded in a fantasy narrative to test generalization.

Setup

1. Clone the repository:

git clone https://github.com/TekinGunasar/CPSY-1950-Project-Demo
cd CPSY-1950-Project-Demo

2. Create and activate a conda environment:

conda create -n cpsy1950 python=3.11
conda activate cpsy1950

3. Install dependencies:

pip install -r requirements.txt

4. Create a .env file in the root directory with your API key (Or just copy and paste the API key into your notebook while making sure it stays on your machine):

echo OPENAI_API_KEY=your_api_key_here > .env

5. After installing, you usually also want to register the environment as a Jupyter kernel:

python -m ipykernel install --user --name cpsy1950

6. Open the notebook by entering the following into your terminal (make sure to select cpsy1950 as your kernel):

jupyter lab

Connecting to the API

In order for your API calls to work, you must be connected to the Brown University VPN: https://it.brown.edu/services/virtual-private-network-vpn

References

• Binz et al. (2025). Centaur: a foundation model of human cognition
• Feher da Silva et al. (2023). Rethinking model-based and model-free influences on mental effort and striatal prediction errors. Nature Human Behaviour, 7(6), 956–969.