johnathansun/llm-cognitive-choice

Do LLMs Choose Like Humans?

Data for Do LLMs Choose Like Humans? Using Cognitive Theory to Evaluate LLM Decision-Making, by Johnathan Sun, Andrei Shleifer, and Yonatan Belinkov.

The dataset contains the product choice trials, model responses, stimuli, and ratings used in the paper. The files follow the layout expected by the analysis code.

The download is about 3.42 GB and includes 2,240,800 recorded model responses across 51 files.

Browse the files in Data Studio. Use the subset menu to choose a trial file, a model's responses, a stimulus file, or a ratings table. Each subset corresponds to one data file.

Download and reproduce

From the root of the code repository, download the data:

python -m pip install --upgrade huggingface_hub
hf download johnathansun/llm-cognitive-choice \
  --repo-type dataset --local-dir . --include "data/*"

The analyses were run with Python 3.10.12 and the package versions recorded in the code repository. Install those versions and run the analyses:

python -m pip install -r requirements-analysis.txt -c requirements-analysis-constraints.txt
python analysis/fit_regressions.py
jupyter execute analysis/figures_and_claims.ipynb
jupyter execute analysis/appendix_figures.ipynb

The regression script saves its results in cache/, and the notebooks write figures to figs/. The data are stored as JSON and CSV files; the download command preserves the folder structure used by the analyses.

Files

Path	Contents
data/stimuli/	Seed products, five-tier product ladders, cue definitions, and quality variants.
data/trials_baseline.json	8,800 baseline trials.
data/trials_price.json	88,000 price-sweep trials.
data/trials_quality.json	44,000 quality-sweep trials.
data/trials_icl_mix.json	36,000 in-context learning mixture trials.
data/trials_icl_priming.json	35,200 priming trials for the generation pipeline.
data/responses/<model>/<family>.json	51 response files: 16 baseline, 14 price, 14 quality, and 7 ICL mixture configurations.
data/quality_ratings_gpt54.csv	2,500 GPT-5.4 quality-rating ensembles.
data/quality_ratings_claude46.csv	2,500 Claude Sonnet 4.6 quality-rating ensembles.
data/hedonic_scores.csv	Hedonic scores for 100 product categories.
data/manifest.json	File inventory, row counts, and SHA-256 checksums.
viewer/*.parquet	Copies of the tables for browsing in the dataset viewer.

The baseline, price, and quality experiments contain 140,800 trials in total, with responses from multiple model configurations for each trial. The dataset also includes the in-context learning experiments.

Citation

@misc{sun2026llmschooselikehumans,
  title={Do LLMs Choose Like Humans? Using Cognitive Theory to Evaluate LLM Decision-Making},
  author={Johnathan Sun and Andrei Shleifer and Yonatan Belinkov},
  year={2026},
  eprint={2609.22225},
  archivePrefix={arXiv},
  primaryClass={cs.CL},
  url={https://arxiv.org/abs/2609.22225}
}