Value Aggregation

This project looks at how to choose a collective action when people, groups, or moral views value the available outcomes differently.

The repository includes a Python package for aggregation rules and experiments comparing those rules with human judgments, language-model judgments, and choices in existing datasets. The main comparison is between rules such as Nash bargaining and an inequality-efficiency (IE) family that trades off total value against its distribution across groups.

For details, see Intuitions of Compromise: Utilitarianism vs. Contractualism.

@misc{moore2024intuitions,
  title={Intuitions of Compromise: Utilitarianism vs. Contractualism},
  author={Jared Moore and Yejin Choi and Sydney Levine},
  year={2024},
  eprint={2410.05496},
  archivePrefix={arXiv},
  primaryClass={cs.AI},
  url={https://arxiv.org/abs/2410.05496}
}

What is in the repository

src/value_aggregation/ implements game states, normalization utilities, voting and bargaining rules, inequality and efficiency measures, and game generation helpers. src/tests/ contains its unit tests.
experiments/intuition/ contains the main human and LLM compromise-judgment study: scenario generation, the browser task, model runners, statistical analyses, figures, and reproducibility instructions. Its README.md documents every subdirectory and the current paper workflow.
experiments/prevalence/ contains exploratory notebooks that translate the Kaleido, Moral Machine, and NLPositionality datasets into this project's common game representation.
demos.ipynb demonstrates aggregation behavior on small moral-uncertainty examples.
external_data/ is the local destination for downloaded third-party data; it is not versioned. make data documents and downloads the expected sources.

Setup

Python 3.11 or later is required. Create the canonical virtual environment and install the package in editable mode with:

make init
source env-aggregation/bin/activate

The environment lives at env-aggregation/. Run the library test suite with:

make test

The intuition study additionally uses R with brms and cmdstanr for its mixed-effects models and provider credentials for new LLM runs. See the study README for those workflows.

Running notebooks

With Jupyter installed in the active environment:

jupyter notebook

Most intuition-study commands expect the study directory as the working directory:

cd experiments/intuition