Causal Preference Elicitation (CaPE)

Official code release for **Causal Preference Elicitation (CaPE)** Edwin V. Bonilla, He Zhao, Daniel M. Steinberg (ICML 2026)

Implementation and maintenance of the CaPE codebase by Edwin V. Bonilla.

• Pre-print: https://arxiv.org/abs/2602.01483
• Repository: https://github.com/csiro-funml/CaPE

One-command reproduction

From a fresh clone (assuming python points to python3.14):

python -m venv .venv
source .venv/bin/activate

pip install -r requirements.txt
./reproduce_paper.sh

This runs the paper experiments and regenerates figures using the plotting pipeline.

Repository layout

• dag/, feedback/, generation/, inference/, likelihood/, metrics/, expts/ – core library modules
• experiments/ – paper experiment entrypoints
• scripts/ – helper scripts (cluster submission, plotting, misc)
• configs/ – reference hyperparameters (YAML)
• docs/ – dataset setup notes
• results/, figures/, data/ – generated artifacts (gitignored)

Key entrypoints (paper)

• Sachs (HITL): python experiments/run_sachs_hitl.py
• Synthetic (cluster): bash scripts/cluster/submit_synthetic_cluster.sh
• CausalBench (cluster): bash scripts/cluster/submit_causalbench_cluster.sh
• Plots from saved results: bash scripts/plots/generate_all_plots.sh

Existing results

Results ready to be parsed by scripts/generate_paper_figures.sh and regenerate the figures in the submission can be found in results_final.zip

Cluster Experiments

Scripts for launching the synthetic and causalbench experiments in the cluster are at scripts/cluster/submit_causalbench_cluster.sh and scripts/cluster/submit_synthetic_cluster.sh.

Citation

If you use this code, please cite:

@inproceedings{bonilla2026cape,
  title={Causal Preference Elicitation},
  author={Bonilla, Edwin V. and Zhao, He and Steinberg, Daniel M.},
  booktitle={International Conference on Machine Learning},
  year={2026},
  doi={10.48550/arXiv.2602.01483}
}

Citation metadata is also available via CITATION.cff.

License

Licensed under the Apache License 2.0.