Learning to Decide with AI-Assistance under Human-Alignment

This repository contains the code and data for the paper "Learning to Decide with AI-Assistance under Human-Alignment".

Dependencies

Python

All experiments were performed using Python 3.13.5 To create a virtual environment and install dependencies run:

python3 -m venv env
source env/bin/activate
pip install -r requirements.txt

Repository structure

├── configs/
│   └── human_ai_interactions.json
│   └── human_alignment.json
├── data/
│   ├── human_ai_interactions_data/
│   └── human_alignment_data/
├── figures/
├── notebooks/
├── outputs/
├── scripts/
│   ├── human_ai_interactions.sh
│   └── human_alignment.sh
└── src/
    ├── bandits.py
    ├── experiment.py
    ├── plot.py
    └── utils.py

• configs/ — JSON files specifying experiment parameters
• data/ — Raw and processed data for each of the datasets
• figures/ — Generated figures
• notebooks/ — Jupyter notebooks for data preprocessing
• outputs/ — Experiment results (JSON files)
• scripts/ — Shell scripts to run the experiment for each dataset
• src/ — Source code:
  • bandits.py — Deploys our online learning algorithm, the contextual vanilla baseline and the optimal policy
  • experiment.py — Main experiment loop
  • utils.py — Utility functions to find the optimal decision policy
  • plot.py — Generates plots and legends

Running experiments

To run the online learning algorithms and the optimal decision policy for an entire dataset run:

./scripts/<dataset_name>.sh 

where <dataset_name> is one of human_ai_interactions and human_alignment.

Results and plots

Results are saved to the output_dir specified in the config (default: outputs/<dataset_name>). If the DEBUG flag in experiment.py is set to True logs are saved to log_dir (default: outputs/<dataset_name>/logs) otherwise no logs are printed. The optimal decisions are saved under optimal_path_actions (default: outputs/<dataset_name>/optimal_policies/actions) specified in the respective config file. To generate the plots used in the paper run:

python3 -m src.plot

The plots are saved under figures/<dataset_name>/<num_iterations> by default, where <num_iterations> is specified in each config file by num_iterations.

Citation

If you use parts of the code in this repository for your own research, please consider citing:

@article{benz2026learning,
  title={Learning to Decide with AI Assistance under Human-Alignment},
  author={Benz, Nina Corvelo and Straitouri, Eleni and Gomez-Rodriguez, Manuel},
  journal={arXiv preprint arXiv:2605.12646},
  year={2026}
}