Multicalibration Yields Better Matchings

Code accompanying the paper Multicalibration Yields Better Matchings, accepted at ICML 2026.

Quickstart

We recommend using uv for managing the Python environment and dependencies.

# Run an experiment from a registered config
uv run main.py --cfg sweep/top_k
uv run main.py --cfg sweep/complete_graph

# Rerun even if results already exist
uv run main.py --cfg sweep/top_k --redo

# Run from a config file path
uv run main.py --cfg path/to/config.py

Configs live under wmcal/configs/. Registry keys (e.g. sweep/top_k) are resolved via wmcal.configs.CONFIG_REGISTRY; any .py file with an experiments list and optional WORKERS value can also be passed directly.

License

This project is MIT-licensed. See LICENSE.

Citation

If you use this code in your research, please cite:

@inproceedings{colinibaldeschi2026multicalibration,
  title     = {Multicalibration Yields Better Matchings},
  author    = {Colini Baldeschi, Riccardo and Di Gregorio, Simone and
               Fioravanti, Simone and Fusco, Federico and Guy, Ido and
               Haimovich, Daniel and Leonardi, Stefano and Linder, Fridolin and
               Perini, Lorenzo and Russo, Matteo and Sirin, Cem and Tax, Niek},
  booktitle = {International Conference on Machine Learning},
  year      = {2026},
  url       = {https://openreview.net/forum?id=G1bzbJtcHD},
}

This entry will be updated with the official PMLR citation once the proceedings are published.