@inproceedings{tjuatja25acl,
    title = {BehaviorBox: Automated Discovery of Fine-Grained Performance Differences Between Language Models},
    author = {Lindia Tjuatja and Graham Neubig},
    booktitle = {Annual Meeting of the Association for Computational Linguistics (ACL)},
    address = {Vienna, Austria},
    month = {July},
    url = {https://arxiv.org/abs/2506.02204},
    year = {2025}
}