Initial code release for the paper:

Surgical, Cheap, and Flexible: Mitigating False Refusal in Language Models via Single Vector Ablation (ICLR 2025)

Xinpeng Wang, Chengzhi Hu, Paul Röttger and Barbara Plank.

This code is build on top of the code from the great work Refusal in Language Models Is Mediated by a Single Direction.

source setup.sh

Install the evaluation harness from source

cd lm-evaluation-harness
pip install -e .

To run vector extraction, ablation and evaluation, run the script bellow:

python -m pipeline.run_pipeline --config_path configs/cfg.yaml

We also provide a demo notebook here. We recommend using this as a hands-on intro of how our pipeline works and how the model is changed when doing the (fine-grained) vector ablation.

@inproceedings{wang2025surgical,
    title={Surgical, Cheap, and Flexible: Mitigating False Refusal in Language Models via Single Vector Ablation},
    author={Xinpeng Wang and Chengzhi Hu and Paul R{\"o}ttger and Barbara Plank},
    booktitle={The Thirteenth International Conference on Learning Representations},
    year={2025},
    url={https://openreview.net/forum?id=SCBn8MCLwc}
}