Revealing and Mitigating Over-Attention in Knowledge Editing

This repository contains the official implementation of the paper "Revealing and Mitigating Over-Attention in Knowledge Editing". The instructions for setting up and running the experiments described below. We also provide a demo on Colab for a quick understanding of the attention drift phenomenon in knowledge editing. Feel free to open an issue if you encounter any problems.

Requirements

To install the necessary dependencies, run the following command:

pip install -r requirements.txt

Preparation

Before starting, you need to set up the required data and models:

1. Download Status Data: Fetch the dataset from ROME by downloading it to the ./data/status directory from here.

2. Download Attribute Snippets: Download attribute_snippets.json from this link and place it in the ./data/status directory.

3. Set Up Pretrained Models: Download the checkpoints for the pretrained model, for example, meta-llama/Meta-Llama-3-8B. Update the model paths in ./easyeditor/editors/easy_editor.py to reflect where the models are stored on your system.

Editing and Evaluation

To perform knowledge editing using our ADR method on the counterfact dataset, use the following commands: The hyparams of all experiments are illustrated in out Appendix.

cd ./examples
bash scripts/rome.sh 

To get the metrics in our paper, run following instructions:

python summrize.py \
--metric accuracy \
--mode scores \
--base_path ./results/gpt-j-6B

The results will be shown in results.md.

Results

Our model achieves impressive performance metrics, as shown in the figure below:

Citation

@article{wangrevealing,
  title={Revealing and Mitigating Over-Attention in Knowledge Editing},
  author={Wang, Pinzheng and Tang, Zecheng and Zhou, Keyan and Li, Juntao and Zhu, Qiaoming and Zhang, Min},
  journal={arXiv preprint arXiv:2502.14838},
  year={2025}
}