[ACL 2025] DELMAN: Dynamic Defense Against Large Language Model Jailbreaking with Model Editing

Yi Wang · Fenghua Weng · Sibei Yang · Zhan Qin · Minlie Huang · Wenjie Wang^†

News

• (2026.05) 🎉 Congratulations! Our follow-up work "EVA: Editing for Versatile Alignment against Jailbreaks" has been accepted by IEEE TPAMI 2026 !

Overview

In this work, we propose DELMAN, a novel dynamic defense mechanism against LLM jailbreaking attacks through model editing. Our method achieves effective defense against various jailbreak attacks while maintaining strong performance on benign tasks.

Quickstart

Installation

To install all dependencies, please get into this directory and run the following command:

git clone https://github.com/wanglne/DELMAN.git
cd DELMAN
conda create -n delman python=3.9.20
conda activate delman
pip install -r requirements.txt

We directly provide the "cov" matrix of models that we have already calculated. https://drive.google.com/drive/folders/1uee2b_rti0UlNgQ52hlY2oVB5AduO7Ch?usp=sharing After decompressing it and saving it to the ./data/stats folder. Note that the provided cov matrices were computed on our devices, and differences in hardware or environments may lead to inconsistent results. If you encounter such issues, we recommend recomputing the cov matrices locally on your own device.

Llama 3.1 Configuration

If you are using Llama 3.1 with DELMAN, you need to adjust the offset value in ./rome/repr_tools.py line 106:

# NOTE: For Llama 3.1, set offset to 2
# For other models, use 1
offset = 1

Change offset = 1 to offset = 2 when using Llama 3.1.

Run DELMAN

export model_name=Qwen/Qwen2.5-7B-Instruct
export hparams_fname=Qwen2.5-7B-Instruct.json
export data_name=HarmBench.json
export out_name="DELMAN_qwen2.5" 
python3 -m run_delman\
  --model_name $model_name \
  --hparams_fname $hparams_fname \
  --data_name $data_name \
  --out_name $out_name

export model_name=meta-llama/Llama-3.1-8B-Instruct
export hparams_fname=Llama-3.1-8B-Instruct.json
export data_name=HarmBench.json
export out_name="DELMAN_llama3.1" 
python3 -m run_delman\
  --model_name $model_name \
  --hparams_fname $hparams_fname \
  --data_name $data_name \
  --out_name $out_name

Acknowledgement

Our code is based on MEMIT and BadEdit.

Citation

@inproceedings{wang2025delman,
  title={DELMAN: Dynamic defense against large language model jailbreaking with model editing},
  author={Wang, Yi and Weng, Fenghua and Yang, Sibei and Qin, Zhan and Huang, Minlie and Wang, Wenjie},
  booktitle={Findings of the Association for Computational Linguistics: ACL 2025},
  pages={11465--11481},
  year={2025},
  doi={10.18653/v1/2025.findings-acl.598},
  url={https://aclanthology.org/2025.findings-acl.598/}
}