AdaSteer: Your Aligned LLM is Inherently an Adaptive Jailbreak Defender

The official implementation for the EMNLP 2025 paper
"AdaSteer: Your Aligned LLM is Inherently an Adaptive Jailbreak Defender"

📄 Paper • 🚀 Code

---

🔧 Requirements

• Python 3.9.0
• PyTorch 2.4.1
• Transformers 4.46.3

Installation

git clone https://github.com/MuyuenLP/AdaSteer
cd AdaSteer
pip install -e .

🚀 Quick Start

1. Vector Extraction

We employ a difference-in-means method for vector extraction. Here's an example of extracting the rejection vector for LLaMA-3.1-8B-Instruct:

bash scripts/llama31/RD.sh

2. Activation Steering

Fixed Steering Coefficient

Apply rejection vectors with a fixed steering coefficient λ to reject harmful inputs:

bash scripts/llama31/refusal.sh

Adaptive Steering (AdaSteer)

For dynamic steering, we modify the transformers library to implement adaptive activation steering coefficient calculation. See ./adasteer/models/For_Steering_LlamaModel_adasteer.py for implementation details.

bash scripts/llama31/adasteer.sh

📚 Citation

If you find our work useful for your research, please kindly cite our paper:

@article{zhao2025adasteer,
  title={AdaSteer: Your Aligned LLM is Inherently an Adaptive Jailbreak Defender},
  author={Zhao, Weixiang and Guo, Jiahe and Hu, Yulin and Deng, Yang and Zhang, An and Sui, Xingyu and Han, Xinyang and Zhao, Yanyan and Qin, Bing and Chua, Tat-Seng and others},
  journal={arXiv preprint arXiv:2504.09466},
  year={2025}
}