📄 Paper
MAGIC:
- Authors: Xiaoyu Wen, Zhida He, Han Qi, Ziyu Wan, Ying Wen, Tianhang Zheng, Xingcheng Xu, Chaochao Lu, Qiaosheng Zhang.
- Paper: https://arxiv.org/pdf/2602.01539
- Code & Models: https://huggingface.co/XiaoyuWen/MAGIC-Llama3.1-8B-Instruct
This repository provides the official implementation and model checkpoints described in the paper.
🧠 MAGIC Framework Overview
MAGIC is a co-evolving attacker–defender adversarial game framework designed to improve the robustness and safety of large language models.
Instead of relying on static red-teaming or fixed safety datasets, MAGIC formulates LLM safety alignment as a dynamic game between:
- an attacker, which continuously generates increasingly sophisticated harmful or policy-violating prompts, and
- a defender, which adapts through iterative training to resist these attacks while preserving helpfulness.
Through this co-evolutionary process, both sides improve over time, enabling the defender model to generalize to unseen and adaptive attacks.
This model, MAGIC-Llama3.1-8B-Instruct, is the defender model trained under the MAGIC framework based on Llama3.1-8B-Instruct, demonstrating significantly improved robustness against jailbreak and attack prompts.
🤗 Usage
from transformers import AutoModelForCausalLM, AutoTokenizer
model_name = "XiaoyuWen/MAGIC-Llama3.1-8B-Instruct"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(model_name, device_map="auto")
prompt = "Explain why jailbreaking LLMs is dangerous."
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=8192)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))
📚 Citation
If you find this work useful, please cite:
@article{wen2026magic,
title={MAGIC: A Co-Evolving Attacker-Defender Adversarial Game for Robust LLM Safety},
author={Wen, Xiaoyu and He, Zhida and Qi, Han and Wan, Ziyu and Wen, Ying and Zheng, Tianhang and Xu, Xingcheng and Lu, Chaochao and Zhang, Qiaosheng},
journal={arXiv preprint arxiv:2602.01539},
year={2026}
}