bmbgsj/REVEAL_think_3class

REVEALthink3class

REVEAL-think-3class is a reasoning-driven AI-Generated Content (AIGC) detection model based on Qwen3-8B. It uses a Think-then-Answer paradigm, generating a transparent reasoning chain (<think>...</think>) before outputting the final fine-grained classification (<answer>...</answer>).

This model is introduced in the paper: [Reasoning-Aware AIGC Detection via Alignment and Reinforcement](https://arxiv.org/abs/2604.19172).

🔗 Project Homepage & Code: https://aka.ms/reveal 📚 Associated Dataset: AIGC-text-bank

🌟 Model Overview

This model performs fine-grained detection, discriminating between three categories:

• —Human: Authentic human-authored text.
• —AI-Native: Fully machine-generated text.
• —AI-Polish: Human-authored text refined by AI to improve fluency and style while preserving original semantics.

🚀 How to Use

To run inference, simply use the `think.py` script provided in our GitHub repository. It handles prompt formatting, vLLM acceleration, and automatically extracts the final prediction along with fine-grained confidence scores.

python think.py \
    --model_path "bmbgsj/REVEAL_think_3class" \
    --text "The rapid advancement of Large Language Models has ushered in an era where AI-generated content is increasingly pervasive..."

📖 Citation

If you use this model in your research, please cite:

@misc{wang2026reasoningawareaigcdetectionalignment,
      title={Reasoning-Aware AIGC Detection via Alignment and Reinforcement}, 
      author={Zhao Wang and Max Xiong and Jianxun Lian and Zhicheng Dou},
      year={2026},
      eprint={2604.19172},
      archivePrefix={arXiv},
      primaryClass={cs.AI},
      url={https://arxiv.org/abs/2604.19172}, 
}