SpecDetect: Simple, Fast, and Training-Free Detection of LLM-Generated Text via Spectral Analysis (AAAI 26 Main Track Oral)

This project provides the initial version of the source code for SpecDetect, a simple, fast, and training-free method for detecting LLM-generated text based on spectral analysis. Our method embeds the core implementation of Lastde (from Lastde), and we sincerely thank the developers of the Lastde project for their open-source contribution.

Note: This is the initial release of the SpecDetect codebase. We will update it with more complete implementations and detailed documentation in subsequent versions.

Environment

The key dependencies of our experimental environment are as follows:

• Python 3.8
• Pytorch 2.0.0
• Other dependencies:

  pip install -r requirements.txt

Code Structure & Implementation

The core logic of SpecDetect is organized in the following directories:

• shell_scripts: Contains executable scripts for triggering detection experiments (white-box/black-box settings).
• py_scripts: Includes core implementation code for spectral analysis, detection pipelines, and data processing (inherited and extended from Lastde).

Run Detection

You can run the detection experiments via the scripts in shell_scripts (consistent with Lastde's execution mode):

cd shell_scripts

# White-box detection
./detection_white_box.sh 

# Black-box detection
./detection_black_box.sh

Detection results will be saved in the experiment_results directory (with subdirectories corresponding to different experimental settings).

Acknowledgements

We are grateful to the authors of Lastde for their open-source code and insightful research. Their work laid the foundation for the implementation of SpecDetect.

Citation

If you find SpecDetect useful in your research, please cite our paper:

@article{luo2025specdetect,
  title={SpecDetect: Simple, Fast, and Training-Free Detection of LLM-Generated Text via Spectral Analysis},
  author={Luo, Haitong and Zhang, Weiyao and Wang, Suhang and Zou, Wenji and Lin, Chungang and Meng, Xuying and Zhang, Yujun},
  journal={arXiv preprint arXiv:2508.11343},
  year={2025}
}