Zero-Shot Statistical Tests for LLM-Generated Text Detection and Attribution using Finite Sample Concentration Inequalities

This repository contains the official code to replicate experiments from our paper:

🔗 Live Demo: Try it on Hugging Face Spaces

---

📁 Contents

Notebook	Description
Whiteboxdetection.ipynb	Runs statistical detection tests using access to the LLM (white-box setting).
Whiteboxattribution.ipynb	LLM Attribution: identifies the generating model among two sets of candidates.
Blackbox_adverserial.ipynb	Tests black-box detection and robustness of our detection against adversarial attacks.

All experiments were run on RunPod (B200) and use Hugging Face models and datasets where applicable.

---

🧪 How to Run

Each notebook is fully self-contained and includes:

• Installation and environment setup (in the first cells)
• Data loading and preparation
• Full experiment code with results

💻 No setup beyond running the notebook cells is required — all dependencies are installed inline.

---

📥 Data

Most datasets used in the experiments are downloaded programmatically from 🤗 Hugging Face.

However:

If you'd like to run the WritingPrompts experiments, you'll need to download the WritingPrompts dataset manually from:
🔗 https://github.com/facebookresearch/WritingPrompts

Place the downloaded data in the following path: data/writingPrompts/

Portions of this code are adapted from DetectGPT, available at: https://github.com/eric-mitchell/detect-gpt/tree/main

We thank the authors of DetectGPT for open-sourcing their codebase.

---

🧾 Citation

If you use this code or build upon our work, please cite:

@article{radvand2025zero,
  title={Zero-Shot Statistical Tests for LLM-Generated Text Detection using Finite Sample Concentration Inequalities},
  author={Radvand, Tara and Abdolmaleki, Mojtaba and Mostagir, Mohamed and Tewari, Ambuj},
  journal={arXiv preprint arXiv:2501.02406},
  year={2025}
}