SSG: Logit-Balanced Vocabulary Partitioning for LLM Watermarking

This repository contains the official implementation for the ACL 2026 paper: "SSG: Logit-Balanced Vocabulary Partitioning for LLM Watermarking".

---

🛠️ Framework & Acknowledgement

Our codebase is built upon MarkLLM, a professional and comprehensive open-source toolkit for LLM watermarking. We sincerely thank the authors of MarkLLM for their well-structured framework, which greatly facilitated our implementation and evaluation.

We have modified and extended the vocabulary partitioning and logit modulation modules in MarkLLM to integrate our proposed SSG algorithm.

---

🚀 Quick Start

Prerequisites

Ensure you have installed the required environments as specified in MarkLLM. You can install the dependencies via:

pip install -r requirements.txt

SSG Script

We provide a unified script to run a quick pipeline demo of the SSG watermarking process, including text generation with logit modulation and subsequent watermark detection.

To run the demo, simply execute:

bash main.sh

📝 Citation

If you find this repository, our algorithm, or the experimental results useful in your research, please consider citing our work:

@misc{gu2026ssglogitbalancedvocabularypartitioning,
      title={SSG: Logit-Balanced Vocabulary Partitioning for LLM Watermarking}, 
      author={Chenxi Gu and Xiaoning Du and John Grundy},
      year={2026},
      eprint={2604.22438},
      archivePrefix={arXiv},
      primaryClass={cs.CR},
      url={https://arxiv.org/abs/2604.22438}, 
}