Reproduce “Duwak: Dual Watermarks in Large Language Models” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Chaoyi Zhu, Jeroen Galjaard, Pin-Yu Chen, Lydia Y. Chen
Researchers from TU Delft and IBM Research developed Duwak, a dual watermarking method for large language models that embeds two independent secret patterns: one in the token probability distribution and another through a quality-aware contrastive search sampling scheme. This approach achieves higher text quality, notably enhancing diversity, and provides up to 70% faster watermark detection while showing superior robustness against various post-editing attacks compared to single-watermark methods.
Key insights
- Combining two independent watermarking mechanisms synergistically enhances both detection efficiency and text quality, surpassing limitations of single-watermark approaches.
- The quality-aware contrastive search sampling effectively mitigates text degeneration and improves diversity, counteracting the potential negative impact of probability-biased token selection.
- Fisher's method can reliably combine independent statistical tests from the dual watermarks, providing a unified and robust measure for detection with enhanced statistical significance.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.