Is Multilingual LLM Watermarking Truly Multilingual? Scaling Robustness to 100+ Languages via Back-Translation

1. 1Identify the limitations of current multilingual watermarking methods
2. 2Evaluate the robustness of these methods in medium- and low-resource languages
3. 3Use back-translation to scale robustness to 100+ languages
4. 4Develop new methods that can effectively watermark LLM outputs across languages

ML researchers and engineers working on LLMs and watermarking techniques can benefit from this research to improve the robustness of their models across multiple languages

💡 Current multilingual watermarking methods are not truly multilingual and require improvement to scale to 100+ languages

🚨 Multilingual LLM watermarking methods are not as robust as claimed, especially in medium- and low-resource languages 🚨