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

Name: Is Multilingual LLM Watermarking Truly Multilingual? Scaling Robustness to 100+ Languages via Back-Translation
Uploaded: 2026-09-15T11:07:03.076125+00:00
Duration: 323 s
Description: This episode explores a paper by Asim Mohamed and Martin Gubri on the failures of current multilingual LLM watermarking. Researchers found that tokenization fragments words in low-resource languages, destroying watermark signals. They propose STEAM, a method using back-translation and Bayesian optimization to recover signals across 133 languages.

This episode explores a paper by Asim Mohamed and Martin Gubri on the failures of current multilingual LLM watermarking. Researchers found that tokenization fragments words in low-resource languages, destroying watermark signals. They propose STEAM, a method using back-translation and Bayesian optimization to recover signals across 133 languages.

Open the full episode