Watermarking Diffusion Language Models with Learned Foresight and Balanced Bands
Summary
- Researchers introduce a watermarking method for diffusion language models that uses learned vocabulary bands and a foresight predictor.
- The approach addresses unique challenges in diffusion models where tokens can be chosen before their preceding tokens.
- Experiments on WaterBench with LLaDA-8B-Instruct show improved detection z-scores while maintaining unwatermarked text task accuracy.
Why it matters
As diffusion language models become more prevalent, robust watermarking techniques are essential for provenance tracking and detecting machine-generated text without harming output quality.