Watermarkability
Measuring and improving model compatibility with logit-based watermarks.
Different language models can respond very differently to the same logit-based watermarking rule. This project studies watermarkability: the degree to which a model can satisfy watermark constraints while maintaining generation quality.
The work develops measurements of model–watermark compatibility and methods for improving watermark performance across model families.
- Evaluates compatibility across language models and watermark rules
- Studies the relationship between constraint satisfaction and text quality
- Read the paper on arXiv
The method first improves model–watermark compatibility with LoRA, then uses a residual controller to meet the target watermark strength reliably.