Researchers have introduced a new metric called "watermarkability" to assess how well language models are compatible with logit-based watermarking techniques. This metric quantifies the cost associated with a model assigning sufficient probability to a designated token set. To improve this compatibility, a two-stage method is proposed: first, a low-rank adapter is trained to increase the probability mass of the target set, and second, a controller applies a minimal logit bias to meet the target probability. Evaluations on OPT-125M and Qwen3_8B models demonstrated that this approach significantly reduces perplexity increases while maintaining high reliability for watermark adherence. AI
IMPACT Enhances the reliability of identifying AI-generated text and encoding provenance information in language models.
RANK_REASON Academic paper introducing a new metric and method for language model watermarking. [lever_c_demoted from research: ic=1 ai=1.0]
- alphaXiv
- arXiv
- CatalyzeX
- Connected Papers
- DagsHub
- Gotit.pub
- Hugging Face
- Litmaps
- OPT-125M
- Qwen3_8B
- ScienceCast
- scite Smart Citations
- Watermarkability
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
