Watermarking LLM Outputs
HardConceptualwatermarkingdetectionsafety
Related module: Evaluation and interpretability
Problem statement
Watermarking embeds detectable patterns in LLM-generated text without visibly altering quality.
- How do red-green list watermarks work?
- What is the trade-off between detectability and text quality?
- How robust are watermarks to paraphrasing attacks?