Research map / Large language models / Language, safety and interpretability
A Watermark for Large Language Models
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add soft green-list logit bias delta=2.0 (gamma=0.5) during multinomial sampling vs no watermark | OPT-1.3B soft watermark · C4 RealNewsLike, T=200 | TPR at z=4 | Helped |
| lower green-list bias delta from 2.0 to 1.0 (ablation) | soft watermark, multinomial · C4 RealNewsLike, T=200 | TPR at z=4: 0.984 to 0.767 | Hurt |
| use 8-way beam search with soft watermark (ablation) | soft watermark delta=2.0 · C4 RealNewsLike, T=200 | TPR at z=4: 0.984 to 0.992 (gamma 0.5) | Helped |
| use smaller green list gamma=0.25 instead of 0.5 at delta=2.0 (ablation) | soft watermark, multinomial · C4 RealNewsLike, T=200 | TPR at z=4: 0.984 to 0.994 | Mixed |
| increase delta to 5.0 (ablation) | soft watermark, multinomial · C4 RealNewsLike, T=200 | TPR at z=4: 0.984 to 0.996 | Helped |
| hard red list rule (prohibit red tokens) vs soft bias | proof-of-concept watermark · theory | z-threshold false positive rate | No clear effect |
| skip repeated n-grams when computing green/red counts | detector · text analysis | false positives on repetitive text | Helped |
| T5 span replacement attack with budget epsilon=0.1 | soft watermark gamma=0.5 delta=2.0 · T5 span attack, multinomial | AUC decrease: -0.01 AUC | Hurt |
| emoji and character-substitution generative attacks on ChatGPT | watermark robustness · qualitative | watermark removal | Hurt |
Closest papers by meaning
- On the Learnability of Watermarks for Language Models (ICLR 2024 · similarity 0.93)
- GaussMark: A Practical Approach for Structural Watermarking of Language Models (ICML 2025 · similarity 0.92)
- Adaptive Text Watermark for Large Language Models (ICML 2024 · similarity 0.91)
- StealthInk: A Multi-bit and Stealthy Watermark for Large Language Models (ICML 2025 · similarity 0.91)
- Black-Box Detection of Language Model Watermarks (ICLR 2025 · similarity 0.91)
- Watermarks in the Sand: Impossibility of Strong Watermarking for Language Models (ICML 2024 · similarity 0.91)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.