Research map / Large language models / Language, safety and interpretability
In-Context Watermarks for Large Language Models
Yepeng Liu, Xuandong Zhao, Christopher Kruegel, Dawn Song, Yuheng Bu
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (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 |
|---|---|---|---|
| use stronger LLM (GPT-o3-mini instead of GPT-4o-mini) (ablation) | Initials ICW · ELI5 (DTS) | ROC-AUC: 0.572 -> 0.999 | Improved |
| use stronger LLM (GPT-o3-mini instead of GPT-4o-mini) (ablation) | Acrostics ICW · ELI5 (DTS) | ROC-AUC: 0.590 -> 1.000 | Improved |
| use stronger LLM (GPT-o3-mini instead of GPT-4o-mini) (ablation) | Lexical ICW · ELI5 (DTS) | ROC-AUC: 0.910 -> 0.995 | Improved |
| use stronger LLM (GPT-o3-mini instead of GPT-4o-mini) (ablation) | Unicode ICW · ELI5 (DTS) | ROC-AUC: 1.000 -> 1.000 | No clear effect |
| long-context injection via paper (IPI) instead of system prompt (DTS) (ablation) | Acrostics ICW with GPT-o3-mini · ICLR papers | ROC-AUC: 1.000 -> 0.997 | No clear effect |
| paraphrase watermarked text with ChatGPT (ablation) | Unicode ICW · ELI5 (DTS) | ROC-AUC: 1.000 -> 0.500 | Hurt |
| 30% synonym replacement attack (ablation) | Lexical ICW · ELI5 (DTS) | ROC-AUC: 0.995 -> 0.758 | Hurt |
| 30% deletion attack (ablation) | Initials ICW · ELI5 (DTS) | ROC-AUC: 0.999 -> 0.998 | No clear effect |
| paraphrase attack (ablation) | Acrostics ICW · ELI5 (DTS) | ROC-AUC: 1.000 -> 0.922 | Hurt |
| increase output length from 300 to 1000 words (ablation) | Initials ICW · ELI5 (DTS) | ROC-AUC: 0.999 -> 0.998 | No clear effect |
| ICW instead of PostMark (text quality) | GPT-o3-mini · ELI5 text quality | LLM-judge overall: 2.997 -> 4.813 | Improved |
| Initials ICW instruction on text quality (ablation) | GPT-o3-mini · ELI5 text quality | LLM-judge overall: 4.992 -> 4.282 | Hurt |
Closest papers by meaning
- Can Watermarking Large Language Models Prevent Copyrighted Text Generation and Hide Training Data? (AAAI 2025 · similarity 0.90)
- PRO: Enabling Precise and Robust Text Watermark for Open-Source LLMs (NeurIPS 2026 · similarity 0.90)
- Learning to Follow In-Context Watermark Instructions via Self-Distillation (NeurIPS 2026 · similarity 0.90)
- Unbiased Watermark for Large Language Models (ICLR 2024 · similarity 0.90)
- On the Learnability of Watermarks for Language Models (ICLR 2024 · similarity 0.89)
- Theoretically Grounded Framework for LLM Watermarking: A Distribution-Adaptive Approach (NeurIPS 2025 · similarity 0.89)
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.
