Research map / Large language models / Language, safety and interpretability
How Good is Post-Hoc Watermarking With Language Model Rephrasing?
Pierre Fernandez, Tom Sander, Hady Elsahar, Hongyan Chang, Tomáš Souček, Valeriu Lacatusu, Tuan A. Tran, Sylvestre-Alvise Rebuffi, Alexandre Mourachko
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 |
|---|---|---|---|
| add context-aware chunking (ablation) | post-hoc rephrasing watermark · Dickens excerpts (4000 tokens) | detection -log10 p: 24.9 -> 198.3 | Improved |
| add context-aware chunking (ablation) | post-hoc rephrasing watermark · Dickens excerpts (1500 tokens) | SBERT similarity: 0.85 -> 0.91 | Improved |
| add context-aware chunking (ablation) | post-hoc rephrasing watermark · Dickens excerpts (500 tokens) | detection -log10 p: 19.3 -> 19.3 | No clear effect |
| add context-aware chunking (ablation) | post-hoc rephrasing watermark · Dickens excerpts (4000 tokens) | length ratio: 0.70 -> 0.86 | Improved |
| increase model size from Llama-3.1-8B to Llama-3.3-70B | Gumbel-max code watermarking · HumanEval/MBPP code (T=1.4, passed samples) | -log10 p-value (median): 1.83 -> 0.44 | Hurt |
| increase model size from Llama-3.1-8B to Llama-3.3-70B | Gumbel-max code watermarking · HumanEval/MBPP code (T=1.4) | pass@1: 0.71 -> 0.90 | Improved |
| raise sampling temperature from 1.0 to 1.4 | Gumbel-max code watermarking with Llama-3.1-8B · HumanEval/MBPP code | pass@1: 0.92 -> 0.71 | Hurt |
| raise sampling temperature from 1.0 to 1.4 | Gumbel-max code watermarking with Llama-3.1-8B · HumanEval/MBPP code (passed samples) | -log10 p-value (median): 0.74 -> 1.83 | Improved |
| raise temperature from 0.80 to 1.20 | Gumbel-max watermarking · Wikipedia ru | -log10 p-value (median): 1.355 -> 23.689 | Improved |
| raise temperature from 0.80 to 1.20 | Gumbel-max watermarking · Wikipedia ru | SBERT similarity: 0.908 -> 0.836 | Hurt |
| rephrase non-English (ru) instead of English | Gumbel-max watermarking · Wikipedia (T=1.0) | SBERT similarity: 0.954 -> 0.888 | Hurt |
| add beam search to decoding (ablation) | Green-red, SynthID, MorphMark, DiPMark · Dickens rephrasing | cross entropy vs watermark strength | Improved |
Closest papers by meaning
- Theoretically Grounded Framework for LLM Watermarking: A Distribution-Adaptive Approach (NeurIPS 2025 · similarity 0.90)
- Watermarking Makes Language Models Radioactive (NeurIPS 2024 · similarity 0.89)
- On the Learnability of Watermarks for Language Models (ICLR 2024 · similarity 0.89)
- Black-Box Detection of Language Model Watermarks (ICLR 2025 · similarity 0.89)
- PRO: Enabling Precise and Robust Text Watermark for Open-Source LLMs (NeurIPS 2026 · similarity 0.89)
- BiMark: Unbiased Multilayer Watermarking for Large Language Models (ICML 2025 · similarity 0.88)
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.
