Research map / Large language models / Language, safety and interpretability
An Ensemble Framework for Unbiased Language Model Watermarking
Yihan Wu, Ruibo Chen, Georgios Milis, Heng Huang
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 ensemble of n=5 keys | DiPmark (alpha=0.3) · Detectability, 250 tokens | TPR@0.01% FPR: 32.22% -> 66.77% | Improved |
| add ensemble of n=5 keys | gamma-reweight · Detectability, 500 tokens | TPR@0.01% FPR: 72.45% -> 88.58% | Improved |
| add ensemble of n=3 keys | MCMark (l=20) · Detectability, 250 tokens | TPR@0.01% FPR: 90.37% -> 91.71% | Improved |
| add ensemble of n=3 keys | MCMark (l=20) · Detectability, 500 tokens | median p-value: 8.30e-26 -> 2.58e-31 | Improved |
| increase ensemble size from 5 to 10 (ablation) | ENS-gamma-reweight · Detectability, 250 tokens | TPR@0.01% FPR: 64.14% -> 55.54% | Hurt |
| increase ensemble size from 5 to 10 (ablation) | ENS-DiPmark (alpha=0.4) · Detectability, 250 tokens | TPR@0.01% FPR: 67.15% -> 58.74% | Hurt |
| increase ensemble size from 3 to 5 (ablation) | ENS-MCMark (l=20) · Detectability, 500 tokens | TPR@0.01% FPR: 99.57% -> 98.90% | Hurt |
| vary ensemble size n=1..10 (ablation) | DiPmark and gamma-reweight · C4 subset, Llama-3.2-3B-Instruct | TPR@0.01% FPR and median p-value | Mixed |
| increase generation length (ablation) | all watermarking methods · C4 subset, Llama-3.2-3B-Instruct | TPR@0.01% FPR and median p-value: 250 -> 500 tokens, e.g. 32.22% -> 61.68% (DiPmark) | Improved |
| add ensemble n=3 under GPT-4o-mini paraphrase attack | SynthID (m=30) · GPT-4o-mini paraphrase attack | TPR@0.01% FPR: 13.47% -> 29.44% | Improved |
| compare ENS-MCMark vs SynthID under back translation | SynthID (m=30) · Back translation (En-Fr) | TPR@0.01% FPR: 64.53% -> 76.43% | Improved |
| add ensemble to MCMark (n=3) | MCMark · WMT En-Ro translation | BLEU: 20.35 -> 20.43 | No clear effect |
Closest papers by meaning
- BiMark: Unbiased Multilayer Watermarking for Large Language Models (ICML 2025 · similarity 0.90)
- Unbiased Watermark for Large Language Models (ICLR 2024 · similarity 0.90)
- GaussMark: A Practical Approach for Structural Watermarking of Language Models (ICML 2025 · similarity 0.89)
- Theoretically Grounded Framework for LLM Watermarking: A Distribution-Adaptive Approach (NeurIPS 2025 · similarity 0.89)
- Black-Box Detection of Language Model Watermarks (ICLR 2025 · similarity 0.89)
- Provable Robust Watermarking for AI-Generated Text (ICLR 2024 · 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.
