Research map / Large language models / Language, safety and interpretability
Analyzing and Evaluating Unbiased Language Model Watermark
Yihan Wu, Xuehao Cui, Ruibo Chen, Heng Huang
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 |
|---|---|---|---|
| raise KGW bias strength delta from 0.5 to 2.0 | KGW · WMT16 ro-en | BLEU: 20.6198 -> 17.9401 | Hurt |
| raise Unigram bias strength delta from 0.5 to 2.0 | Unigram · WMT16 ro-en | BLEU: 20.0175 -> 18.4838 | Hurt |
| raise KGW delta from 0.5 to 2.0 | KGW · open-ended detection | TPR@FPR=1%: 43.9% -> 96.57% | Improved |
| raise Unigram delta from 0.5 to 2.0 | Unigram · open-ended detection | TPR@FPR=1%: 53.59% -> 97.85% | Improved |
| use unbiased MCmark(n=10) instead of DiPmark(alpha=0.4) | unbiased watermark · open-ended detection | TPR@FPR=1%: 74.11% -> 97.09% | Improved |
| increase MCmark list length n from 10 to 100 | MCmark · open-ended detection | TPR@FPR=1%: 97.09% -> 92.2% | Hurt |
| use DIPPER paraphrase attack instead of 10% random token replacement | MCmark(n=10) · robustness | TPR@1%FPR: 96.11% -> 5.10% | Hurt |
| use DIPPER paraphrase attack | KGW(delta=2.0) · robustness | TPR@1%FPR: 91.68% -> 1.35% | Hurt |
| use unbiased SynthID instead of Unigram(delta=2.0) for robustness | SynthID · robustness | TPR@1%FPR (random 30%): 95.63% -> 7.71% | Hurt |
| use ITS-Edit instead of EXP-Edit | edit-distance unbiased watermark · WMT16 ro-en | BLEU: 20.4879 -> 17.9904 | Hurt |
| raise Unigram delta from 0.5 to 2.0 | Unigram · unbiasedness, repeated prompts (SPMG) | BLEU deviation: 0.8311 -> 3.9869 | Hurt |
| use SynthID instead of no watermark | SynthID · unbiasedness, repeated prompts (SPMG) | BLEU deviation: 0.1199 -> 2.5266 | Hurt |
Closest papers by meaning
- Unbiased Watermark for Large Language Models (ICLR 2024 · similarity 0.91)
- Provable Robust Watermarking for AI-Generated Text (ICLR 2024 · similarity 0.90)
- Black-Box Detection of Language Model Watermarks (ICLR 2025 · similarity 0.90)
- BiMark: Unbiased Multilayer Watermarking for Large Language Models (ICML 2025 · similarity 0.89)
- Edit Distance Robust Watermarks via Indexing Pseudorandom Codes (NeurIPS 2024 · similarity 0.89)
- Theoretically Grounded Framework for LLM Watermarking: A Distribution-Adaptive Approach (NeurIPS 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.
