Research map / Large language models / Language, safety and interpretability
LLM Watermark Evasion via Bias Inversion
Jeongyeon Hwang, Sangdon Park, Jungseul Ok
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 |
|---|---|---|---|
| apply negative logit bias beta=-4 to proxy suppression set during rewriting (ablation) | LLM paraphrasing attack (Llama-3.1-8B) · C4, SIR watermark, Llama-3.1-8B-Instruct | ASR: 54.2 -> 99.6 | Improved |
| strengthen negative bias beyond -4 (to -9) (ablation) | BIRA · C4, SIR watermark, Llama-3.1-8B-Instruct | LLM judgment score: 4.24 -> 2.94 | Hurt |
| strengthen negative bias beyond -4 (ablation) | BIRA · C4, SIR watermark, Llama-3.1-8B-Instruct | ASR: 99.6 -> 99.6 | No clear effect |
| raise percentile q of suppression set (suppress fewer, higher-surprisal tokens) to 0.9 (ablation) | BIRA · C4, SIR watermark, Llama-3.1-8B-Instruct | ASR: 99.6 -> 77.2 | Hurt |
| raise percentile q to 0.9 (ablation) | BIRA · C4, SIR watermark, Llama-3.1-8B-Instruct | LLM judgment score: 4.24 -> 4.60 | Improved |
| suppress all tokens (q=0) instead of q=0.5 (ablation) | BIRA · C4, SIR watermark, Llama-3.1-8B-Instruct | LLM judgment score: 4.24 -> 4.15 | Hurt |
| select suppression tokens by self-information instead of random (ablation) | BIRA · token selection, ratio 0.1-0.9 | ASR | Improved |
| use BIRA instead of SIRA on 600-token texts | rewriting attack · 600-token SIR text from Qwen2.5-32B | LLM judge: 1.33 -> 4.14 | Improved |
| use BIRA instead of SIRA on 600-token texts | rewriting attack · 600-token SIR text from Qwen2.5-32B | TPR@FPR=1%: 20.8 -> 4.0 | Improved |
| use BIRA instead of vanilla paraphrasing (GPT-4o-mini) | rewriting attack · C4, seven watermarks | average ASR: 49.6 -> 99.5 | Improved |
| use BIRA instead of vanilla paraphrasing | rewriting attack · SIR z-score | z-score: 0.19 -> -0.06 | Improved |
| shift SIR watermark signal to low-entropy tokens (ablation) | SIR watermark defense · BIRA robustness, appendix | detectability-quality tradeoff | Hurt |
Closest papers by meaning
- Revealing Weaknesses in Text Watermarking Through Self-Information Rewrite Attacks (ICML 2025 · similarity 0.88)
- Break the Trade-off Between Watermark Strength and Speculative Sampling Efficiency for Language Models (ICLR 2026 · similarity 0.88)
- PRO: Enabling Precise and Robust Text Watermark for Open-Source LLMs (NeurIPS 2026 · similarity 0.87)
- Optimizing Adaptive Attacks against Watermarks for Language Models (ICML 2025 · similarity 0.87)
- Watermark Stealing in Large Language Models (ICML 2024 · similarity 0.87)
- Enhancing LLM Watermark Resilience Against Both Scrubbing and Spoofing Attacks (NeurIPS 2025 · similarity 0.87)
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.
