Research map / Large language models / Language, safety and interpretability
Beyond Heuristic Tuning: Power-Calibrated LLM Watermarking
Xiaopu Wang, Zelin He, Chengyuan Liu, Runze Li
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 |
|---|---|---|---|
| choose (gamma, delta) by power-calibrated optimization under KL budget instead of grid search | KGW watermark · C4, WinMax-C attack 10%, GPT-2 | largest -log(KL) at TPR>=0.95: 2.15 -> 2.70 | Improved |
| choose (gamma, delta) by power calibration instead of grid search | KGW watermark · C4, WinMax-C attack 10%, OPT-125M | largest -log(KL) at TPR>=0.95: 1.86 -> 2.53 | Improved |
| choose (gamma, delta) by power calibration instead of DP (KL difference-in-probability) | KGW watermark · C4, WinMax-C attack 10%, Pythia-160M | largest -log(KL) at TPR>=0.95: 1.34 -> 2.70 | Improved |
| choose (gamma, delta) by power calibration instead of OPT-style log-perplexity optimization | KGW watermark · C4, WinMax-C attack 20%, GPT-2 | largest -log(KL) at TPR>=0.95: 2.01 -> 2.69 | Improved |
| choose the higher root gamma_h of the KL budget equation instead of the lower root (ablation) | Power-calibrated watermark · Theory, alpha=0.05, delta=2, n=50 | KL at equal power | Improved |
| use WinMax-C detector instead of full-sequence z-score under local signal removal | KGW watermark detection · signal-removal attack | expected signal | Improved |
| theoretical gamma' formula versus observed green-token rate | Dirichlet-prior theory · C4/LFQA/Wikipedia with OPT, Pythia, GPT-2 | R2: R2 >= 0.98 | Improved |
| use KL divergence as proxy for text quality | KL distortion metric · LFQA quality metrics | R2 vs BLEU: R2 0.566 (BLEU), 0.870 (ROUGE), 0.791 (BERTScore) | Mixed |
Closest papers by meaning
- Theoretically Grounded Framework for LLM Watermarking: A Distribution-Adaptive Approach (NeurIPS 2025 · similarity 0.90)
- WaterMax: breaking the LLM watermark detectability-robustness-quality trade-off (NeurIPS 2024 · similarity 0.89)
- Black-Box Detection of Language Model Watermarks (ICLR 2025 · similarity 0.89)
- Break the Trade-off Between Watermark Strength and Speculative Sampling Efficiency for Language Models (ICLR 2026 · similarity 0.89)
- PRO: Enabling Precise and Robust Text Watermark for Open-Source LLMs (NeurIPS 2026 · similarity 0.89)
- No Free Lunch in LLM Watermarking: Trade-offs in Watermarking Design Choices (NeurIPS 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.
