Beyond Heuristic Tuning: Power-Calibrated LLM Watermarking

Xiaopu Wang, Zelin He, Chengyuan Liu, Runze Li
Proceedings of the 43rd International Conference on Machine Learning, PMLR 306:126614-126643, 2026.

Abstract

Logit-based watermarking is a widely used mechanism for identifying LLM generated content, yet its effectiveness is governed by a fundamental trade-off between detectability and semantic distortion. Existing analyses provide limited guidance for principled hyperparameter selection, leaving practical deployments reliant on heuristic tuning. In this work, we develop a power-calibrated statistical framework that establishes explicit quantitative relationships between watermark hyperparameters, detection power, and distortion. This characterization transforms watermark design into a guided optimization problem. Building on these results, we derive practical parameter selection procedures that achieve optimal trade-offs under constraints. Extensive experiments across multiple language models and datasets validate the theory and demonstrate that the proposed framework consistently identifies Pareto-optimal points.

Cite this Paper


BibTeX
@InProceedings{pmlr-v306-wang26bu, title = {Beyond Heuristic Tuning: Power-Calibrated {LLM} Watermarking}, author = {Wang, Xiaopu and He, Zelin and Liu, Chengyuan and Li, Runze}, booktitle = {Proceedings of the 43rd International Conference on Machine Learning}, pages = {126614--126643}, year = {2026}, editor = {Zhang, Tong and Dudik, Miroslav and Jaggi, Martin and Agarwal, Alekh and Li, Sharon and Schuurmans, Dale and Zhu, Jerry and Berkenkamp, Felix and Dong, Hanze and Bietti, Alberto}, volume = {306}, series = {Proceedings of Machine Learning Research}, month = {06--11 Jul}, publisher = {PMLR}, pdf = {https://raw.githubusercontent.com/mlresearch/v306/main/assets/wang26bu/wang26bu.pdf}, url = {https://proceedings.mlr.press/v306/wang26bu.html}, abstract = {Logit-based watermarking is a widely used mechanism for identifying LLM generated content, yet its effectiveness is governed by a fundamental trade-off between detectability and semantic distortion. Existing analyses provide limited guidance for principled hyperparameter selection, leaving practical deployments reliant on heuristic tuning. In this work, we develop a power-calibrated statistical framework that establishes explicit quantitative relationships between watermark hyperparameters, detection power, and distortion. This characterization transforms watermark design into a guided optimization problem. Building on these results, we derive practical parameter selection procedures that achieve optimal trade-offs under constraints. Extensive experiments across multiple language models and datasets validate the theory and demonstrate that the proposed framework consistently identifies Pareto-optimal points.} }
Endnote
%0 Conference Paper %T Beyond Heuristic Tuning: Power-Calibrated LLM Watermarking %A Xiaopu Wang %A Zelin He %A Chengyuan Liu %A Runze Li %B Proceedings of the 43rd International Conference on Machine Learning %C Proceedings of Machine Learning Research %D 2026 %E Tong Zhang %E Miroslav Dudik %E Martin Jaggi %E Alekh Agarwal %E Sharon Li %E Dale Schuurmans %E Jerry Zhu %E Felix Berkenkamp %E Hanze Dong %E Alberto Bietti %F pmlr-v306-wang26bu %I PMLR %P 126614--126643 %U https://proceedings.mlr.press/v306/wang26bu.html %V 306 %X Logit-based watermarking is a widely used mechanism for identifying LLM generated content, yet its effectiveness is governed by a fundamental trade-off between detectability and semantic distortion. Existing analyses provide limited guidance for principled hyperparameter selection, leaving practical deployments reliant on heuristic tuning. In this work, we develop a power-calibrated statistical framework that establishes explicit quantitative relationships between watermark hyperparameters, detection power, and distortion. This characterization transforms watermark design into a guided optimization problem. Building on these results, we derive practical parameter selection procedures that achieve optimal trade-offs under constraints. Extensive experiments across multiple language models and datasets validate the theory and demonstrate that the proposed framework consistently identifies Pareto-optimal points.
APA
Wang, X., He, Z., Liu, C. & Li, R.. (2026). Beyond Heuristic Tuning: Power-Calibrated LLM Watermarking. Proceedings of the 43rd International Conference on Machine Learning, in Proceedings of Machine Learning Research 306:126614-126643 Available from https://proceedings.mlr.press/v306/wang26bu.html.

Related Material