… Experiments across six watermarking schemes and three LLMs show that averaging across 3 models suppresses detection z-scores from 5-300 to below 2 below the detection threshold of 4 and reduces TPR at 5% FPR to below 50%, while improving quality by 27.5% and running 6 times faster than the best bas… …
Papers arxiv:2605.10977 PASA: A Principled Embedding-Space Watermarking Approach for LLM-Generated Text under Semantic-Invariant Attacks Published on May 9 Submitted by Sicheng Feng on May 13 HKUSTGZ Authors: Zhenxin Ai , Haiyun He Abstract PASA is a robust watermarking algorithm for large language… …
… Evaluation on four QA datasets and four attack types demonstrates the effectiveness of spectral entropy for backdoor detection. …
… Its standalone watermarks further achieve 93.6–100.0% detection with at most 1.9% false alarms . 🚀 Agent traces are not just logs — they are security interfaces. 🌐 Project: https://xushuwenn.github.io/RedAct Website/ 💻 Code: https://github.com/XuShuwenn/RedAct 📄 Paper: https://arxiv.org/abs/2606.10… …