TLDRocket
Sign in

AI Robustness & Adversarial

1 summarised story about AI Robustness & Adversarial, each linking back to the original source. Browse all topics →

Friday, 17 July 2026

SAMark: A Self-Anchored Text Watermarking with Paragraph-Level Paraphrase Robustness

arXiv cs.CL 6 hours ago

SAMark is a text watermarking method that maintains robustness against paragraph-level paraphrasing by using position-independent semantic anchors rather than relying on sentence order. The method achieves 90.2% true positive rate at 1% false positive rate on paraphrasing attacks, outperforming prior methods by over 30% on average. This approach improves watermark detection while preserving text generation quality without the quality-robustness tradeoff seen in earlier techniques.