detectability

This paper formalizes when benchmark contamination is detectable, deriving information-theoretic limits and proposing power-calibrated audits that distinguish a clean benchmark from a powerless detector. It reports two-sided empirical findings on calibration efficacy and validity gates.

本文介绍了STELA，一个语言学感知的LLM水印框架，通过POS n-gram的句法可预测性来平衡文本质量和检测鲁棒性。该方法无需访问模型logits即可实现公开可验证的水印检测，在类型学多样化的语言（英语、中文、韩语）上展示了优异性能。