Title: A Distribution-Free Framework for Rewrite-Based Human-text Detection via Knockoff Filtering

Abstract: This study introduces a statistical framework that is independent of data distribution, designed to transform any existing rewrite-based detector into one with finite-sample False Discovery Rate (FDR) guarantees, all without the need for retraining. The central insight of this approach is that rewrite-based detection methods inherently generate knockoff samples. By leveraging this property, the task of detecting text generated by Large Language Models (LLMs) can be recast as a multiple hypothesis testing problem characterized by a knockoff structure. This conceptual shift decouples the creation of detection statistics from the management of false discoveries. Consequently, current rewrite detectors can attain finite-sample FDR guarantees through a straightforward calibration step. We validate the efficacy of this method, demonstrating robust FDR control alongside significant detection power across three distinct detection models, nineteen different domains, and four various LLMs.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC