MGTEVAL: An interactive platform for systematically evaluating machine-generated text detectors

The researchers introduce MGTEVAL, a scalable platform for systematically evaluating machine-generated text (MGT) detectors, addressing the problem that existing evaluations are scattered across different datasets, preprocessing, attacks, and metrics, making results difficult to compare and reproduce. The platform organizes the workflow into four components: dataset construction, dataset attack, detector training, and performance evaluation; it supports building custom benchmarks by generating machine text with configurable LLMs, applies 12 text attacks to the test set, trains detectors through a unified interface, and reports effectiveness, robustness, and efficiency. The platform provides both command-line and Web interfaces, enabling experiments without rewriting code.