SAMP: source-aware multi-prototype learning for machine-generated text detection

Abstract

Machine-generated text detection is commonly formulated as a binary classification problem that separates human-written texts from machine-generated ones. In heterogeneous settings involving diverse generators, domains, languages, and adversarial perturbations, binary classes are often not structurally simple or unimodal. Human-written texts typically exhibit broad dispersion, while machine-generated texts frequently contain both dominant modes and long-tailed substructures. However, standard binary detectors mainly learn a coarse human–machine decision boundary, providing limited constraints on the internal geometry of the learned representation space and leaving intra-class structures insufficiently organized. Moreover, by collapsing all machine-generated texts into a single homogeneous category, such detectors can obscure generator-dependent regularities and weaken generalization under distribution shifts. To address these issues, we propose SAMP, a Source-Aware Multi-Prototype learning framework for robust machine-generated text detection. SAMP represents each binary class with multiple class-conditional prototypes to capture intra-class heterogeneity, thereby explicitly accommodating complex internal data distributions. It further uses source-model identity strictly as auxiliary training supervision. This signal preserves generator-dependent relations within the machine-generated region and prevents fine-grained source regularities from being overshadowed by standard binary supervision. Experiments on MAGE, M4-multilingual, and RAID show that SAMP achieves competitive and robust performance across heterogeneous evaluation settings. It remains highly competitive on MAGE and achieves the strongest reported performance on M4-multilingual and RAID. Notably, on the attacked RAID setting, SAMP achieves 96.31% AUROC, 99.86% AUPR, and 16.11% FPR95 averaged over three independent training seeds, demonstrating strong reliability under challenging perturbation conditions.