MD-ProTector: Multiple Data-Driven Prototype-Based LLM-Generated Text Detection

ACL ARR 2026 January Submission7915 Authors

Keywords: LLM-generated text detection, contrastive learning, prototype-based representation learning, robustness

Abstract: As LLM-generated content becomes more sophisticated, detection systems for distinguishing those texts from human-written must operate at scale while handling diverse writing styles and domains. Relying only on standard binary classification formulations can neglect diversity in writing style and domain within each class. We propose MD-ProTector, an encoder-based detector that models internal diversity in both classes using multiple prototypes, which are representative vectors summarizing groups of instances. MD-ProTector represents each class with multiple learnable prototypes that are updated during training. A prototype-to-instance contrastive loss, combined with instance-level contrastive loss, data-driven initialization and dynamic prototype updates, effectively leverages class-internal diversity. Experiments show that MD-ProTector improves detection performance while avoiding bias toward either human-written or LLM-generated texts, and exhibits robustness under distribution shifts from unseen domains and generator models.

Paper Type: Long

Research Area: Machine Learning for NLP

Research Area Keywords: LLM-generated text detection, prototype-based learning, contrastive learning

Contribution Types: NLP engineering experiment

Languages Studied: English

Submission Number: 7915