Title: DEER: Disentangled Mixture of Experts with Instance-Adaptive Routing for Generalizable Machine-Generated Text Detection

Abstract:

As Large Language Models (LLMs) advance rapidly, detecting machine-generated text has emerged as a crucial challenge. However, current detection methods suffer significant performance drops when facing domain shifts. Our systematic pilot studies reveal that this vulnerability stems from two core issues in existing generalization strategies: the failure to fully retain domain-specific knowledge during multi-domain training, and the disconnect between knowledge retrieval and the detection goal during inference.

To bridge these gaps, we introduce DEER, a Disentangled mixturE-of-ExpeRts framework. DEER explicitly separates domain-local and domain-invariant knowledge into distinct expert modules. Rather than relying on static domain matching, the framework utilizes a reinforcement learning-based router that chooses expert pathways according to instance-level detection rewards. This approach creates a task-aligned, domain-agnostic mechanism that prioritizes detection utility over stylistic similarity, thereby ensuring robust adaptation to unseen distributions.

Extensive experimental results show that DEER consistently surpasses state-of-the-art detectors. It achieves average F1 score improvements of 1.28% and 2.92%, alongside accuracy gains of 1.35% and 2.26% on in-domain and out-of-domain datasets, respectively. These findings highlight DEER’s potential for providing reliable generalization in open-world deployment scenarios.

Source: arXiv Generated at: 2026-06-04 00:00:00 UTC