Brief

Researchers have developed DEER, a novel framework for detecting machine-generated text that aims to overcome the limitations of current methods which degrade under domain shifts. DEER utilizes a Disentangled Mixture-of-Experts approach to separate domain-specific and domain-invariant knowledge, allowing for more robust adaptation to unseen text distributions. An instance-adaptive routing mechanism, driven by reinforcement learning, selects expert pathways based on detection rewards, leading to improved generalization and performance over existing state-of-the-art detectors. AI