Reinforcement-Guided Adaptive Manifold Perturbation Framework for Robust AI-Generated Text Detection

Large-scale language models (LLMs) have enabled AI-generated text to match human writing in fluency and semantic coherence, delivering substantial application value while posing severe security and trustworthiness risks. Existing detection frameworks primarily treat robustness as a static noise augmentation problem, overlooking the structured semantic manifold of natural language and the dynamic interaction between perturbation modeling and representation learning. In this work, we reconceptualize robust AI-generated text detection as a manifold-aware adaptive robustness learning problem, where perturbations are not predefined but dynamically optimized to expose semantically consistent yet decision-boundary-sensitive variations. Based on this perspective, we propose a novel AI generated text detection framework based on dynamic perturbation modeling and multi-view robust representation learning, termed the reinforcement learning-guided adaptive manifold perturbation network (RAMP-Net). The method integrates a reinforcement-guided dynamic perturbation generation mechanism, which adaptively adjusts perturbation strength under se mantic manifold constraints to construct semantically consistent yet discriminatively challenging training samples. A dual-stream residual fusion architecture is further designed to explicitly model and fuse clean and perturbed semantic representations, enhancing detection stability under strong perturbations. Moreover, a hard-parameter-sharing multi-task learning strategy is adopted to jointly optimize robust detection and semantic preservation objectives, mitigating semantic forgetting during strong perturbation training. Extensive experiments on representative cross domain detection tasks show that RAMP-Net outperforms state of-the-art methods in accuracy and generalization, verifying its effectiveness in complex environments. © 2026 IEEE.