The rapid advancement of Large Language Models (LLMs) has made distinguishing between LLM-generated and human-written text increasingly difficult, raising concerns about authenticity and security. Although supervised training can yield effective LLM-generated text detectors for specific LLMs, the continuous emergence of new models renders the process of labeling data and training individual models for each LLM and application scenario impractical. To address this challenge, we propose a novel framework that leverages generalized features. Specifically, we introduce LLM-conditional feature alignment (LCFA) to guide the model in learning domain-invariant features characteristic of LLM-generated text. Furthermore, we incorporate dynamic contrastive learning (DCL) to enhance the model’s robustness to data perturbations, thereby improving the generalization of learned representations. To facilitate evaluation under realistic conditions, we construct a new dataset, MLS, comprising text generated by state-of-the-art LLMs across multiple scenarios and languages. Experimental results on the MLS dataset demonstrate the efficacy of our proposed approach.