# Breaking the Generator Barrier: Disentangled Representation for Generalizable AI-Text Detection Source: [Breaking the Generator Barrier: Disentangled Representation for Generalizable AI-Text Detection]() Collection index: [llms.txt]() Data updated: 2026-04-17T05:00:04.441Z Authors: Xiao Pu, Zepeng Cheng, Lin Yuan, Yu Wu, Xiuli Bi Published: 2026-04-15T10:14:06.000Z ## Abstract As large language models (LLMs) generate text that increasingly resembles human writing, the subtle cues that distinguish AI-generated content from human-written content become increasingly challenging to capture. Reliance on generator-specific artifacts is inherently unstable, since new models emerge rapidly and reduce the robustness of such shortcuts. This generalizes unseen generators as a central and challenging problem for AI-text detection. To tackle this challenge, we propose a progressively structured framework that disentangles AI-detection semantics from generator-aware artifacts. This is achieved through a compact latent encoding that encourages semantic minimality, followed by perturbation-based regularization to reduce residual entanglement, and finally a discriminative adaptation stage that aligns representations with task objectives. Experiments on MAGE benchmark, covering 20 representative LLMs across 7 categories, demonstrate consistent improvements over state-of-the-art methods, achieving up to 24.2% accuracy gain and 26.2% F1 improvement. Notably, performance continues to improve as the diversity of training generators increases, confirming strong scalability and generalization in open-set scenarios. Our source code will be publicly available at https://github.com/PuXiao06/DRGD. ## Recorded benchmark results A research result applies to its stated dataset, split, metric and evaluation configuration. Compare like-for-like settings, preserve the metric's unit and direction, and distinguish paper-reported results from independent reproduction. The best result recorded here is not a claim of exhaustive, current global coverage. - [AI-Generated Text Detection on OPT unseen category](): AI-generated text detection; OPT unseen category; Accuracy: 89.1. - [AI-generated text detection on MAGE FLAN-T5 1.0 (test)](): AI-generated text detection; MAGE FLAN-T5 1.0 (test); Accuracy: 68.9. - [AI-generated text detection on MAGE GPT 1.0 (test)](): AI-generated text detection; MAGE GPT 1.0 (test); Accuracy: 82.7. - [AI-generated text detection on MAGE (LLaMA) 1.0 (test)](): AI-generated text detection; MAGE (LLaMA) 1.0 (test); Accuracy: 88. - [AI-generated text detection on MAGE OPT 1.0 (test)](): AI-generated text detection; MAGE OPT 1.0 (test); Accuracy: 89.1. - [AI-generated text detection on MAGE GLM 1.0 (test)](): AI-generated text detection; MAGE GLM 1.0 (test); Accuracy: 94.1. - [AI-generated text detection on MAGE BigScience 1.0 (test)](): AI-generated text detection; MAGE BigScience 1.0 (test); Accuracy: 96.7. - [AI-generated text detection on MAGE (EleutherAI) 1.0 (test)](): AI-generated text detection; MAGE (EleutherAI) 1.0 (test); Accuracy: 97.2. - [Machine-generated text detection on MEGA FLAN-T5](): Machine-generated text detection; MEGA FLAN-T5; Accuracy: 77.4. - [Machine-generated text detection on MEGA (GPT split)](): Machine-generated text detection; MEGA (GPT split); Accuracy: 86.9. - [Machine-generated text detection on MEGA LLaMA](): Machine-generated text detection; MEGA LLaMA; Accuracy: 91.7. - [Machine-generated text detection on MEGA OPT](): Machine-generated text detection; MEGA OPT; Accuracy: 93.8. - [Machine-generated text detection on MEGA GLM](): Machine-generated text detection; MEGA GLM; Accuracy: 95.7. - [Machine-generated text detection on MEGA BigScience](): Machine-generated text detection; MEGA BigScience; Accuracy: 98.7. - [Machine-generated text detection on MEGA EleutherAI](): Machine-generated text detection; MEGA EleutherAI; Accuracy: 98.3. - [AI-Generated Text Detection on MAGE GPT-4 OOD](): AI-generated text detection; MAGE GPT-4 OOD; Accuracy: 61.8. - [AI-Generated Text Detection on MAGE GPT-5 OOD](): AI-generated text detection; MAGE GPT-5 OOD; Accuracy: 68.7. - [AI-Generated Text Detection on MAGE Claude-sonnet-4-5 OOD](): AI-generated text detection; MAGE Claude-sonnet-4-5 OOD; Accuracy: 57.1. - [AI-Generated Text Detection on MAGE DeepSeek-R1 OOD](): AI-generated text detection; MAGE DeepSeek-R1 OOD; Accuracy: 71. ## Original sources and implementation [Original paper]() [PDF]()