Robust Detection of AI-Generated Text: Insights on Evolving LLMs and Adversarial Data

Large language models (LLMs) can generate highly fluent text, making it increasingly difficult to distinguish AI-generated text from human-written text and raising concerns about undetected misuse. Existing detectors face two key challenges: (i) as LLMs evolve, detectors struggle to generalize to unseen models and their probability distributions, and (ii) adversaries can paraphrase AI-generated text to evade detection. To tackle these problems, we propose a twofold strategy that (a) leverages LLMs themselves as detectors and (b) employs data augmentation via back-translation to improve robustness. We systematically evaluate several LLMs on the AI-generated text detection task. Moreover, augmenting the training data with paraphrases produced by back-translation exposes the detectors to adversarial variations and significantly enhances their resilience. Our best model achieves state-of-the-art performance, winning first place in the CCKS-2025 LLM-generated Text Detection Task with a macro-F1 score of 89.59%.