LLM-as-Critic: Contrastive and Adversarial Strategies for Authentic Text Verification

The rapid proliferation of sophisticated large language models (LLMs) has revolutionized content generation but concurrently poses significant challenges for distinguishing human-authored from AI-generated text. Traditional detection methods often struggle with the increasing fluency of LLM outputs and their vulnerability to adversarial manipulations. In response, we propose LLM-as-Critic, a novel discriminative framework that fine-tunes a pre-trained LLM to act as an expert judge of textual authenticity. Our method integrates a multi-objective training paradigm encompassing Binary Cross-Entropy loss for fundamental classification, a bespoke Contrastive Learning loss to maximize inter-class separation, and an Adversarial Training scheme to bolster robustness against sophisticated AI-generated content. Extensive experiments across diverse datasets, including news, creative writing, and academic papers, consistently demonstrate LLM-as-Critic's superior performance, achieving F1 scores up to 0.97, significantly outperforming baselines such as Perplexity-based Detectors, Stylometric Feature Analyzers, and Fine-tuned RoBERTa Classifiers. Furthermore, ablation studies validate the incremental contribution of each training component, while human evaluation confirms a higher agreement rate with our model's classifications, reinforcing its practical utility. LLM-as-Critic establishes a new state-of-the-art in AI-generated text detection, particularly excelling in generalization to unseen generators and resilience against adversarial attacks.