融合自一致性与对数秩的大语言模型生成文本检测OA

The rapid advancement of large language models(LLMs)in text generation has brought significant conve-nience but also raised concerns about potential misuse.Traditional detection methods reliant on training data face limita-tions in adapting to new domains and typically require access to the source model of the generated text.To address these challenges,this paper proposes a novel detection method that requires neither extensive training data nor access to the source model.This method integrates the self-consistency feature of LLMs with the ln-rank feature.Random partial dele-tions are performed at the sentence level on the input text.The modified text is fed into a large language model to recon-struct the missing content.A semantic similarity model is used to compute the semantic similarity between the original and reconstructed texts,from which a self-consistency score is derived.This self-consistency score is combined with the ln-rank feature to construct a comprehensive metric.Based on this metric,an effective classification threshold is estab-lished to determine the origin of text.Experimental results demonstrate that this method achieves superior performance on two real-world datasets,XSum and HC3,with an average detection accuracy of 92.52%and an AUROC of 96.13%.Fur-thermore,the method exhibits robust performance on Chinese datasets,achieving an average AUROC of 84.57%.In cross-domain detection scenarios,it outperforms the current state-of-the-art method by an average AUROC improvement of 6.53 percentage points,highlighting its exceptional generalization capability.