When Less is More: Minimal Prompts with LoRA for LLM Text Detection

Shiquan Wang, Ruiyu Fang, Mengxiang Li, Zhongjiang He, Shuangyong Song

Abstract: This paper presents a method for detecting LLM-generated Chinese text in the NLPCC 2025 Task 1. We utilize the Qwen2.5-72B-Instruct model as the base, applying parameter-efficient fine-tuning using Low-Rank Adaptation (LoRA) to optimize its performance on the detection task. Our approach integrates multiple models trained with different configurations and combines their predictions using an ensemble learning strategy based on majority voting. Additionally, we evaluate GECScore, a semantic-aware unsupervised baseline, for comparison. Through comprehensive experiments, we show that our ensemble strategy significantly enhances the model’s robustness and generalization ability across different evaluation scenarios. In the final test, our method achieved an average Macro-F1 score of 0.9665, securing second place in the competition. The results validate the effectiveness of the proposed method in distinguishing between human-written and LLM-generated texts.

External IDs:dblp:conf/nlpcc/WangFLHS25