Reproduce “Unveiling Large Language Models Generated Texts: A Multi-Level Fine-Grained Detection Framework” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Zhen Tao, Zhiyu Li, Runyu Chen, Dinghao Xi, Wei Xu
Researchers from Renmin University of China and collaborators developed the Multi-level Fine-grained Detection (MFD) framework, which precisely quantifies the degree of LLM involvement at the sentence level across lexical, grammatical, and syntactic dimensions. The framework achieved superior detection accuracy and robustness, outperforming state-of-the-art models with a Mean Absolute Error of 0.1347 and maintaining strong performance against texts generated by advanced LLMs like GPT-4.
Key insights
- Integrating low-level statistical, high-level semantic (via contrastive learning), and deep-level linguistic (via LLM analysis) features significantly enhances AI-generated text detection.
- A contrastive learning approach with adversarially rewritten texts improves detector robustness against common evasion tactics like paraphrasing and advanced LLM outputs.
- Fine-grained, sentence-level analysis across lexical, grammatical, and syntactic dimensions offers a more precise understanding of LLM involvement in mixed-content documents.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.