In the current context of widespread application of Large Language Models (LLMs), the detection of LLM-generated text and authorship attribution has become increasingly prominent. Previous research has mainly focused on surface linguistic features, which are susceptible to paraphrasing and other obfuscation techniques. This paper proposes an analysis method that goes beyond surface language by extracting and analyzing reasoning structures in LLM-generated texts to capture more complex signals of LLM authorship.
We introduce a graph neural network approach that leverages reasoning graphs extracted by an argument mining pipeline, demonstrating improved robustness and generalization over a traditional Longformer baseline. Our approach outperforms the baseline by up to 27 percentage points under obfuscation attacks such as paraphrasing and backtranslation, and 19 percentage points when evaluated on texts generated by unseen model versions, simulating real-world conditions where new LLM versions are continuously released.
Blogger's Review: This paper significantly enhances the accuracy of LLM authorship attribution through the introduction of reasoning graphs, showcasing the potential of graph neural networks in natural language processing, which deserves attention and further exploration from researchers.