Show Me How You Reason and I'll Tell You Who You Are: Reasoning Graphs for Robust LLM Authorship Attribution

2026-07-16

AI Summary

This work addresses the vulnerability of existing large language model (LLM) authorship attribution methods, which rely on surface-level linguistic features and are thus susceptible to paraphrasing, back-translation, and other obfuscation attacks, limiting their generalizability. To overcome this limitation, the authors propose a novel approach that integrates reasoning graphs with graph neural networks for the first time. Specifically, they employ an argument mining pipeline to extract deep reasoning structures from LLM-generated texts, construct corresponding reasoning graphs, and apply graph neural networks for authorship attribution. Experimental results demonstrate that the proposed method achieves up to a 27-percentage-point improvement in accuracy under obfuscation attacks and a 19-percentage-point gain on unseen LLM versions, substantially outperforming Longformer-based baselines and surpassing the robustness ceiling of conventional surface-feature approaches.

Abstract

Given the current trend to employ large language models (LLMs) in almost any imaginable context, LLM-generated text detection and authorship attribution have become a pressing issue. Prior work has primarily focused on surface-level linguistic features, an approach shown to be susceptible to paraphrasing and other obfuscation techniques. In this paper, we go beyond the linguistic surface, extracting and analysing reasoning structures in LLM-generated texts with the goal of capturing more complex signals of LLM authorship. We propose a graph neural network approach that leverages reasoning graphs extracted by an argument mining pipeline, demonstrating improved robustness and generalisation over a traditional Longformer baseline. Our approach outperforms the baseline by up to 27 percentage points under the obfuscation attacks such as paraphrasing and backtranslation, and 19 percentage points when evaluated on the texts generated by the unseen model versions, simulating real-world conditions in which new LLM versions are continuously released.

Problem

• LLM authorship attribution
• reasoning graphs
• obfuscation attacks
• text generation detection
• model generalization

Innovation

• reasoning graphs
• graph neural network
• authorship attribution
• LLM detection
• argument mining