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

arXiv cs.CL · 6 hours ago

Researchers developed a method using graph neural networks to detect whether text was written by large language models by analyzing reasoning structures rather than surface-level linguistic features. The approach achieved 27 percentage points higher accuracy than baseline methods when tested against paraphrasing and backtranslation attacks, and 19 percentage points better performance on unseen LLM versions. This enables more robust detection of AI-generated text that resists common obfuscation techniques.