Adnan KhanAdnanProduct & AI
San Francisco Bay Area
Get in touch

Source note3 min read

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

Source

arXiv

arxiv.org/abs/2607.14905v1

Read the source

Source

Plain-Language Brief

Researchers propose a graph neural network approach to detect and analyze reasoning structures in large language models (LLMs) generated texts for improved robustness against paraphrasing and obfuscation techniques. The method outperforms traditional Longformer baseline by up to 27 percentage points under obfuscation attacks, and 19 percentage points when tested on unseen model versions.

What the Document Says

  • The paper discusses the issue of detecting LLM-generated text and authorship attribution as a pressing concern given the increasing use of large language models in various contexts.
  • Previous work primarily focuses on surface-level linguistic features, which are susceptible to paraphrasing and other obfuscation techniques.
  • The authors aim to go beyond the linguistic surface by extracting and analyzing reasoning structures in LLM-generated texts to capture more complex signals of LLM authorship.
  • They propose a graph neural network approach using reasoning graphs extracted by an argument mining pipeline.
  • This approach demonstrates improved robustness and generalization over a traditional Longformer baseline.
  • The proposed method outperforms the baseline by up to 27 percentage points under obfuscation attacks such as paraphrasing and backtranslation.
  • It also outperforms by 19 percentage points when evaluated on texts generated by unseen model versions, simulating real-world conditions.
  • The researchers note that their approach can help in identifying the authorship of LLM-generated text and potentially improve the understanding of how these models reason.

Key Concepts Mentioned

  • large language models (LLMs)
  • authorship attribution
  • reasoning structures
  • graph neural network
  • argument mining pipeline
  • obfuscation attacks
  • paraphrasing
  • backtranslation

Claims Made by the Document

  • Previous work on LLM-generated text detection and authorship attribution focuses primarily on surface-level linguistic features.
  • The authors propose a method that captures more complex signals of LLM authorship by going beyond the linguistic surface.
  • The proposed graph neural network approach using reasoning graphs extracted by an argument mining pipeline demonstrates improved robustness and generalization over a traditional Longformer baseline.
  • The method outperforms the baseline by up to 27 percentage points under obfuscation attacks such as paraphrasing and backtranslation.
  • It also outperforms by 19 percentage points when evaluated on texts generated by unseen model versions.

Evidence Used

  • The authors state that given the increasing use of large language models in various contexts, LLM-generated text detection and authorship attribution have become a pressing issue.
  • They mention that previous work primarily focuses on surface-level linguistic features, which are susceptible to paraphrasing and other obfuscation techniques.
  • The authors describe their approach of using a graph neural network and reasoning graphs extracted by an argument mining pipeline for capturing more complex signals of LLM authorship.
  • They provide results showing improved robustness and generalization over a traditional Longformer baseline, with performance gains under obfuscation attacks and on unseen model versions.

Methods / Approach

UNKNOWN

Limitations Stated by the Authors

UNKNOWN

Limitations Noted by the Agent

Agent interpretation — not from the source.

UNKNOWN

Open Questions

UNKNOWN

Notes for Future Q&A

May include interpretation. Labeled where it does.

UNKNOWN

Raw Abstract / Excerpt

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.