Beyond the Final Actor: Modeling the Dual Roles of Creator and Editor for Fine-Grained LLM-Generated Text Detection

1. Problem Statement

The rapid proliferation of Large Language Models (LLMs) has created a critical need for detecting synthetic text. However, existing detection methods suffer from significant limitations:

• Oversimplification: Most current approaches use binary (Human vs. LLM) or ternary (Human, LLM, Mixed) classification.
• Regulatory Insufficiency: These coarse categories fail to distinguish between nuanced collaboration modes that carry different ethical and policy implications. Specifically, they cannot differentiate between:
  1. LLM-Polished Human Text: A human writes a draft, and an LLM refines it (often considered legitimate assistance).
  2. Humanized LLM Text: An LLM generates text, and a human edits it to evade detection (often considered academic dishonesty or fraud).
• The Core Challenge: Distinguishing these four classes requires understanding not just the final output, but the dual roles involved in the text's creation: the Creator (who establishes the logical foundation) and the Editor (who modifies the linguistic surface).

2. Methodology: RACE (Rhetorical Analysis for Creator-Editor Modeling)

The authors propose RACE, a fine-grained detection framework that explicitly models the distinct signatures of the creator and the editor using Rhetorical Structure Theory (RST).

Key Components:

1. Dual Trace Extraction:

   • The framework utilizes an end-to-end RST parser to segment text into Elementary Discourse Units (EDUs) and construct a logical tree of rhetorical relations (e.g., Cause, Contrast, Elaboration).
   • This separates the text into its logical organization (Creator's trace) and surface-level linguistic expression (Editor's trace).

2. Logic-Aware Graph Initialization:

   • The text is formalized as a multi-relational graph where nodes represent EDUs and rhetorical relations.
   • Descendant Span Pooling: Node features are initialized by pooling semantic embeddings from the underlying text spans, ensuring relation nodes capture the global semantic context of the segments they govern.
   • Information Bottleneck Projection: A dimensionality reduction layer filters out surface-level lexical noise, forcing the model to focus on structural features relevant to authorship rather than superficial word choices.

3. Rhetoric-Guided Message Passing:

   • A Relational Graph Convolutional Network (RGCN) is employed to propagate information across the graph.
   • Unlike standard GCNs, RGCN uses relation-specific weight matrices to learn distinct propagation rules for different rhetorical logics.
   • Basis Decomposition: To prevent parameter explosion given the large number of rhetorical relations, the model uses basis decomposition to share weights across relations, acting as a regularizer.

4. Graph Readout and Classification:

   • The final representation is derived from the root node of the RST tree (representing the global rhetorical intent).
   • The model is optimized using a joint loss function combining Supervised Contrastive Loss (to cluster similar classes and separate distinct ones) and Cross-Entropy Loss.

3. Key Contributions

• Task Definition: The paper introduces a rigorous four-class detection setting that explicitly separates "Mixed" text into LLM-Polished Human and Humanized LLM, addressing a gap in current regulatory frameworks.
• Novel Framework (RACE): It proposes the first method to decouple the Creator (logical structure) from the Editor (linguistic style) using RST-based graph learning.
  • Insight: Experiments show that the Creator's logical structure (e.g., Attribution, Background) remains persistent even after editing, while the Editor's influence is primarily on surface-level features.
• Performance: RACE achieves state-of-the-art results with low false alarm rates, outperforming 12 baselines.

4. Experimental Results

The model was evaluated on the HART dataset (re-split for the four-class setting) across diverse domains (News, Essays, ArXiv, Writing).

• Main Metrics:
  • TPR@1% FPR (True Positive Rate at 1% False Positive Rate): RACE achieved 83.06% average, significantly outperforming the best baseline (CoCo at 79.70%).
  • Macro-AUROC: RACE achieved 97.99%, demonstrating robust probability estimation.
• Ablation Studies:
  • Removing the Relation types (using vanilla GCN) caused a significant drop, confirming the necessity of modeling rhetorical logic.
  • Removing the Information Bottleneck hurt performance on LLM-Polished texts, proving the module's role in preventing overfitting to surface-level editing patterns.
• Robustness:
  • Out-of-Distribution (OOD): RACE maintained superior performance across unseen domains compared to baselines, suggesting its reliance on logical structure is more generalizable than entity-based or purely statistical methods.
  • Text Length: RACE showed particular advantage on shorter texts (200-600 tokens), where capturing nuanced structural differences is harder.
• Feature Discriminability: Clustering metrics (Davies-Bouldin and Calinski-Harabasz indices) confirmed that RACE learns a more compact and distinct embedding space than competitors.

5. Significance and Impact

• Policy Alignment: By distinguishing between legitimate assistance (polishing) and malicious evasion (humanizing), RACE provides a technical foundation for nuanced AI regulation in academic and professional settings.
• Paradigm Shift: The paper moves the field from "detecting the final actor" to "modeling the generative process." It demonstrates that the logical architecture of a text is a more stable fingerprint of the original creator than the surface-level style, which is easily altered by editors.
• Future Directions: The work opens avenues for Logic-Based Model Attribution (tracing specific LLMs even after heavy editing) and Fine-Grained Forensics (identifying exactly which parts of a text were human-edited vs. machine-generated).

In conclusion, RACE represents a significant advancement in AI text detection by leveraging rhetorical structure to disentangle the complex interplay between human and machine contributions, offering a more precise and ethically aligned solution for the era of collaborative AI writing.