Search papers, labs, and topics across Lattice.
This paper introduces a novel approach to detecting AI-generated text by identifying a structural signal at the sentence-pair level, termed Relational Over-Regularization (ROR), which captures the inflated inter-sentence transition variance characteristic of LLMs compared to human writing. By formalizing this signal, the authors develop the Cross-Source Stylometric Fingerprint Graph (CSFG), a graph-based framework that leverages various deviation signals as learnable features. The method achieves a remarkable 97.14% accuracy in binary detection, significantly outperforming existing graph-based methods and demonstrating robust generalization to unseen LLMs.
LLMs exhibit a distinct inter-sentence transition variance that can be exploited for highly accurate AI-generated text detection, achieving over 97% accuracy.
Detecting AI-generated text (AIGT) remains challenging because existing approaches rely on token-level statistical signals or independent stylometric features, causing them to overfit to specific generators and fail under distribution shift. We identify a structural signal at the sentence-pair level: LLMs produce inter-sentence transition variance that deviates from human writing through inflated variance driven by recurring similarity bursts at paragraph boundaries and templated transitions. We formalize this as Relational Over-Regularization (ROR) and validate it across four benchmarks (p<0.001). The central contribution is this relational problem formulation, not a novel GNN architecture; CSFG is one concrete instantiation for operationalizing ROR. To exploit this signal, we propose the Cross-Source Stylometric Fingerprint Graph (CSFG), a graph-based framework that encodes positional, sequential, semantic, and transition deviation signals as learnable GNN edge features. The per-edge signed deviation {\delta}_ij operationalizes ROR without hand-crafted thresholds and acts as a false-positive calibrator. CSFG achieves 97.14% accuracy under binary detection, outperforming the strongest graph-based baseline by 11.14 pp, with a false-positive rate of 1.57% and robust generalization to unseen LLMs in the inflated-variance regime; detection degrades for generators whose transition variance falls at or below the human baseline.
