Detecting AI‑generated text (AIGT) remains difficult because most existing methods rely on token‑level statistics or independent stylometric cues, which cause over‑fitting to specific generators and degrade under distribution shift. We uncover a structural signal at the sentence‑pair level: large language models (LLMs) exhibit inflated inter‑sentence transition variance compared with human writing, manifested as similarity bursts at paragraph boundaries and templated transitions. We formalize this phenomenon as Relational Over‑Regularization (ROR) and validate it on four public benchmarks.
The approach builds a directed graph where each node is a sentence and edge weights are cosine similarities between adjacent sentence embeddings $s_{i}=\cos(\mathbf{v}_{i},\mathbf{v}_{i+1})$. The transition variance of a document is then computed as $\sigma^{2}=\frac{1}{N-1}\sum_{i=1}^{N-1}(s_{i}-\bar{s})^{2}$, with $\bar{s}$ the mean similarity. Empirically, AI‑generated texts have significantly larger $\sigma^{2}$ than human texts, and this gap persists across models and languages.
To capture higher‑order patterns we feed the similarity graph into a Graph Convolutional Network (GCN). The pipeline is straightforward: sentence embeddings → similarity graph → two GCN layers → global pooling → binary classifier. On benchmarks covering OpenAI, Anthropic, Google, and Meta generators, ROR improves detection accuracy by 4.2%–7.8% and retains about 85% recall on unseen generators.
Blogger's Review: ROR offers a novel detection angle by exploiting sentence‑level structural fluctuations, complementary to traditional token‑based signals. Its graph‑based implementation is lightweight and transferable, making it a promising component for real‑world AI‑text detectors.