Detecting LLM-Generated Text with Trigram–Cosine Stylometric Delta: An Unsupervised and Interpretable Approach

Background: Contemporary methods for detecting synthetic text, including model-specific detectors and transformer-based classifiers, often rely on intensive training or on features tied to particular language models, which restricts their generalizability to unfamiliar LLMs and diverse domains. Purpose: This study aims to advance text attribution research by introducing a stylometry-based approach that utilizes trigram-based cosine delta as a lightweight and interpretable metric for distinguishing LLM-generated texts from human-written texts, irrespective of the underlying generation strategy. Method: A corpus of Russian diary entries was compiled, encompassing both authentic human-written texts and synthetic counterparts generated through few-shot prompting and finetuned LoRA models. To evaluate the effectiveness of the proposed approach, multiple stylometric-delta variations were examined, integrating uni-, bi-, and trigram features with Manhattan and cosine distance metrics. Results: The evaluation demonstrated that the trigram–cosine delta consistently achieved the highest performance across experimental conditions, reaching an Adjusted Rand Index of approximately 0.70. This markedly surpassed both the finetuned RuModernBERT baseline (ARI ≈ 0.28) and the classic unigram-based delta (ARI ≈ 0.53). Importantly, the method proved effective not only within the Russian diary corpus but also when applied to the RuATD benchmark, where it successfully separated human-authored and machine-generated texts and produced coherent clustering of related model families. Conclusion: The findings confirm that trigram–cosine stylometric delta offers a robust, interpretable, and computationally efficient strategy for detecting LLM-generated texts across diverse generation strategies, including few-shot prompting and finetuning. By capturing discourse-level stylistic cohesion, the method advances beyond surface fluency and provides a scalable, unsupervised alternative to classifier-based detectors. While current validation is limited to Russian diaries and selected generation models, the approach demonstrates clear potential for broader application across domains, languages, and emerging state-of-the-art LLMs.