Elamur Research⌕
☰
home / papers / 2026 / Synergizing Stylometrics with Semantics: Dual-Path Framework for LLM Detection and Attribution

Synergizing Stylometrics with Semantics: Dual-Path Framework for LLM Detection and Attribution

Paper · 2026 · Xingyu Lu, Yumeng Ma, Xiang Zhou et al.

Summary

This work addresses attributing a text either to a human or to the specific LLM that wrote it. Likelihood-based detectors such as perplexity or curvature scores, and fine-tuned semantic classifiers, both lose reliability on short texts, derivative content, and cross-domain data. The authors observe that LLMs keep stable, topic-independent writing habits ("style inertia"). Building on this, they propose Stylometric-Semantic LLM Attribution (SSLA), which builds explicit stylistic profiles from three kinds of features:

  • static linguistic metrics drawn from dependency parsing and lexical statistics;
  • a rigidity feature computed by having a probe LLM rewrite the text;
  • interpretable syntactic n-grams (SN-Grams) taken from dependency trees.

An attention-based dual-path module merges these signals with RoBERTa contextual embeddings. On text from six LLM families across domains including Wikipedia and essays, SSLA reaches 95.6% Macro-F1 on Wikipedia, outperforming baselines such as OTBDetector. It also stays robust when models are pushed into a shared persona. Its feature-level explanations make it well suited to offline forensic attribution where transparency matters.

Key points

  • The static linguistic vector has six dimensions, including average dependency depth and distance, Type-Token Ratio, and sentence length. It is combined with the rewrite-based feature and SN-Grams.
  • Macro-F1 is 0.956 ± 0.011 for standard attribution over 6 LLMs. In an extreme adversarial setting, where 4 SOTA LLMs are forced into an identical persona, it remains 0.952.
  • Replacing the default rewriter, DeepSeek-V3, with Gemini-1.5-Flash leaves Average F1 almost unchanged (0.954 vs. 0.956). This suggests the rigidity signal does not depend on the choice of probe.
  • In ablations, attention fusion generalizes better than simply concatenating RoBERTa and SN-Gram features, especially on Wikipedia. SHAP analysis shows SN-Gram and lexical-density features carry substantial decision weight.
  • The authors list several limitations:
    • extra computational cost;
    • reliance on high-quality parsers, which is an obstacle for low-resource languages;
    • evaluation restricted to closed-set attribution;
    • weaker results on texts under 50 tokens.

Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.

 Institutions

🎓Chongqing University of Posts and TelecommunicationsChongqing · China2 works · 1 citation

Also on the byline: Chongqing Medical University.

 Fields, topics and concepts

📚 Shelves 5The field of science this work belongs to (OpenAlex fields).Computer Science · 16,667 worksArtificial Intelligence · 16,194 worksArts and Humanities · 16,039 worksLanguage and Linguistics · 15,891 worksComputer Vision and Pattern Recognition · 2,869 works

Cite

Cite as Xingyu Lu, Yumeng Ma, Xiang Zhou, Shengli Gan, Guiying Deng, Yang Wen and Yanbing Liu. 2026. Synergizing Stylometrics with Semantics: Dual-Path Framework for LLM Detection and Attribution. Findings of the Association for Computational Linguistics: ACL 2026.

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.18653/v1/2026.findings-acl.1855

@inproceedings{lu2026synergizing,
  title = {{Synergizing Stylometrics with Semantics: Dual-Path Framework for LLM Detection and Attribution}},
  author = {Xingyu Lu, Yumeng Ma, Xiang Zhou, Shengli Gan, Guiying Deng, Yang Wen and Yanbing Liu},
  booktitle = {Findings of the Association for Computational Linguistics: ACL 2026},
  year = {2026},
  doi = {10.18653/v1/2026.findings-acl.1855},
  url = {https://doi.org/10.18653/v1/2026.findings-acl.1855}
}

References

24 references, 8 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Harika Abburi, Sanmitra Bhattacharya, Edward Bowen, and Nirmala Pudota. 2025. Ai-generated text detection: A multifaceted approach to binary and multiclass classification. arXiv preprint arXiv:2505.11550. — [arXiv:2505.11550](https://arxiv.org/abs/2505.11550) find
  2. Guangsheng Bao, Yanbin Zhao, Zhiyang Teng, Linyi Yang, and Yue Zhang. 2023. Fast-detectgpt: Efficient zero-shot detection of machine-generated text via conditional probability curvature. arXiv preprint arXiv:2310.05130.
  3. Amrita Bhattacharjee, Tharindu Kumarage, Raha Moraffah, and Huan Liu. 2023. Conda: Contrastive domain adaptation for ai-generated text detection. In Proceedings of the 13th International Joint Conference on Natural Language Processing and the 3rd Conference of the Asia-Pacific Chapter of the Associa… find
  4. Yehonatan Bitton, Elad Bitton, and Shai Nisan. 2025. Detecting stylistic fingerprints of large language models. arXiv preprint arXiv:2503.01659. — [arXiv:2503.01659](https://arxiv.org/abs/2503.01659) find
  5. Sebastian Gehrmann, Hendrik Strobelt, and Alexander M. Rush. 2019. Gltr: Statistical detection and visualization of generated text. In Annual Meeting of the Association for Computational Linguistics.
  6. Xinlei He, Xinyue Shen, Zeyuan Chen, Michael Backes, and Yang Zhang. 2023. Mgtbench: Benchmarking machine-generated text detection. Proceedings of the 2024 on ACM SIGSAC Conference on Computer and Communications Security. find
  7. Mihailo korić, Ranka Stanković, Milica Ikonić Neić, Joanna Byszuk, and Maciej Eder. 2022. Parallel stylometric document embeddings with deep learning based language models in literary authorship attribution. Mathematics. find
  8. Tharindu Kumarage, Garima Agrawal, Paras Sheth, Raha Moraffah, Amanat Chadha, Joshua Garland, and Huan Liu. 2024. A survey of ai-generated text forensic systems: Detection, attribution, and characterization. arXiv preprint arXiv:2403.01152. — [arXiv:2403.01152](https://arxiv.org/abs/2403.01152) find
  9. Tharindu Kumarage and Huan Liu. 2023a. Neural authorship attribution: Stylometric analysis on large language models. 2023 International Conference on Cyber-Enabled Distributed Computing and Knowledge Discovery (CyberC), pages 51–54. find
  10. Tharindu Kumarage and Huan Liu. 2023b. Neural authorship attribution: Stylometric analysis on large language models. 2023 International Conference on Cyber-Enabled Distributed Computing and Knowledge Discovery (CyberC), pages 51–54. find
  11. Lucio La Cava and Andrea Tagarelli. 2025. Openturingbench: an open-model-based benchmark and framework for machine-generated text detection and attribution. In Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing, pages 26666–26682. find
  12. Dongfang Li, Zetian Sun, Xinshuo Hu, Zhenyu Liu, Ziyang Chen, Baotian Hu, Aiguo Wu, and Min Zhang. 2023a. A survey of large language models attribution. arXiv preprint arXiv:2303.11666. — [arXiv:2303.11666](https://arxiv.org/abs/2303.11666) find
  13. Yafu Li, Qintong Li, Leyang Cui, Wei Bi, Zhilin Wang, Longyue Wang, Linyi Yang, Shuming Shi, and Yue Zhang. 2023b. Mage: Machine-generated text detection in the wild. In Annual Meeting of the Association for Computational Linguistics.
  14. Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. 2019. Roberta: A robustly optimized bert pretraining approach. arXiv preprint arXiv:1907.11692.
  15. Eric Mitchell, Yoonho Lee, Alexander Khazatsky, Christopher D. Manning, and Chelsea Finn. 2023. Detectgpt: Zero-shot machine-generated text detection using probability curvature. In International Conference on Machine Learning.
  16. Pablo Francisco Posadas-Durán, Germán Ríos-Toledo, Erick Velázquez-Lozada, A De Jesús Osuna-Coutiño, Madaín Pérez-Patricio, and Fernando Pech May. 2025. Learning the style via mixed sn-grams: An evaluation in authorship attribution. AI, 6(5):104. find
  17. Grigori Sidorov, Francisco Velasquez, Efstathios Stamatatos, Alexander Gelbukh, and Liliana Chanona-Hernández. 2014. Syntactic n-grams as machine learning features for natural language processing. Expert Systems with Applications, 41:853–860. find
  18. Irene Solaiman, Miles Brundage, Jack Clark, Amanda Askell, Ariel Herbert-Voss, Jeff Wu, Alec Radford, and Jasmine Wang. 2019. Release strategies and the social impacts of language models. arXiv preprint arXiv:1908.09203.
  19. Jinyan Su, Terry Yue Zhuo, Di Wang, and Preslav Nakov. 2023. Detectllm: Leveraging log rank information for zero-shot detection of machine-generated text. In Conference on Empirical Methods in Natural Language Processing.
  20. Ruixiang Tang, Yu-Neng Chuang, and Xia Hu. 2023. The science of detecting llm-generated text. Communications of the ACM, 67:50 – 59. find
  21. Saranya Venkatraman, Adaku Uchendu, and Dongwon Lee. 2023. Gpt-who: An information density-based machine-generated text detector. In NAACL-HLT. find
  22. Junchao Wu, Shu Yang, Runzhe Zhan, Yulin Yuan, Lidia S. Chao, and Derek F. Wong. 2025a. A survey on llm-generated text detection: Necessity, methods, and future directions. Computational Linguistics, 51:275–338.
  23. Zehao Wu, Yanjie Zhao, and Haoyu Wang. 2025b. Gradient-based model fingerprinting for llm similarity detection and family classification. arXiv preprint arXiv:2506.01631. — [arXiv:2506.01631](https://arxiv.org/abs/2506.01631) find
  24. Xiao Yu, Kejiang Chen, Qi Yang, Weiming Zhang, and Neng H. Yu. 2024. Text fluoroscopy: Detecting llm-generated text through intrinsic features. In Conference on Empirical Methods in Natural Language Processing. find

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/lu-2026-synergizing-stylometrics-semantics-dual-path/