Elamur Research⌕
☰
home / papers / 2026 / UMPIRE: Unveiling LLM-generated Posts via Redundant Expressions

UMPIRE: Unveiling LLM-generated Posts via Redundant Expressions

Paper · 2026 · Xiaoquan Yi, Haixing Wu, Haozhao Wang et al.

Summary

Social media is increasingly full of image–text posts written by large (multimodal) language models. Detectors that rely on text statistics or visual artifacts tend to fail on short captions, paraphrased text and compressed images. UMPIRE builds on a behavioral observation: people treat the attached picture as given and use the caption to add context, emotion or backstory, while models mostly restate what the image shows.

The authors support this on SoMe, where cross-modal redundancy averages roughly 0.30 for human posts and 0.80 for generated ones. Their method splits the text embedding into two parts: one aligned with the visual embedding (redundant) and one orthogonal to it (complementary). An adaptive gate weighs the two parts. A latent contrastive redundancy regularization loss (L_LCRR) then pushes generated posts toward high redundancy and human posts toward low redundancy.

On TwiBot-22, Fakeddit and Hateful Memes, UMPIRE outperforms unimodal, multimodal and zero-shot LLM baselines. It also transfers well to unseen domains. The detection cue matters because it comes from how models are trained, not from fragile surface traces.

Key points

  • Annotation of visual dependence: 89.3% of objects mentioned in LLM-generated captions are visible in the image, versus 28.4% for human captions. Emotions or opinions appear in 68.7% of human captions but only 12.4% of LLM-generated ones.
  • In-domain accuracy and F1:
    • TwiBot-22: 84.82% accuracy and 88.26% F1. The F1 is 5.38 points above SIDA, which matches the abstract's "up to 5.38%", although the abstract labels this an accuracy gain.
    • Fakeddit: 90.05% accuracy, 5.34 points above SIDA.
    • Hateful Memes: 87.92% accuracy.
  • AUROC: TwiBot-22 reaches 92.97%, Fakeddit 98.39%, and Hateful Memes 95.67%.
  • Cross-domain test: models were trained on NewsCLIPpings and evaluated without fine-tuning. On TwiBot-22, UMPIRE reached 82.78% accuracy against 70.29% for SIDA.
  • Ablation: each component adds to the result: orthogonal decomposition beats plain concatenation, and the adaptive gate beats static fusion. Adding L_LCRR raises accuracy by about 1.8%. It also widens the "Geo-Gap" between human and LLM embeddings, from 0.29 to 0.45 on TwiBot-22 and from 0.33 to 0.48 on Fakeddit.

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

🎓Huazhong University of Science and TechnologyWuhan · China59 works · 8,310 citations🎓Wuhan University of TechnologyWuhan · China6 works · 261 citations

 Fields, topics and concepts

📚 Shelves 4The field of science this work belongs to (OpenAlex fields).Computer Science · 18,614 worksArtificial Intelligence · 17,965 worksArts and Humanities · 17,450 worksLanguage and Linguistics · 17,296 works

🧭 Topics 1Research subjects and languages the work is about.Language models & pretraining · 3,894 works

Cite

Cite as Xiaoquan Yi, Haixing Wu, Haozhao Wang, Yichen Li, Yuhua Li, Rui Zhang and Ruixuan Li. 2026. UMPIRE: Unveiling LLM-generated Posts via Redundant Expressions. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers).

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.18653/v1/2026.acl-long.1427

@inproceedings{yi2026umpire,
  title = {{UMPIRE: Unveiling LLM-generated Posts via Redundant Expressions}},
  author = {Xiaoquan Yi, Haixing Wu, Haozhao Wang, Yichen Li, Yuhua Li, Rui Zhang and Ruixuan Li},
  booktitle = {Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)},
  year = {2026},
  doi = {10.18653/v1/2026.acl-long.1427},
  url = {https://doi.org/10.18653/v1/2026.acl-long.1427}
}

References

47 references, 10 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, and 1 others. 2023. Gpt-4 technical report. arXiv preprint arXiv:2303.08774.
  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. In The Twelfth International Conference on Learning Representations.
  3. Yutian Chen, Hao Kang, Vivian Zhai, Liangze Li, Rita Singh, and Bhiksha Raj. 2023. Token prediction as implicit classification to identify llm-generated text. arXiv preprint arXiv:2311.08723. find
  4. Colin Cherry. 1966. On human communication. find
  5. Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, and 1 others. 2024. The llama 3 herd of models. arXiv e-prints, pages arXiv–2407.
  6. Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, and 1 others. 2024. Scaling rectified flow transformers for high-resolution image synthesis. In Forty-first international conference on machine learning. find
  7. Kanika Goswami, Puneet Mathur, Ryan Rossi, and Franck Dernoncourt. 2025. Plotgen: Multi-agent llm-based scientific data visualization via multimodal feedback. arXiv preprint arXiv:2502.00988. — [arXiv:2502.00988](https://arxiv.org/abs/2502.00988) find
  8. Xun Guo, Yongxin He, Shan Zhang, Ting Zhang, Wan-quan Feng, Haibin Huang, and Chongyang Ma. 2024. Detective: Detecting ai-generated text via multi-level contrastive learning. Advances in Neural Information Processing Systems, 37:88320–88347. find
  9. Robert L Heath and Jennings Bryant. 2013. Human communication theory and research: Concepts, contexts, and challenges. Routledge. find
  10. Runsheng Huang, Liam Dugan, Yue Yang, and Chris Callison-Burch. 2024. MiRAGeNews: Multimodal realistic AI-generated news detection. In Findings of the Association for Computational Linguistics: EMNLP 2024, pages 16436–16448, Miami, Florida, USA. Association for Computational Linguistics. find
  11. Zhenglin Huang, Jinwei Hu, Xiangtai Li, Yiwei He, Xingyu Zhao, Bei Peng, Baoyuan Wu, Xiaowei Huang, and Guangliang Cheng. 2025. Sida: Social media image deepfake detection, localization and explanation with large multimodal model. In Proceedings of the Computer Vision and Pattern Recognition Confere… find
  12. Oana Ignat, Gayathri Ganesh Lakshmy, and Rada Mihalcea. 2025. Inspaired: Cross-cultural inspiration detection and analysis in real and llm-generated social media data. In Proceedings of the 3rd Workshop on Cross-Cultural Considerations in NLP (C3NLP 2025), pages 35–49. find
  13. Minsuk Jang, Hyeonseo Jeong, Minseok Son, and Changick Kim. 2025. Cinemae: Leveraging frozen masked autoencoders for cross-generator ai image detection. arXiv preprint arXiv:2511.06325. — [arXiv:2511.06325](https://arxiv.org/abs/2511.06325) find
  14. Douwe Kiela, Hamed Firooz, Aravind Mohan, Vedanuj Goswami, Amanpreet Singh, Pratik Ringshia, and Davide Testuggine. 2020. The hateful memes challenge: Detecting hate speech in multimodal memes. In Advances in Neural Information Processing Systems, volume 33, pages 2611–2624. Curran Associates, Inc.
  15. Clyde Kluckhohn. 1950. Human behavior and the principle of least effort. find
  16. Shangbin Feng, Zhaoxuan Tan, Herun Wan, Ningnan Wang, Zilong Chen, Binchi Zhang, Qinghua Zheng, Wenqian Zhang, Zhenyu Lei, Shujie Yang, and 1 others. 2022. Twibot-22: Towards graph-based twitter bot detection. Advances in Neural Information Processing Systems, 35:35254–35269. find
  17. An Lao, Qi Zhang, Chongyang Shi, Longbing Cao, Kun Yi, Liang Hu, and Duoqian Miao. 2024. Frequency spectrum is more effective for multimodal representation and fusion: A multimodal spectrum rumor detector. In Proceedings of the AAAI conference on artificial intelligence, volume 38, pages 18426–18434… find
  18. Team GLM, Aohan Zeng, Bin Xu, Bowen Wang, Chenhui Zhang, Da Yin, Dan Zhang, Diego Rojas, Guanyu Feng, Hanlin Zhao, and 1 others. 2024. Chatglm: A family of large language models from glm-130b to glm-4 all tools. arXiv preprint arXiv:2406.12793.
  19. Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. 2022. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In International conference on machine learning, pages 12888–12900. PMLR.
  20. Wei Li, Jiawen Deng, Jiali You, Yuanyuan He, Yan Zhuang, and Fuji Ren. 2025. Ets-mm: A multi-modal social bot detection model based on enhanced textual semantic representation. In Proceedings of the ACM on Web Conference 2025, pages 4160–4170. find
  21. Kai Nakamura, Sharon Levy, and William Yang Wang. 2020. Fakeddit: A new multimodal benchmark dataset for fine-grained fake news detection. In Proceedings of the twelfth language resources and evaluation conference, pages 6149–6157. find
  22. Zijie Lin, Yiqing Shen, Qilin Cai, He Sun, Jinrui Zhou, and Mingjun Xiao. 2025. Autop2c: An llm-based agent framework for code repository generation from multimodal content in academic papers. arXiv preprint arXiv:2504.20115. — [arXiv:2504.20115](https://arxiv.org/abs/2504.20115) find
  23. Nicolò Pagan, Petter Törnberg, Christopher Bail, An-csa Hannak, and Christopher Barrie. 2025. Can llms imitate social media dialogue? techniques for calibration and bert-based turing-test. In First Workshop on Social Simulation with LLMs. find
  24. Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. 2023a. Visual instruction tuning. Advances in neural information processing systems, 36:34892– 34916.
  25. Xiaoming Liu, Zhaohan Zhang, Yichen Wang, Hang Pu, Yu Lan, and Chao Shen. 2023b. CoCo: Coherence-Enhanced Machine-Generated Text Detection Under Data Limitation With Contrastive Learning. arXiv preprint. ArXiv:2212.10341 [cs]. — [arXiv:2212.10341](https://arxiv.org/abs/2212.10341) find
  26. Yuhan Liu, Zhaoxuan Tan, Heng Wang, Shangbin Feng, Qinghua Zheng, and Minnan Luo. 2023c. Botmoe: Twitter bot detection with community-aware mixtures of modal-specific experts. In Proceedings of the 46th International ACM SIGIR Conference on Research and Development in Information Retrieval, pages 48… find
  27. Zhuoran Lu, Gionnieve Lim, and Ming Yin. 2025. Understanding the effects of large language model (llm)-driven adversarial social influences in online information spread. In Proceedings of the Extended Abstracts of the CHI Conference on Human Factors in Computing Systems, pages 1–7. find
  28. Grace Luo, Trevor Darrell, and Anna Rohrbach. 2021. Newsclippings: Automatic generation of out-of-context multimodal media. arXiv preprint arXiv:2104.05893. find
  29. Dominik Macko, Aashish Anantha Ramakrishnan, Jason Samuel Lucas, Robert Moro, Ivan Srba, Adaku Uchendu, and Dongwon Lee. 2025. Beyond speculation: Measuring the growing presence of llm-generated texts in multilingual disinformation. arXiv preprint arXiv:2503.23242. find
  30. Elyas Meguellati, Assaad Zeghina, Shazia Sadiq, and Gianluca Demartini. 2025. Llm-based semantic augmentation for harmful content detection. In Proceedings of the International AAAI Conference on Web and Social Media, volume 19, pages 1190–1209. find
  31. Eric Mitchell, Yoonho Lee, Alexander Khazatsky, Christopher D. Manning, and Chelsea Finn. 2023. Detectgpt: zero-shot machine-generated text detection using probability curvature. In Proceedings of the 40th International Conference on Machine Learning, ICML’23. JMLR.org.
  32. Shaghayegh Najari, Mostafa Salehi, and Reza Farahbakhsh. 2022. Ganbot: a gan-based framework for social bot detection. Social Network Analysis and Mining, 12(1):4. find
  33. Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, and 1 others. 2021. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748–8…
  34. Amine Sallah, Said Agoujil, Mudasir Ahmad Wani, Mohamed Hammad, Yassine Maleh, Ahmed A Abd El-Latif, and 1 others. 2024. Fine-tuned understanding: Enhancing social bot detection with transformer-based classification. IEEE Access, 12:118250– 118269. find
  35. S Sorensen. Claude 4.0, with manus ai (2025).". Consciousness Assessment in Large Language Models: A Comparative Analysis of Response Patterns to Recursive Self-Examination and Temporal Discontinuity." Human-AI Collaborative Research Series, 1(1). find
  36. Zhen Sun, Zongmin Zhang, Xinyue Shen, Ziyi Zhang, Yule Liu, Michael Backes, Yang Zhang, and Xinlei He. 2025. Are we in the ai-generated text world already? quantifying and monitoring aigt on social media. In Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Vol… find
  37. Sahar Tahmasebi, Eric Müller-Budack, and Ralph Ewerth. 2024. Multimodal misinformation detection using large vision-language models. In Proceedings of the 33rd ACM International Conference on Information and Knowledge Management, pages 2189–2199. find
  38. Gemma Team, Aishwarya Kamath, Johan Ferret, Shreya Pathak, Nino Vieillard, Ramona Merhej, Sarah Perrin, Tatiana Matejovicova, Alexandre Ramé, Morgane Rivière, and 1 others. 2025. Gemma 3 technical report. arXiv preprint arXiv:2503.19786. find
  39. Irina Tolstykh, Aleksandra Tsybina, Sergey Yakubson, Aleksandr Gordeev, Vladimir Dokholyan, and Maksim Kuprashevich. 2024. Gigacheck: Detecting llm-generated content. arXiv preprint arXiv:2410.23728. — [arXiv:2410.23728](https://arxiv.org/abs/2410.23728) find
  40. Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, and 1 others. 2024a. Qwen2-vl: Enhancing vision-language model’s perception of the world at any resolution. arXiv preprint arXiv:2409.12191.
  41. Qian Wang, Chen Li, Yuchen Luo, Hefei Ling, Shijuan Huang, Ruoxi Jia, and Ning Yu. 2025. Detecting adversarial data using perturbation forgery. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 13917–13926. find
  42. Rongsheng Wang, Haoming Chen, Ruizhe Zhou, Han Ma, Yaofei Duan, Yanlan Kang, Songhua Yang, Baoyu Fan, and Tao Tan. 2024b. Llm-detector: Improving ai-generated chinese text detection with open-source llm instruction tuning. arXiv preprint arXiv:2402.01158. — [arXiv:2402.01158](https://arxiv.org/abs/2402.01158) find
  43. Dizhan Xue, Jing Cui, Shengsheng Qian, Chuanrui Hu, and Changsheng Xu. 2025. Some: A realistic benchmark for llm-based social media agents. arXiv preprint arXiv:2512.14720. — [arXiv:2512.14720](https://arxiv.org/abs/2512.14720) find
  44. Hong Zhang, Mohamed Meyer Kana Kone, Xiao-Qian Ma, and Nan-Run Zhou. 2025. Frequency-domain attention-guided adaptive robust watermarking model. Journal of the Franklin Institute, 362(3):107511. find
  45. Chuqing Zhao and Yisong Chen. 2025. Llm-powered topic modeling for discovering public mental health trends in social media. In IFIP International Conference on Artificial Intelligence Applications and Innovations, pages 119–132. Springer. find
  46. Yuying Zhao, Yu Wang, Xueqi Cheng, Anne Marie Tum-lin, Yunchao Liu, Damin Xia, Meng Jiang, and Tyler Derr. 2025. Amplifying your social media presence: Personalized influential content generation with llms. arXiv preprint arXiv:2505.01698. — [arXiv:2505.01698](https://arxiv.org/abs/2505.01698) find
  47. George Kingsley Zipf. 2016. Human behavior and the principle of least effort: An introduction to human ecology. Ravenio books. find

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/yi-2026-umpire-unveiling-llm-generated-posts/