Elamur Research⌕
☰
home / papers / 2026 / U-RoCX: An xLSTM based Approach to AI-Generated Urdu Text Detection

U-RoCX: An xLSTM based Approach to AI-Generated Urdu Text Detection

Paper · 2026 · Rabee Adel Al-Qasem

Summary

The paper addresses telling human-written Urdu text apart from LLM output, a hard problem in a low-resource language. It is a system submission to Task 2 of the AbjadGenEval shared task. The dataset has 6,800 samples: 4,800 for training and 2,000 for testing. Human texts come from news sites, and machine texts come from several LLMs.

The author proposes U-RoCX, a hybrid classifier with four parts:

  • a frozen XLM-RoBERTa encoder that supplies token embeddings,
  • a small 1D CNN that captures trigram-level local patterns,
  • a short stack of Extended LSTM (xLSTM) blocks for sequence modelling,
  • a compact dense head that makes the binary decision.

Before training, the texts are cleaned with UrduHack. This covers normalization and stopword removal, followed by a rule-based stemmer that strips affixes.

On the official test set, the system scores about 88% on both macro F1 and balanced accuracy, and the scores are similar across the two classes. The author presents this as the first use of xLSTM for binary Urdu text classification. The paper suggests that a lightweight recurrent component on top of a frozen multilingual encoder can be competitive for detection tasks in low-resource languages. No baseline comparisons are reported.

Key points

  • Test results: accuracy 88.02%, balanced accuracy 88.03%, precision 87.47%, recall 88.61%, macro F1 88.04%.
  • Parameters: the model has roughly 278M parameters in total, but only 399,114 are trained, because XLM-RoBERTa stays frozen. The author reports that freezing improved accuracy and shortened training.
  • Architecture details:
    • 768-d embeddings are projected down to 128 channels.
    • Conv1D uses 128 filters of width 3, followed by max-pooling of size 2.
    • Two xLSTM blocks (an sLSTM, then an mLSTM) with 4 heads.
    • A 256-unit ReLU layer with dropout 0.3.
  • Training setup: cross-entropy loss, the AdamW optimizer and a OneCycleLR schedule, run on a single NVIDIA Tesla T4. Maximum sequence length is set at the 99th percentile of training-text lengths.
  • Preprocessing: regex removal of URLs, hashtags and emojis; Unicode and whitespace normalization; stopword filtering; and an affix stemmer that always keeps a stem of at least two characters.

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

 Fields, topics and concepts

📚 Shelves 4The 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 works

Cite

Cite as Rabee Adel Al-Qasem. 2026. U-RoCX: An xLSTM based Approach to AI-Generated Urdu Text Detection. Proceedings of the 2nd Workshop on NLP for Languages Using Arabic Script.

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.18653/v1/2026.abjadnlp-1.53

@inproceedings{alqasem2026u,
  title = {{U-RoCX: An xLSTM based Approach to AI-Generated Urdu Text Detection}},
  author = {Rabee Adel Al-Qasem},
  booktitle = {Proceedings of the 2nd Workshop on NLP for Languages Using Arabic Script},
  year = {2026},
  doi = {10.18653/v1/2026.abjadnlp-1.53},
  url = {https://doi.org/10.18653/v1/2026.abjadnlp-1.53}
}

References

16 references, 2 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Shadi Abudalfa, Saad Ezzini, Ahmed Abdelali, Hamza Alami, Abdessamad Benlahbib, Salmane Chafik, Mo El-Haj, Abdelkader El Mahdaouy, Mustafa Jarrar, Salima Lamsiyah, and Hamzah Luqman. 2025. The AraGenEval shared task on Arabic authorship style transfer and AI generated text detection. In Proceedings…
  2. Naeem Ahmed, Rashid Amin, Hamza Aldabbas, Muhammad Saeed, Muhammad Bilal, and Houbing Song. 2024. A novel approach for sentiment analysis of a low resource language using deep learning models. ACM Transactions on Asian and Low-Resource Language Information Processing. find
  3. Alexis Conneau, Kartikay Khandelwal, Naman Goyal, Vishrav Chaudhary, Guillaume Wenzek, Francisco Guzmán, Edouard Grave, Myle Ott, Luke Zettlemoyer, and Veselin Stoyanov. 2020. Unsupervised cross-lingual representation learning at scale. In Proceedings of the 58th annual meeting of the association fo… find
  4. Hamed Abdelreheem Ead. 2024. Exploring the impact of artificial intelligence on academic writing: Perspectives of ph. d. students in the faculty of science at cairo university. find
  5. Saad Ezzini, Irfan Ahmed, Salmane Chafik, Shadi Abudalfa, Mo El-Haj, Ahmed Abdelali, Mustafa Jarrar, Nadir Durrani, Hassan Sajjad, and Farah Adeeba. 2026. Abjadgeneval: Abjad ai generated text detection shared task for languages using arabic script at abjadnlp 2026. In Proceedings of the 2nd Worksho…
  6. Dana Hasan, Amal Nazzal, and Sulafa Zidani. 2025. Beating algorithmic discrimination: Maneuvering digital surveillance to indigenize the narrative. International Journal of Communication, 19:23. find
  7. Salima Lamsiyah, Saad Ezzini, Abdelkader El Mahdaouy, Hamza Alami, Abdessamad Benlahbib, Samir El amrany, Salmane Chafik, and Hicham Hammouchi. 2025. M-DAIGT: A Shared Task on Multi-Domain Detection of AI-Generated Text. In Proceedings of the Shared Task on Multi-Domain Detection of AI-Generated Tex… find
  8. Shih-Yang Liu, Xin Dong, Ximing Lu, Shizhe Diao, Peter Belcak, Mingjie Liu, Min-Hung Chen, Hongxu Yin, Yu-Chiang Frank Wang, Kwang-Ting Cheng, and 1 others. 2026. Gdpo: Group reward-decoupled normalization policy optimization for multi-reward rl optimization. arXiv preprint arXiv:2601.05242. — [arXiv:2601.05242](https://arxiv.org/abs/2601.05242) find
  9. Ilya Loshchilov and Frank Hutter. 2019. Decoupled weight decay regularization. In International Conference on Learning Representations. find
  10. Sabri Odeh and Emad Natsheh. 2025. Toward better solar pv panel fault detection: A multi-ml approach for series and parallel hotspot analysis. Engineering Research Express. find
  11. Husnain Saleem, Muhammad Javed, and Junaid Khan. 2025. Hate speech identification in formal and informal social media text using roberta-base and xlm-roberta-base models. BRAIN: Broad Research in Artificial Intelligence & Neuroscience, 16(4). find
  12. Leslie N Smith. 2018. A disciplined approach to neural network hyper-parameters: Part 1–learning rate, batch size, momentum, and weight decay. arXiv preprint arXiv:1803.09820. — [arXiv:1803.09820](https://arxiv.org/abs/1803.09820) find
  13. Zhenda Xie, Yixuan Wei, Huanqi Cao, Chenggang Zhao, Chengqi Deng, Jiashi Li, Damai Dai, Huazuo Gao, Jiang Chang, Liang Zhao, and 1 others. 2025. mhc: Manifold-constrained hyper-connections. arXiv preprint arXiv:2512.24880. — [arXiv:2512.24880](https://arxiv.org/abs/2512.24880) find
  14. Lirong Yin, Lei Wang, Zhuohang Cai, Siyu Lu, Ruiyang Wang, Ahmed AlSanad, Salman A AlQahtani, Xiaobing Chen, Zhengtong Yin, Xiaolu Li, and 1 others. 2024. Dpal-bert: A faster and lighter question answering model. CMES-Computer Modeling in Engineering & Sciences, 141(1). find
  15. Yifan Zhang, Yifeng Liu, Mengdi Wang, and Quanquan Gu. 2026. Deep delta learning. arXiv preprint arXiv:2601.00417. — [arXiv:2601.00417](https://arxiv.org/abs/2601.00417) find
  16. Wenfeng Zheng, Siyu Lu, Zhuohang Cai, Ruiyang Wang, Lei Wang, and Lirong Yin. 2023. Pal-bert: an improved question answering model. Computer Modeling in Engineering & Sciences, 10. find

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/qasem-2026-u-rocx-xlstm-based-approach/