Elamur Research⌕
☰
home / papers / 2026 / Islamic Large Language Models: From Knowledge Acquisition to Trustworthy and Hallucination-Resistant AI

Islamic Large Language Models: From Knowledge Acquisition to Trustworthy and Hallucination-Resistant AI

Paper · 2026 · Mohammed Amine Mouhoub

Summary

This survey asks what it would take for large language models to answer questions about Islam reliably. In this domain a fabricated Hadith, a wrong verse number or an invented ruling can mislead people about religious practice. The author traces a line from classical Arabic NLP tools (MADAMIRA, Farasa, CAMeL Tools), through Arabic pretrained encoders (AraBERT, ARBERT/MARBERT, CAMeLBERT), to Arabic-centric LLMs (Jais, ALLaM, AceGPT, Fanar, AraLLaMA). It then reviews Islamic resources and benchmarks: Qur'anic QA shared tasks, IslamicMMLU, IslamicEval, QIAS and MAWARITH. The central claim is that writing good Arabic and being trustworthy on Islamic content are separate abilities. Reliability comes from the surrounding system rather than model size: curated sources, retrieval, citation checking, school-aware (madhhab) reasoning, symbolic calculators for structured law, and scholar oversight. The paper offers a trustworthiness framework and a matching set of evaluation families. It ends with an agenda of open challenges and future directions. This matters for any high-stakes domain where provenance and legitimate disagreement belong to the answer itself.

Key points

  • Table 1 lists six requirements that set Islamic AI apart from open-domain QA:
    • source hierarchy
    • citation precision
    • bridging to Classical Arabic
    • juristic disagreement
    • rule-based legal reasoning
    • a safety boundary against autonomous Fatwa issuance
  • Qur'an QA 2022 framed Qur'anic QA as machine reading comprehension. Qur'an QA 2023 split it into passage retrieval and reading comprehension, which showed how much answer extraction depends on retrieval quality.
  • QIAS 2026 uses the MIR-E metric for inheritance problems. MIR-E scores four stages separately:
    • heirs and blocking
    • share assignment
    • adjustment handling
    • final distribution
  • CAMeLBERT released separate variants for MSA, dialectal, Classical and mixed Arabic. Fanar reports an Islamic RAG and attribution service, and Fanar-Sadiq is cited as an agentic Islamic QA design.
  • Table 7 maps metric families to their best use:
    • accuracy for MCQ
    • EM/F1 for extractive QA
    • Recall@k/MRR/nDCG for retrieval
    • stage-wise scores for legal pipelines
    • span F1 for hallucination localization
    • citation validity for referenced answers
    • expert rating for Fatwa quality

Interesting ideas

  • I1. Scoring a multi-step reasoning chain stage by stage, as MIR-E does, separates an interpretation error from an arithmetic error and shows where a mistake propagated.
  • I2. Neuro-symbolic split: the language model only parses the input and explains the result, while a deterministic module enforces rules and does exact arithmetic.
  • I3. Every claim carries its provenance and authority status, so a direct source quote is kept apart from an inference drawn from it.
  • I4. Legitimate disagreement is kept as several labelled positions instead of being collapsed into one "correct" answer, and benchmarks must tell a false answer from a valid minority view.
  • I5. Hallucination is evaluated at span level, because the damaging error is often a short fragment such as a wrong number or an invented name inside an otherwise fine answer.

For Mova

How these ideas could serve Mova, our small open language model — its files, stages and tests:

  • I1 (answer): Maths word-problem solving could be graded separately on extracted quantities, chosen operation and final number, so failures point to the responsible stage.
  • I2 (answer): This matches Mova's philosophy; word problems should hand computation to a symbolic solver over the world state rather than letting any learned component produce numbers.
  • I3 (world): Facts in the world of a text could be tagged as stated in the text, inferred by rule, or taken from compiled knowledge, so answers can show which kind of evidence they rest on.
  • I4 (frames): When parses or implicature readings are genuinely ambiguous, Mova could keep alternative frames with their assumptions rather than forcing one reading early.
  • I5 (back): The text → graph → text compressor could be checked with span-level alignment, flagging any generated fragment that has no node in the source graph.

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

🎓Université Paris Dauphine-PSLParis · France4 works · 16,983 citations

 Fields, topics and concepts

📚 Shelves 6The 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 worksSocial Sciences · 1,772 worksPolitical Science and International Relations · 182 works

🧭 Topics 2Research subjects and languages the work is about.Language models & pretraining · 3,320 worksSemitic languages · 2,908 works

Cite

Cite as Mohammed Amine Mouhoub. 2026. Islamic Large Language Models: From Knowledge Acquisition to Trustworthy and Hallucination-Resistant AI. arXiv:2606.16629.

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.48550/arxiv.2606.16629 · arXiv 2606.16629

@misc{mouhoub2026islamic,
  title = {{Islamic Large Language Models: From Knowledge Acquisition to Trustworthy and Hallucination-Resistant AI}},
  author = {Mohammed Amine Mouhoub},
  year = {2026},
  doi = {10.48550/arxiv.2606.16629},
  eprint = {2606.16629},
  archivePrefix = {arXiv},
  url = {https://doi.org/10.48550/arxiv.2606.16629}
}

References

39 references, 29 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Ummar Abbas, Mourad Ouzzani, Mohamed Y. Eltabakh, Omar Sinan, Gagan Bhatia, Hamdy Mubarak, Majd Hawasly, Mohammed Qusay Hashim, Kareem Darwish, and Firoj Alam. 2026. Fanar-Sadiq: A Multi-Agent Architecture for Grounded Islamic QA. arXiv preprint arXiv:2603.08501. ↗
  2. Ahmed Abdelali, Kareem Darwish, Nadir Durrani, and Hamdy Mubarak. 2016. Farasa: A Fast and Furious Segmenter for Arabic. In Proceedings of NAACL Demonstrations, pages 11–16.
  3. Ali Abdelaal, Mohammed Nader Al Haffar, Mahmoud Fawzi, and Walid Magdy. 2026. IslamicMMLU: A Benchmark for Evaluating LLMs on Islamic Knowledge. arXiv preprint arXiv:2603.23750. ↗
  4. Muhammad Abdul-Mageed, AbdelRahim Elmadany, and El Moatez Billah Nagoudi. 2021. AR-BERT and MARBERT: Deep Bidirectional Transformers for Arabic. In Proceedings of ACL, pages 7088–7105. find
  5. Fakhraddin Alwajih, Abdellah El Mekki, Hamdy Mubarak, Majd Hawasly, Abubakr Mohamed, and Muhammad Abdul-Mageed. 2025. PalmX 2025: The First Shared Task on Benchmarking LLMs on Arabic and Islamic Culture. arXiv preprint arXiv:2509.02550. ↗
  6. Wissam Antoun, Fady Baly, and Hazem Hajj. 2020. AraBERT: Transformer-based Model for Arabic Language Understanding. In Proceedings of OS-ACT, pages 9–15. ↗
  7. M. Saiful Bari, Yazeed Alnumay, Norah A. Alzahrani, Nouf M. Alotaibi, Hisham A. Alyahya, and others. 2024. ALLaM: Large Language Models for Arabic and English. arXiv preprint arXiv:2407.15390. ↗
  8. Gael de Francony, Victor Guichard, Praveen Joshi, Haithem Afli, and Abdessalam Bouchekif. 2019. Hierarchical Deep Learning for Arabic Dialect Identification. In Proceedings of WANLP, pages 249–253.
  9. Abdessalam Bouchekif, Samer Rashwani, Heba Sbahi, Shahd Gaben, Mutaz Al-Khatib, and Mohammed Ghaly. 2025a. Assessing Large Language Models on Islamic Legal Reasoning: Evidence from Inheritance Law Evaluation. In Proceedings of ArabicNLP.
  10. Abdessalam Bouchekif, Samer Rashwani, Mohammed Ghaly, Mutaz Al-Khatib, Shahd Gaben, Heba Sbahi, Emad Mohamed, Wajdi Zaghouani, and Aiman Erbad. 2025b. QIAS 2025: Overview of the Shared Task on Islamic Inheritance Reasoning and Knowledge Assessment. In Proceedings of ArabicNLP Shared Tasks.
  11. Abdessalam Bouchekif, Somaya Eltanbouly, Samer Rashwani, Shahd Gaben, Mutaz Al-Khatib, Heba Sbahi, Emad Mohamed, and Mohammed Ghaly. 2026a. QIAS 2026: Overview of the Shared Task on Islamic Inheritance Reasoning. arXiv preprint arXiv:2606.13756. ↗
  12. Abdessalam Bouchekif, Shahd Gaben, Samer Rashwani, Somaya Eltanbouly, Mutaz Al-Khatib, Heba Sbahi, Mohammed Ghaly, and Emad Mohamed. 2026b. MAWARITH: A Dataset and Benchmark for Legal Inheritance Reasoning with LLMs. arXiv preprint arXiv:2603.07539. ↗
  13. Tom B. Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, and others. 2020. Language Models are Few-Shot Learners. In Advances in Neural Information Processing Systems.
  14. Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding. In Proceedings of NAACL, pages 4171– 4186.
  15. Kais Dukes and Nizar Habash. 2010. Morphological Annotation of Quranic Arabic. In Proceedings of LREC, pages 2530–2536.
  16. Ezieddin Elmahjub, Junaid Qadir, Abdullah Mushtaq, Rafay Naeem, Ibrahim Ghaznavi, and Waleed Iqbal. 2026. IslamicLegalBench: Evaluating LLMs Knowledge and Reasoning of Islamic Law Across 1,200 Years of Islamic Pluralist Legal Traditions. arXiv preprint arXiv:2602.21226. ↗
  17. Fanar Team, Ummar Abbas, Mohammad Shahmeer Ahmad, Firoj Alam, Enes Altinisik, Ehsannedin Asgari, and others. 2025. Fanar: An Arabic-Centric Multimodal Generative AI Platform. arXiv preprint arXiv:2501.13944. ↗
  18. Dan Hendrycks, Collin Burns, Steven Basart, Andy Zou, Mantas Mazeika, Dawn Song, and Jacob Steinhardt. 2021. Measuring Massive Multitask Language Understanding. In Proceedings of ICLR. ↗
  19. Huang Huang and others. 2024. AceGPT, Localizing Large Language Models in Arabic. In Proceedings of NAACL.
  20. Go Inoue, Bashar Alhafni, Nurpeiis Baimukan, Houda Bouamor, and Nizar Habash. 2021. The Interplay of Variant, Size, and Task Type in Arabic Pre-trained Language Models. In Proceedings of WANLP, pages 92–104.
  21. Vladimir Karpukhin, Barlas Oguz, Sewon Min, Patrick Lewis, Ledell Wu, Sergey Edunov, Danqi Chen, and Wen-tau Yih. 2020. Dense Passage Retrieval for Open-Domain Question Answering. In Proceedings of EMNLP, pages 6769–6781.
  22. Patrick Lewis, Ethan Perez, Aleksandra Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich Kuttler, Mike Lewis, Wen-tau Yih, Tim Rocktaschel, and others. 2020. Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks. In Advances in Neural Information Processing Systems.
  23. Junyi Li, Xiaoxue Cheng, Wayne Xin Zhao, Jian-Yun Nie, and Ji-Rong Wen. 2023. HaluEval: A Large-Scale Hallucination Evaluation Benchmark for Large Language Models. arXiv preprint arXiv:2305.11747. ↗
  24. Stephanie Lin, Jacob Hilton, and Owain Evans. 2022. TruthfulQA: Measuring How Models Mimic Human Falsehoods. In Proceedings of ACL, pages 3214–3252.
  25. Rana Malhas, Watheq Mansour, and Tamer Elsayed. 2022. Qur’an QA 2022: Overview of The First Shared Task on Question Answering over the Holy Qur’an. In Proceedings of OSACT.
  26. Rana Malhas, Watheq Mansour, and Tamer Elsayed. 2023. Qur’an QA 2023 Shared Task: Overview of Passage Retrieval and Reading Comprehension Tasks over the Holy Qur’an. In Proceedings of ArabicNLP, pages 690–701.
  27. Potsawee Manakul, Adian Liusie, and Mark J. F. Gales. 2023. SelfCheckGPT: Zero-Resource Black-Box Hallucination Detection for Generative Large Language Models. In Proceedings of EMNLP, pages 9004–9017. ↗
  28. Sewon Min, Kalpesh Krishna, Xinxi Lyu, Mike Lewis, Wen-tau Yih, Pang Wei Koh, Mohit Iyyer, Luke Zettlemoyer, and Hannaneh Hajishirzi. 2023. FActScore: Fine-grained Atomic Evaluation of Factual Precision in Long Form Text Generation. In Proceedings of EMNLP. ↗
  29. Hamdy Mubarak, Rana Malhas, Watheq Mansour, and others. 2025. IslamicEval 2025: The First Shared Task of Capturing LLMs Hallucination in Islamic Content. In Proceedings of ArabicNLP Shared Tasks, pages 480–493.
  30. Cheng Niu, Jibao Wu, Yifei Zhang, and others. 2024. RAGTruth: A Hallucination Corpus for Developing Trustworthy Retrieval-Augmented Language Models. In Proceedings of ACL, pages 10862–10878.
  31. Ossama Obeid, Nasser Zalmout, Salam Khalifa, Dima Taji, Mai Oudah, Bashar Alhafni, Go Inoue, Fadhl Eryani, Alexander Erdmann, and Nizar Habash. 2020. CAMeL Tools: An Open Source Python Toolkit for Arabic Natural Language Processing. In Proceedings of LREC, pages 7022– 7032.
  32. A. Omayrah and others. 2025. HUMAIN at IslamicEval 2025 Shared Task 1. In Proceedings of ArabicNLP Shared Tasks. find
  33. Arfath Pasha, Mohamed Al-Badrashiny, Mona Diab, Ahmed El Kholy, Ramy Eskander, Nizar Habash, Manoj Pooleery, Owen Rambow, and Ryan Roth. 2014. MADAMIRA: A Fast, Comprehensive Tool for Morphological Analysis and Disambiguation of Arabic. In Proceedings of LREC, pages 1094–1101.
  34. Stephen Robertson and Hugo Zaragoza. 2009. The Probabilistic Relevance Framework: BM25 and Beyond. Foundations and Trends in Information Retrieval, 3(4):333–389. ↗
  35. Neha Sengupta, Sunil Kumar Sahu, Bokang Jia, Satheesh Katipomu, Haonan Li, Fajri Koto, William Marshall, Gurpreet Gosal, Cynthia Liu, Zhiming Chen, and others. 2023. Jais and Jais-chat: Arabic-Centric Foundation and Instruction-Tuned Open Generative Large Language Models. arXiv preprint arXiv:2308.16149. ↗
  36. Nandan Thakur, Nils Reimers, Andreas Rueckle, Abhishek Srivastava, and Iryna Gurevych. 2021. BEIR: A Heterogeneous Benchmark for Zero-shot Evaluation of Information Retrieval Models. In Proceedings of NeurIPS Datasets and Benchmarks. find
  37. Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothee Lacroix, Baptiste Roziere, Naman Goyal, Eric Hambro, and others. 2023. LLaMA: Open and Efficient Foundation Language Models. arXiv preprint arXiv:2302.13971. ↗
  38. Raul Vazquez, Timothee Mickus, Elaine Zosa, Teemu Vahtola, Jorg Tiedemann, Aman Sinha, Vincent Segonne, Fernando Sanchez-Vega, Alessandro Raganato, and others. 2025. SemEval-2025 Task 3: Mu-SHROOM, the Multilingual Shared Task on Hallucinations and Related Observable Overgeneration Mistakes. In Proceedings of SemEval. find
  39. Jianqing Zhu, Huang Huang, Zhihang Lin, Juhao Liang, Zhengyang Tang, Khalid Almubarak, Abdulmohsen Alharthik, Bang An, Juncai He, Xiangbo Wu, and others. 2025. Second Language (Arabic) Acquisition of Large Language Models via Progressive Vocabulary Expansion. In Proceedings of ACL. find

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/mouhoub-2026-islamic-large-language-models-knowledge/