--- title: "CVPD at QIAS 2025 Shared Task: An Efficient Encoder-Based" description: "The authors present the work as a measured trade-off. 12 references, cited by 4 papers here." authors: ["Salah Eddine Bekhouche", "Abdellah Zakaria Sellam", "Telli Hichem", "Cosimo Distante", "Abdenour Hadid"] year: 2025 venue: "Proceedings of The Third Arabic Natural Language Processing Conference: Shared Tasks" source: "https://aclanthology.org/2025.arabicnlp-sharedtasks.128/" doi: "10.18653/v1/2025.arabicnlp-sharedtasks.128" licence: "CC BY 4.0 (ACL Anthology)" licence_url: "https://creativecommons.org/licenses/by/4.0/" summary_licence: "Elamur Cosmic License" canonical: "https://research.elamur.ai/free/en/papers/bekhouche-2025-cvpd-qias-2025-shared-task/" --- [Elamur Research](https://md.research.elamur.ai/free/en/) · [this page in HTML](https://research.elamur.ai/free/en/papers/bekhouche-2025-cvpd-qias-2025-shared-task/) [home](https://md.research.elamur.ai/free/en/) / [papers](https://md.research.elamur.ai/free/en/papers/) / [2025](https://md.research.elamur.ai/free/en/papers/2025/page/11/) / CVPD at QIAS 2025 Shared Task: An Efficient Encoder-Based Approach for Islamic Inheritance Reasoning # CVPD at QIAS 2025 Shared Task: An Efficient Encoder-Based Approach for Islamic Inheritance Reasoning **Authors:** [Salah Eddine Bekhouche](https://md.research.elamur.ai/free/en/authors/salah-eddine-bekhouche/), Abdellah Zakaria Sellam, Telli Hichem, Cosimo Distante and [Abdenour Hadid](https://md.research.elamur.ai/free/en/authors/abdenour-hadid/) · **Year:** 2025 · **Venue:** [Proceedings of The Third Arabic Natural Language Processing Conference: Shared Tasks](https://md.research.elamur.ai/free/en/venues/arabicnlp/) ([ArabicNLP 2025](https://md.research.elamur.ai/free/en/venues/arabicnlp/2025/)) · **Publisher:** [Association for Computational Linguistics](https://md.research.elamur.ai/free/en/publishers/association-for-computational-linguistics/) · **Licence:** [CC BY 4.0 (ACL Anthology)](https://creativecommons.org/licenses/by/4.0/) **Links:** [Publisher's page](https://aclanthology.org/2025.arabicnlp-sharedtasks.128/) · [PDF](https://aclanthology.org/2025.arabicnlp-sharedtasks.128.pdf) · [DOI 10.18653/v1/2025.arabicnlp-sharedtasks.128](https://doi.org/10.18653/v1/2025.arabicnlp-sharedtasks.128) · [OpenAlex](https://openalex.org/W4416037636) **Citations:** 0 citations (OpenAlex, 2026-10-14) · [4](https://md.research.elamur.ai/free/en/papers/bekhouche-2025-cvpd-qias-2025-shared-task/cited-by/) in our library, of the 149,439 papers whose bibliographies we read ## Summary The paper addresses multiple-choice questions on Islamic inheritance law ('Ilm al-Mawārīth) in SubTask 1 of the QIAS 2025 shared task. These questions require identifying heirs and computing exact shares, including the 'awl and radd adjustments. The authors do not generate a chain of reasoning. They treat the task as ranking: a pretrained Arabic encoder embeds the question and each candidate answer separately, and a small trainable head called Attentive Relevance Scoring (ARS) gives each pair a relevance score. Training combines three losses: contrastive alignment, direct supervision of the scores, and a regularizer that keeps the score logits from collapsing. Five encoders were compared. MARBERT performed best, with 69.87% test accuracy. API-based LLMs scored higher, up to 87.60% for Gemini-2.5-flash, but only when each question was sent on its own. When 50 questions were packed into one prompt, its accuracy fell to 68.65%. The authors present the work as a measured trade-off. A small model that runs on-device and keeps data private can serve as an assistant to experts, while the large cloud models remain more accurate. ## Key points - **Data:** QIAS 2025 SubTask 1 has 9,446 training items, 1,000 validation items and 1,000 unlabeled test items. Items are graded Beginner, Intermediate and Advanced, and each has six options (A–F). - **Encoders tested with ARS:** ArabicBERT-Mini, ArabicBERT, AraBERT, MARBERT and QARiB. - MARBERT (162.9M params, 124.5 GFlops) reached 77.32% on validation and 69.87% on test. - ArabicBERT-Mini, with only 11.6M params, still reached 64.23% on test. - **ARS architecture:** - The [CLS] vectors of question and answer are ℓ2-normalized. - Each is projected into a shared space, and the two projections are multiplied element-wise and passed through tanh. - A learned attention vector followed by a sigmoid produces the score. - **Loss weights:** contrastive 0.4, dynamic relevance 0.4, logit regularization 0.2. - **LLM results with batched prompts:** deepseek-chat 66.40%, deepseek-reasoner 69.40%, gemini-2.0-flash 60.44%, gemini-2.5-flash 68.65%. Gemini-2.5-flash rose to 87.60% when questions were sent one at a time. ## Interesting ideas - Packing many questions into one prompt cut a strong LLM's accuracy by about 19 points, apparently because neighbouring questions interfere with each other, so the prompt format is part of the result. - A shallow model that only scores how well a question matches each option, without any reasoning, still gets about 65–70% on a task that supposedly needs multi-step arithmetic, which shows how much a benchmark can leak through surface cues. - A regularization term that pushes for variance in the pre-sigmoid logits keeps a learned scorer from giving every candidate nearly the same score. - Results are broken down by difficulty tier (simple share lookup, radd adjustment, full monetary distribution), which separates language understanding from exact calculation. - The authors explain the accuracy gap by the lack of explicit symbolic computation and plan to add it, since fixed shares with proportional reduction or redistribution are pure rule arithmetic. ## For Mova How these ideas could serve Mova, our small open language model — its files, stages and tests: - **I1** (learn): Query LLM teacher tiers one item per call when producing silver labels or grading, and record the prompt format with every evaluation number. - **I2** (learn): Run such a matcher, or the TF-IDF baseline the authors propose, as a negative control on our test sets; items it solves easily do not prove that the reasoning pipeline works. - **I3** (tree): gparse's arc or label scorer could use a similar variance penalty to stop its candidate scores collapsing early in training. - **I4** (answer): Report maths word problem and school-test accuracy per reasoning tier, so we can see whether failures come from reading or from computing. - **I5** (world): Model inheritance-style division as holdings with fraction measures plus normalization rules in the world layer, so every share becomes an explainable derived state. *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 - [University of the Basque Country](https://md.research.elamur.ai/free/en/universities/university-of-the-basque-country/) — University, Leioa, Spain · 57 works · 16,520 citations - [National Research Council](https://md.research.elamur.ai/free/en/ngos/national-research-council/) — Nonprofit, Rome, Italy · 15 works · 15,649 citations Also on the byline: [Sorbonne University Abu Dhabi](https://ror.org/03e1ymy32). 📚 Shelves (6): [🖥️ Computer Science](https://md.research.elamur.ai/free/en/shelves/computer-science/) · 16,667 works · [🤖 Artificial Intelligence](https://md.research.elamur.ai/free/en/shelves/computer-science/artificial-intelligence/) · 16,194 works · [🎭 Arts and Humanities](https://md.research.elamur.ai/free/en/shelves/arts-and-humanities/) · 16,039 works · [🗣️ Language and Linguistics](https://md.research.elamur.ai/free/en/shelves/arts-and-humanities/language-and-linguistics/) · 15,891 works · [🤝 Social Sciences](https://md.research.elamur.ai/free/en/shelves/social-sciences/) · 1,772 works · [🤝 Political Science and International Relations](https://md.research.elamur.ai/free/en/shelves/social-sciences/political-science-and-international-relations/) · 182 works 🧭 Topics (2): [🧮 Reasoning](https://md.research.elamur.ai/free/en/topics/reasoning/) · 1,426 works · [Semitic languages](https://md.research.elamur.ai/free/en/topics/semitic-languages/) · 2,908 works ## Cite Cite as: Salah Eddine Bekhouche, Abdellah Zakaria Sellam, Telli Hichem, Cosimo Distante and Abdenour Hadid. 2025. CVPD at QIAS 2025 Shared Task: An Efficient Encoder-Based Approach for Islamic Inheritance Reasoning. Proceedings of The Third Arabic Natural Language Processing Conference: Shared Tasks. The paper itself, not this page. BibTeX from the library's record of the work. ```bibtex @inproceedings{bekhouche2025cvpd, title = {{CVPD at QIAS 2025 Shared Task: An Efficient Encoder-Based Approach for Islamic Inheritance Reasoning}}, author = {Salah Eddine Bekhouche, Abdellah Zakaria Sellam, Telli Hichem, Cosimo Distante and Abdenour Hadid}, booktitle = {Proceedings of The Third Arabic Natural Language Processing Conference: Shared Tasks}, year = {2025}, doi = {10.18653/v1/2025.arabicnlp-sharedtasks.128}, url = {https://doi.org/10.18653/v1/2025.arabicnlp-sharedtasks.128} } ``` - [← previous CTCC: A Robust and Stealthy Fingerprinting Framework for Large Language Models via Cross-Turn Contextual Correlation Backdoor](https://md.research.elamur.ai/free/en/papers/xu-2025-ctcc-robust-stealthy-fingerprinting-framework/) - [next → DaKultur: Evaluating the Cultural Awareness of Language Models for Danish with Native Speakers](https://md.research.elamur.ai/free/en/papers/eberstein-2025-dakultur-evaluating-cultural-awareness-language/) - [↑ back to papers of 2025](https://md.research.elamur.ai/free/en/papers/2025/page/11/) ## References 12 references, 10 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet. 1. OpenAI. 2023. [Gpt-4 technical report](https://md.research.elamur.ai/free/en/papers/openai-2023-gpt-4-technical-report/). 2. Ahmed Abdelali, Sabit Hassan, Hamdy Mubarak, Kareem Darwish, and Younes Samih. 2021. [Pre-training bert on arabic tweets: Practical considerations](https://md.research.elamur.ai/free/en/papers/abdelali-2021-pre-training-bert-arabic-tweets/). arXiv preprint arXiv:2102.10684. [↗](https://arxiv.org/abs/2102.10684) 3. Arthur Phillips and Roland Knyvet Wilson. 1995. A Treatise on the Muhammadan Law. Kegan Paul International. [find](https://scholar.google.com/scholar?q=Arthur+Phillips+and+Roland+Knyvet+Wilson%2E+1995%2E+A+Treatise+on+the+Muhammadan+Law%2E+Kegan+Paul+International%2E) 4. Muhammad Abdul-Mageed, AbdelRahim Elmadany, and El Moatez Billah Nagoudi. 2020. [Arbert & marbert: Deep bidirectional transformers for arabic](https://md.research.elamur.ai/free/en/papers/abdul-mageed-2021-arbert-marbert-deep-bidirectional-transformers/). arXiv preprint arXiv:2101.01785. [↗](https://arxiv.org/abs/2101.01785) 5. Wissam Antoun, Fady Baly, and Hazem Hajj. 2020. [Arabert: Transformer-based model for arabic language understanding](https://md.research.elamur.ai/free/en/papers/antoun-2020-arabert-transformer-based-model-arabic/). arXiv preprint arXiv:2003.00104. [↗](https://arxiv.org/abs/2003.00104) 6. Khalid Arabi and Samira Hassan. 2023. Large language models in the arabic-speaking world: A case study on domain-specific challenges. Journal of Natural Language Processing Arabia, 5(2):45–61. [find](https://scholar.google.com/scholar?q=Khalid+Arabi+and+Samira+Hassan%2E+2023%2E+Large+language+models+in+the+arabic%2Dspeaking+world%3A+A+case+study+on+domain%2Dspecific+challenges%2E+Journal+of+Natural+Language+Processing+Arabia%2C+5%282%29%3A45%E2%80%9361%2E) 7. Salah Eddine Bekhouche, Azeddine Benlamoudi, Yazid Bounab, Fadi Dornaika, and Abdenour Hadid. 2025. [Enhanced arabic text retrieval with attentive relevance scoring](https://md.research.elamur.ai/free/en/papers/bekhouche-2025-enhanced-arabic-text-retrieval-attentive/). arXiv preprint arXiv:2507.23404. [↗](https://arxiv.org/abs/2507.23404) 8. Abdessalam Bouchekif, Samer Rashwani, Emad Mohamed, Mutaz Al-Khatib, Heba Sbahi, Shahd Gaben, Wajdi Zaghouani, Aiman Erbad, and Mohammed Ghaly. 2025a. [Qias 2025: Overview of the shared task on islamic inheritance reasoning and knowledge assessment](https://md.research.elamur.ai/free/en/papers/bouchekif-2025-qias-2025-overview-shared-task/). In Proceedings of The Third Arabic Natural Language Processing Conference, ArabicNLP 2025, Suzhou, China, November 5–9, 2025. Association for Computational Linguistics. 9. Abdessalam Bouchekif, Samer Rashwani, Heba Sbahi, Shahd Gaben, Mutaz Al-Khatib, and Mohammed Ghaly. 2025b. [Assessing large language models on islamic legal reasoning: Evidence from inheritance law evaluation](https://md.research.elamur.ai/free/en/papers/bouchekif-2025-assessing-large-language-models-islamic/). In Proceedings of The Second Arabic Natural Language Processing Conference (Arabic-NLP 2025), Suzhou, China. Association for Computational Linguistics. 10. Sébastien Bubeck, Varun Chandrasekaran, Ronen Eldan, Johannes Gehrke, Eric Horvitz, Ece Kamar, Peter Lee, Yin Tat Lee, Yuanzhi Li, Scott Lundberg, and 1 others. 2023. [Sparks of artificial general intelligence: Early experiments with gpt-4](https://md.research.elamur.ai/free/en/papers/bubeck-2023-sparks-artificial-general-intelligence-early/). arXiv preprint arXiv:2303.12712. [↗](https://arxiv.org/abs/2303.12712) 11. Ali Safaya, Moutasem Abdullatif, and Deniz Yuret. 2020. [Kuisail at semeval-2020 task 12: Bert-cnn for offensive speech identification in social media](https://md.research.elamur.ai/free/en/papers/safaya-2020-kuisail-semeval-2020-task-12/). arXiv preprint arXiv:2007.13184. [↗](https://arxiv.org/abs/2007.13184) 12. Gemini Team and 1 others. 2023. [Gemini: A family of highly capable multimodal models](https://md.research.elamur.ai/free/en/papers/gemini-team-2023-gemini-family-highly-capable-multimodal/). *Updated 31 October 2026*