HIAST at QIAS 2025: Retrieval-Augmented LLMs with Top-Hit Web Evidence for Arabic Islamic Reasoning QA
Year: 2025
Type: conference-paper
Abstract: We describe our participation in the QIAS 2025 Shared Task on Islamic Studies Question Answering, comprising two subtasks: (1) Islamic Inheritance Reasoning and (2) General Islamic Knowledge Assessment.Both were solved using the Claude 4 Opus LLM via API with tailored prompting.For Subtask 1, we implemented a lightweight Retrieval-Augmented Generation (RAG) pipeline, which retrieves the top Google Search result (often from Is-lamWeb), preprocesses it, and appends it to a structured few-shot Arabic prompt, thereby boosting reasoning accuracy.For Subtask 2, where web-retrieval was not feasible due to closed-book sources, we applied topic-diverse few-shot prompting to leverage the model's internal knowledge.Our systems achieved 4th/15 (0.895) in Subtask 1 and 3rd/10 (0.9259) in Subtask 2, demonstrating the effectiveness of targeted retrieval in open-web contexts and structured prompting in closed-domain Arabic QA.
Authors: Mohamed Motasim Hamed, Nada Ghneim, Riad Sonbol
Language: English
Topic: Topic Modeling
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W4416037634