PuxAI at QIAS 2025: Multi-Agent Retrieval-Augmented Generation for Islamic Inheritance and Knowledge Reasoning
Year: 2025
Type: conference-paper
Abstract: This paper addresses the challenge of applying Large Language Models (LLMs) to Islamic jurisprudence, a domain that requires both textual retrieval and precise rule-based reasoning.We focus on the QIAS 2025 shared task, which evaluates LLMs on two subtasks: Islamic inheritance reasoning and general Islamic knowledge assessment.Prior works in Arabic NLP and religious QA largely emphasize retrieval and classification, but they do not evaluate multi-step procedural reasoning.To fill this gap, we propose a hybrid multi-agent framework, termed Retrieval-Augmented Reasoning (RAR).For inheritance problems, our Virtual Inheritance Expert parses natural language cases into structured JSON, retrieves relevant fatwas, and applies rule-based synthesis.For general knowledge, our Proponent-Critic Debate simulates dialectical reasoning, with a head scholar model providing final judgment.Using an ensemble of Gemini, Fanar, and Mistral, our system achieved 2nd place in Subtask 1 and 1st place in Subtask 2. These results demonstrate that decomposing complex reasoning into specialized pipelines supports robustness and accuracy in high-stakes domains.
Authors: Nguyen Xuan Phuc, Thìn Đặng Văn
Language: English
Topic: Topic Modeling
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W4416037645