ADAPT–MTU HAI at QIAS2025: Dual-Expert LLM Fine-Tuning and Constrained Decoding for Arabic Islamic Inheritance Reasoning
Year: 2025
Type: conference-paper
Abstract: We present ADAPT-MTU HAI's submission to Subtask 1 of the QIAS 2025 shared task, which focuses on Arabic multiple-choice question answering (MCQ) for Islamic inheritance law.This domain presents unique challenges, requiring models to navigate precise fractional computations, exclusion rules, and doctrinal nuances under strict format constraints.Our system employs a dual-expert architecture based on ALLaM-7B, integrating a LoRA-fine-tuned model specialised for inheritance reasoning with its generalist base counterpart.A custom constrained decoding mechanism ensures output compliance, while arbitration between the two models enhances answer stability.Our system achieves 60.0% accuracy on the development set and 54.7% on the official blind test set-substantially improving upon the baseline.We analyse common failure modes and discuss implications for structured legal reasoning using large language models.
Authors: Shehenaz Hossain, Haithem Afli
Institution: Munster Technological University
Language: English
Topic: Topic Modeling
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
Awards: G2399253094, G3251807
OpenAlex ID: W4416037643