Which Models Perform Better in Inheritance Reasoning?

The paper evaluates the performance of commercial versus open-source large language models in the QIAS 2026 Shared Task on Arabic Islamic inheritance reasoning, focusing on legal interpretation and multi-step reasoning. The results indicate that commercial models, particularly \textit{Gemini 2.5 Flash}, outperform open-source models, achieving a mean relative error (MRE) of 0.989, highlighting a significant reliability gap in structured legal reasoning tasks. This study is crucial for practitioners as it underscores the importance of model selection in domains requiring precise legal reasoning and numerical computation.