This paper evaluates the knowledge and reasoning capabilities of large-scale language models (LLMs) for 'ilm al-mawarith' (Islamic inheritance law). Using a benchmark of 1,000 multiple-choice questions covering various inheritance scenarios, we evaluate the performance of seven LLMs (o3, Gemini 2.5, ALLaM, Fanar, LLaMA, and Mistral). This benchmark was designed to test the models' ability to understand and calculate inheritance distributions as prescribed in Islamic jurisprudence. While o3 and Gemini 2.5 achieved over 90% accuracy, ALLaM, Fanar, LLaMA, and Mistral achieved less than 50% accuracy. This discrepancy reflects significant differences in reasoning ability and domain adaptation. A detailed error analysis was conducted to identify recurring failure patterns across models, including misunderstanding inheritance scenarios, misapplication of legal rules, and insufficient domain knowledge. These findings highlight the limitations of structured legal reasoning and suggest directions for improving Islamic legal reasoning performance.
