Silah: Research on Large Language Models for Islamic Inheritance Law Reasoning

A research project named Silah Inheritance Reasoning LLM aims to use large language models to solve complex reasoning problems in Islamic inheritance law (Faraid). Combining three technologies—supervised fine-tuning, RAG, and RAG-FT—this project opens up a new path for the intelligent processing of religious legal texts and demonstrates the potential of deep AI applications in vertical domains.

Islamic inheritance law (Al-Faraid) is one of the most precise and mathematical branches of Islamic law, involving complex distribution rules, hierarchical kinship relationships, and handling of special cases. Traditional manual calculations are prone to omissions; the increasing complexity of modern family structures and the rise in cross-border inheritance cases have put manual processing under dual challenges of efficiency and accuracy—this is the core problem the Silah project aims to address.

Silah uses a combination of three cutting-edge technologies: 1. Supervised Fine-tuning (SFT): Learning the basic rule framework of inheritance law through massive case data; 2. Retrieval-Augmented Generation (RAG): Retrieving provisions from authoritative sharia literature to solve the model hallucination problem; 3. Retrieval-Augmented Fine-Tuning (RAG-FT): Integrating retrieval capabilities into the fine-tuning process to establish an end-to-end "retrieval-reasoning-generation" capability chain and improve reasoning coherence.

Silah's core capability is to generate structured and rule-compliant solutions, featuring traceable steps (each decision has a clear reasoning chain), verifiable rules (calculation basis can be traced back to specific sharia provisions), and explainable results (non-technical users can understand the decision logic)—reflecting the principles of transparency, interpretability, and auditability.

Silah's potential application scenarios include: for religious scholars (assisting in verifying calculations for complex cases), legal practitioners (reference for cross-border inheritance cases), ordinary Muslim families (lowering the threshold for professional consultation), and the AI research community (providing a technical paradigm for the "AI + religious law" interdisciplinary field).

Silah proves that for highly specialized, rule-intensive domains, a pure general large model is insufficient; the combination of "domain fine-tuning + knowledge retrieval + reasoning enhancement" is a feasible path. The exploration of RAG-FT provides new ideas for AI applications in other vertical domains—integrating retrieval capabilities deeply into training rather than as a post-hoc patch.

Silah represents a microcosm of AI's deep penetration into traditional cultural fields, and its value lies in solving complex real-world problems. With technological iteration, AI is expected to play a role in more traditional knowledge domains, becoming a bridge connecting ancient wisdom and modern life.