Explainable Rule Application via Structured Prompting: A Neural-Symbolic Approach

LLMs need structured guidance for consistent rule application: The paper demonstrates that large language models struggle with applying rules consistently and handling exceptions on their own, but can be significantly improved through a structured prompting framework that breaks reasoning into clear, verifiable steps (entity identification, property extraction, and rule application).

Hybrid neural-symbolic approaches improve both accuracy and explainability: By combining LLMs' natural language understanding with formal symbolic logic, the framework achieved substantial performance gains—for example, OpenAI's o1 model improved from an F1 score of 0.714 to 0.929 on legal reasoning tasks—while making the reasoning process transparent and auditable.

Domain experts can maintain control without technical modifications: The framework externalizes task definitions, allowing legal experts and other domain specialists to refine the logical rules and structures governing AI decisions without requiring changes to the underlying AI system architecture, supporting more responsible human oversight.

Explainability and accuracy are complementary, not competing goals: Rather than trading off transparency for performance, this approach shows that making AI reasoning more explainable through structured decomposition actually improves accuracy, suggesting that responsible AI design can enhance both trustworthiness and reliability.

Formal verification enables accountability in high-stakes domains: By incorporating symbolic rule application with formal verification, the system provides a pathway for auditable decision-making in sensitive fields like legal analysis, where stakeholders need to understand and verify how conclusions were reached.