LLMs for Explainable Business Decision-Making: A Reinforcement Learning Fine-Tuning Approach

Natural language explanations can be both accurate and faithful: The paper demonstrates that LLMs can generate explanations for business decisions (like mortgage approvals) that are simultaneously correct about the actual decision and truthful about which factors drove that decision—addressing a key gap where previous AI explanation methods provided numbers without coherent narratives.

One AI system can explain itself differently to different audiences without changing its decisions: LEXMA successfully tailors explanations for multiple audiences with different needs and backgrounds while keeping the underlying decision rule consistent, meaning a mortgage approval system could explain its decision differently to a loan officer versus a consumer without being deceptive.

Human-labeled explanations aren't necessary for training: The framework uses reward signals that don't require large amounts of human-written explanations, making it more practical and cost-effective to implement than approaches that demand extensive manual annotation of "good" explanations.

This approach addresses real-world accountability needs: By focusing on mortgage decisions—a high-stakes domain with regulatory requirements—the research tackles a genuine business problem where consumers and regulators need to understand why they were approved or denied, not just receive a yes/no answer.

Explainability requires intentional design, not just add-ons: The paper shows that generating trustworthy explanations requires careful technical choices (combining supervised fine-tuning with reinforcement learning) rather than simply asking an AI to explain itself after the fact.