Explaining why a risk score differs from another is a common but difficult task in domains like finance and insurance. A new preprint on arXiv proposes a controlled large language model (LLM) workflow that addresses this by grounding explanations in similar historical transactions.
The workflow combines two components: reference retrieval, which finds comparable past cases, and contrastive SHAP, a technique for attributing differences in model outputs. Together, they let the LLM generate explanations that are tied to concrete evidence rather than relying on ungrounded reasoning.
The authors frame this as a step toward more trustworthy and auditable risk-score justifications. Because the approach is controlled, it may reduce the risk of hallucinated or irrelevant explanations, though the abstract does not report experimental results.