The AI Auditability Dilemma: A Comparative Analysis of Retrieval-Augmented Generation and Fine-Tuning Large Language Models in Credit Risk Management

The integration of Large Language Models (LLMs) into credit risk management offers transformative capabilities for parsing vast arrays of unstructured corporate financial data, market indices, and qualitative disclosures. However, applying generative AI to high-stakes lending decisions introduces a critical "auditability dilemma": balancing model performance and behavioral alignment against the strict traceability required by financial regulators. This paper presents a comparative analysis of the two dominant LLM adaptation strategies within credit risk mitigation: Parameter-Efficient Fine-Tuning (PEFT) and Retrieval-Augmented Generation (RAG). We evaluate both frameworks across four core dimensions essential to model risk management (MRM): factual accuracy (hallucination mitigation), data freshness, operational cost, and mathematical explainability. Our findings indicate that while fine-tuning effectively optimizes behavioral patterns, specialized financial terminology, and strict credit memo formatting, its reliance on parametric memory acts as a black box that fails strict regulatory audits and struggles with real-time macroeconomic shifts. Conversely, RAG provides an explicit, auditable trail directly to source documents (such as 10-K filings or loan applications), making it significantly more resilient to hallucinations and highly adaptive to dynamic borrower data. To resolve the trade-offs inherent to both methods, we propose a hybrid enterprise blueprint that integrates a lightweight, format-fine-tuned LLM with a Knowledge Graph-powered RAG pipeline (GraphRAG). This unified architecture ensures strict adherence to internal risk classification templates while preserving the deterministic, verifiable evidence pathways required to satisfy global financial compliance standards.