Researchers have explored using large language models (LLMs) to translate complex credit risk model explanations into more understandable narratives for stakeholders. A study using Freddie Mac loan data compared three pipelines: standard tabular (XGBoost + SHAP), a network-based approach (GNN + GNNExplainer), and a bimodal combination. These were paired with LLMs including Gemma 3 4B, DeepSeek R1 70B, and Gemini 2.5. The findings indicate that the evidence representation method is a greater constraint on explanation quality than the LLM used, and while LLMs can identify influential factors, they are less reliable in stating the direction of influence, which is critical for adverse action communications. Professionals also applied stricter evidentiary standards than non-professionals. AI
IMPACT LLMs can improve communication of complex financial model outputs, but their reliability in conveying directional influence requires further development for regulated environments.
RANK_REASON Research paper published on arXiv detailing the use of LLMs for credit risk explanation. [lever_c_demoted from research: ic=1 ai=1.0]
- DeepSeek R1 70B
- Freddie Mac
- Gemini 2.5
- Gemma 3 4B
- GNNExplainer: Generating Explanations for Graph Neural Networks
- Graph Neural Networks
- large-language models
- María Óskarsdóttir
- Shap
- XGBoost
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
