Gold-Guided Programmatic Distillation for Financial Reasoning over Hybrid Tables and Text

arXiv cs.CL · 6 hours ago

Researchers developed a method to distill financial reasoning capabilities from large language models to smaller ones using execution-verified Python programs instead of natural language explanations. A 7-billion-parameter student model achieved 87.00 EM on the TAT-QA benchmark, outperforming a 72-billion-parameter teacher model that scored 78.46 EM. The approach enables compact models to perform reliable numerical reasoning over hybrid tabular and textual financial data by ensuring supervisory signals are verified through correct code execution.