Towards Interpretable Tabular Reasoning: Enhancing LLM Reasoning on Tabular Data with Pre-Constructed Logic Graph

Tabular data analysis is critical in highstakes domains like finance and healthcare, where both accuracy and interpretability are paramount. Traditional tree-based models excel in performance but suffer from poor interpretability and reliance on rigid feature engineering. Large Language Models (LLMs), while transparent in reasoning, often fail to capture the precise statistical distributions needed for accurate tabular prediction. To bridge this gap, we propose LogGER (Logic-Graph-Enhanced LLM Reasoning), a novel framework that harmonizes the predictive power of trees with the reasoning transparency of LLMs. Unlike existing hybrids, LogGER explicitly distills statistical patterns from data priors into a human-readable Logic Graph, transforming numerical decision into correlation semantic chains. This graph then serves as a verifier for our Process Supervision mechanism. Specifically, we employ a process reward model that steers the LLM's intermediate reasoning steps, ensuring alignment with the established logic graph. Extensive experiments demonstrate that LogGER not only achieves state-of-the-art accuracy, consistently outperforming both treebased models and LLM baselines, but also provides verifiable, transparent decision paths, establishing a new paradigm for interpretable tabular reasoning.