Enhancing Financial Reasoning via Program-of-Thought Learning

Large Language Models (LLMs) have achieved remarkable progress in reasoning capabilities, yet financial reasoning remains a significant challenge due to the demand for both numerical reasoning and understanding of domain-specific terminology. While prompt engineering techniques have improved reasoning performance, they often suffer from inconsistency, sensitivity to example selection, and limited generalizability. Consequently, supervised fine-tuning (SFT) and reinforcement learning (RL) are widely adopted to post-train LLMs using chain-of-thought to improve financial reasoning. A fundamental question remains: how can we more effectively represent and measure the reasoning steps and enable models to learn more efficiently, particularly for numerically intensive financial reasoning tasks? In this study, we propose a learning framework that enhances financial reasoning in LLMs through a novel program-of-thought (PoT) paradigm. PoT structures reasoning as step-by-step programmatic operations, such as code, pseudo-code, formulas, or symbolic steps, offering greater clarity, consistency, and executability. Our framework begins with SFT followed by RL using Group Relative Policy Optimization (GRPO), with a reward designed to balance accuracy, reasoning steps, and format adherence. Experiments on the FinQA benchmark demonstrate that our method outperforms state-of-the-art reasoning models. Our findings suggest that PoT learning is more effective than long natural-language CoT in the FinQA task, offering a promising direction for post-training LLMs for complex, domain-specific reasoning tasks.