Improving Math Problem Solving in Large Language Models Through Categorization and Strategy Tailoring

TL;DR

This work shows that structuring math problem solving for LLMs via problem categorization into four domains and assigning category-specific solving strategies substantially improves accuracy and efficiency. A lightweight indicator-based neural model for categorization, aided by carefully curated training data, achieves ~84% accuracy—close to the ~92% of SOTA models—and enables meaningful gains over random strategy selection. The study contrasts simple categorization with a more advanced in-context learning approach, demonstrating data quality as a critical factor in performance. Downstream evaluation confirms that even with imperfect categorization, the approach outperforms baselines, illustrating practical impact for real-time math problem solving in LLM-driven workflows. The findings highlight the value of data-centric design and targeted prompting for reliable mathematics reasoning in LLMs.

Abstract

In this paper, we explore how to leverage large language models (LLMs) to solve mathematical problems efficiently and accurately. Specifically, we demonstrate the effectiveness of classifying problems into distinct categories and employing category-specific problem-solving strategies to improve the mathematical performance of LLMs. We design a simple yet intuitive machine learning model for problem categorization and show that its accuracy can be significantly enhanced through the development of well-curated training datasets. Additionally, we find that the performance of this simple model approaches that of state-of-the-art (SOTA) models for categorization. Moreover, the accuracy of SOTA models also benefits from the use of improved training data. Finally, we assess the advantages of using category-specific strategies when prompting LLMs and observe significantly better performance compared to non-tailored approaches.