Step-by-Step Correction of LLM-based Math Word Problems Solutions

Following the success of Large Language Models (LLMs) in language tasks, LLMs have been adapted for reasoning in math word problems (MWPs). MWP is a complex task that requires both semantic understanding of text and mathematical reasoning, such that achieving high accuracy in MWP remains a challenge. We find that MWP performance can be improved by step-by-step reasoning where the LLM is trained to generate smaller and more manageable steps. We further propose a post-processing correction model to edit the initial solutions given by the LLM. Our correction model, designed to detect and rectify mistakes in these steps, is firstly pretrained using heuristically generated model-agnostic error data and further finetuned with model-specific errors generated through self-supervised augmentation. The correction model iteratively refines the solution step-by-step by analyzing the problem statement and steps up until the current one, making corrections as needed, and repeating the process until all steps in the solution are processed. Experimental results demonstrate that the step-by-step reasoning significantly improves MWP performance compared to one-step solutions. The combination of pretraining and finetuning effectively aligns the correction model with the error patterns of the reasoning model, resulting in further accuracy improvements through error correction.