Beyond Training Data: Zero-Shot Mathematical Problem Solving with Large Language Models

This study investigates the potential of Large Language Models (LLMs) for mathematical problem solving using a zero-shot multi-path self-consistency technique combined with programmatic solution verification. We create Python code that is evaluated by a majority vote by providing multiple reasoning pathways using differences in temperature and top-k values, as well as improving inquiry comprehension through phrasing. When applied to the GSM8K and MATH datasets, this approach achieved $\mathbf{45 \%}$ accuracy on GSM8K and $\mathbf{14.57 \%}$ on MATH with the Granite 8B code model. These findings show that even in the absence of task-specific training data, LLMs can greatly enhance their performance by breaking down complicated tasks and employing a consistent multipath technique.