Brain-Inspired Two-Stage Approach: Enhancing Mathematical Reasoning by Imitating Human Thought Processes

Yezeng Chen,Zui Chen,Yi Zhou

2024-02-24

Abstract:Although large language models demonstrate emergent abilities in solving math word problems, there is a challenging task in complex multi-step mathematical reasoning tasks. To improve model performance on mathematical reasoning tasks, previous work has conducted supervised fine-tuning on open-source models by improving the quality and quantity of data. In this paper, we propose a novel approach, named Brain, to imitate human thought processes to enhance mathematical reasoning abilities, using the Frontal Lobe Model to generate plans, and then employing the Parietal Lobe Model to generate code and execute to obtain answers. First, we achieve SOTA performance in comparison with Code LLaMA 7B based models through this method. Secondly, we find that plans can be explicitly extracted from natural language, code, or formal language. Our code and data are publicly available at

Computation and Language,Artificial Intelligence,Machine Learning

What problem does this paper attempt to address?

The problem that this paper attempts to solve is that in complex multi - step mathematical reasoning tasks, large language models (LLMs) lack strong logical reasoning abilities, especially the challenges faced when solving complex multi - step mathematical problems. Although existing large language models have shown a certain ability in solving mathematical word problems, they still have deficiencies when dealing with complex multi - step mathematical reasoning tasks. These deficiencies are mainly reflected in the fact that the model cannot effectively manage the two processes of planning and reasoning calculation simultaneously, resulting in poor performance in solving complex problems.

To overcome this limitation, the paper proposes a new method named Brain, which imitates the thinking process of the human brain to enhance mathematical reasoning ability. Specifically, the Brain method solves mathematical reasoning tasks by constructing a two - stage technical framework, which simulates the way humans solve problems:

1. **Plan Generation**: Generate a plan based on the problem. This step is completed by the Frontal Lobe Model, simulating the decision - making process of the human brain.
2. **Code Generation and Execution**: Generate code based on the plan and execute it to obtain the answer. This step is completed by the Parietal Lobe Model, simulating the process of the human brain dealing with code structures and logical flows.

Through this method, the paper aims to answer the following research questions:
- Does the output of LLMs contain a plan? If so, can it be explicitly extracted?
- How can these plans be used to solve complex reasoning tasks?
- How can higher - quality plans be constructed? Are high - quality plans necessarily useful?

The main contributions of the paper are:
- Proposing a new method, Brain, which imitates the thinking process of the human brain to enhance mathematical reasoning ability.
- Brain has achieved state - of - the - art performance compared to models based on Code LLaMA 7B in the zero - sample case.
- Discovering that plans can be explicitly extracted from natural languages, codes, or formal languages.

In general, by introducing a two - stage framework, the paper has successfully improved the performance of LLMs in complex mathematical reasoning tasks, providing new ideas and directions for future related research.