Core Concepts
Bootstrapping mathematical questions by rewriting them from multiple perspectives, including forward and backward reasoning, can significantly improve the mathematical problem-solving abilities of large language models.
Abstract
The paper proposes a novel method called MetaMath to enhance the mathematical reasoning capabilities of large language models (LLMs). The key idea is to bootstrap the available mathematical questions in the training set by rewriting them from multiple perspectives, including forward and backward reasoning.
Specifically, the authors first apply answer augmentation to generate more reasoning paths for each question. Then, they propose three types of question bootstrapping:
Rephrasing: Generating rephrased questions using the LLM, where the quality of the rephrased questions is evaluated by comparing their reasoning paths with the ground-truth answers.
Self-Verification (SV): Rewriting the question with the answer into a declarative statement and then asking for the value of the unknown variable.
FOBAR (Forward-Backward Reasoning): Directly appending the answer to the original question and asking for the value of the unknown variable.
The authors combine all the augmented data, including answer-augmented data and bootstrapped questions, to create a new dataset called MetaMathQA. They then finetune the state-of-the-art open-source LLM, LLaMA-2, on MetaMathQA to obtain the MetaMath model.
Experimental results on two popular mathematical reasoning benchmarks, GSM8K and MATH, demonstrate that MetaMath significantly outperforms existing open-source LLMs. Specifically, MetaMath-7B achieves 66.5% on GSM8K and 19.8% on MATH, exceeding the previous best open-source LLMs by 11.5% and 8.7%, respectively. The authors also show that the diversity of the training data, especially the backward reasoning questions, is crucial for improving the mathematical reasoning abilities of LLMs.
Stats
James buys 5 packs of beef that are 4 pounds each.
The price of beef is $5.50 per pound.
James paid $110 for the beef.
Quotes
"Question bootstrapping can be viewed as a form of multi-view augmentation in order to enable the transfer of the meta-knowledge."
"Backward reasoning questions are very helpful for LLMs to understand mathematical knowledge without memorization."