Minerva: Solving Quantitative Reasoning Problems with Language Models

Abstract

This article argues that large language models are still far from achieving human-level performance on quantitative reasoning tasks. The tasks require combining skills such as interpreting questions with natural language and mathematical notations, recalling relevant formulas, and performing numerical calculations. Minerva, a large language model trained on a wide range of scientific and mathematical texts, demonstrates significant performance gains on a variety of difficult quantitative reasoning tasks. By focusing on collecting problem-relevant data, training models at scale, and implementing best-in-class inference techniques, the model logra to achieve state-of-the-art results on STEM reasoning tasks. Future research on this line aims to develop models that understand mathematics more deeply and can help researchers and students in STEM fields. – AI-generated abstract.