Minerva

Minerva is Google Research's 2022 language model for quantitative reasoning, introduced in the paper 'Solving Quantitative Reasoning Problems with Language Models' on June 30, 2022. It starts from PaLM, at 8B, 62B and 540B parameter sizes, and continues training on a technical dataset built from arXiv papers and web pages carefully processed to preserve LaTeX and MathJax notation. The model generates step-by-step solutions in correct mathematical notation and uses few-shot prompting, chain-of-thought and majority voting at inference, with no external calculator.

The results were the best reported at the time: state of the art on MATH, GSM8k and the STEM subset of MMLU, and nearly a third of more than 200 undergraduate-level problems in physics, biology, chemistry and economics answered correctly. Google was candid about limits, noting Minerva sometimes reaches the right answer through incorrect reasoning steps. Minerva was a research demonstration, never released as weights or a product, so openWeights is false. It set the template for the reasoning-model wave that followed.