Google offers Minerva the first AI capable of solving math-related problems step-by-step.
Minerva is a built-in PaLM ( Pathways Language Model ) in which additional training was added, consisting of the 118 GB of scientific research articles from arXiv and webpages that include mathematical expressions written in LaTeX and MathJax as well as other formats.
The model has been taught the ability to “converse using standard mathematical notation,” according to researchers. In other words, the process is like other language models that generate various solutions, and Minerva gives probabilities for different results.
Each answer arrives (almost always) with the same effect, with distinct stages. The way the model works is to employ majority voting to pick the most commonly-used answer and then give the answer as the final one.
According to Google, Minerva is wrong now and then; however, his errors can be “easily interpretable.” According to the research, “about half are calculation errors, and the other half are reasoning errors, resulting in nonlogical reasoning.”
To better solve math-related problems, Minerva also uses contemporary methods of prompting and grading. They include majority voting and chains of thought, also known as scratchpads. Like many language models, Minerva offers probabilities to various possible outcomes.
Minerva generates a range of possible answers by randomly sampling every viable product while answering the question. Although these techniques’ stages differ, they usually produce the same result. Minerva chooses the most commonly used answer as the answer, making use of majority voting.