“We don’t want to create a language model that just talks like a human,” said Yuhuai (Tony) Wu of Google AI. “We want it to understand what it’s talking about.” Wu has co-authored two interesting and recent papers that explore two different approaches to teaching AI models Math.
Image Credit: Valentin Tkach for Quanta Magazine
For one paper, Wu and a team of academic and industry researchers worked on teaching large language models to auto-formalize (translate mathematical proofs into the “formal language” computer code). “They can do all these different tasks with only a few demonstrations,” said Wenda Li, a computer scientist at the University of Cambridge and a co-author of the work.
The second work sees Wu and a team of other Google researchers training Minerva, another generally-trained large language model on mathematical material like arxiv.org pages, to not only recognize natural language Math problems, but also solve them. To control for the fact that the system could be repeating answers from its massive training datasets rather than solving the problem itself, the researchers assigned Minerva Poland’s 2022 National Math Exam, which came out after Minerva’s training data was set. The system got a 65%.