AI summary: Researchers from UW Madison have been exploring how large language models like GPT-3/4, PaLM, and LaMDA can learn fundamental mathematical operations. They found that factors such as data format and size, model size, pretraining, and prompting style all play a role. The study also revealed that these models struggle to generalize beyond the training digit lengths, suggesting they learn arithmetic more as a mapping function rather than a flexible procedure. The findings provide insights into the rapid development of arithmetic capabilities in these models.
