“Experiments with a 7B Jurassic-1 large model showed that LLMs struggle to reliably extract the correct arguments for basic arithmetic from verbal math problems.”