Adding code to LLM training data is a known method of improving a model’s reasoning skills. But wouldn’t math, the basis of all reasoning, be even better?