Comment by boxed

LLMs don't learn reasoning. At all. They are statistical language models. Nothing else. If they get math right it's because correct math is more statistically probable given the training data, it can't actually do math. This should be pretty clear from all the "how many Rs are there in strawberry" type examples.