Comment by baxfedi

A model that understands it might be wrong with the token it predicted after “a + b = “ and is open to double checking that with a calculator/script is a more intelligent model that is less likely to hallucinate confidently.