Can't al LLM just detect a mathematical reasoning task then produce a formula (not even display it in the production mode) to invoke on an external service engineered for formal logical and mathematical computations?
The problem here is more serious than mathematics: the quantitative reasoning itself is highly unreliable.