Comment by letmevoteplease

This study appears to be evidence against that: the model failed the benchmark but arrived at the correct answer.