Researchers question AI’s ‘reasoning’ ability as models stumble on math problems with trivial changes