Vision Language Models Know Law of Conservation without Understanding More-or-Less