Small Models Struggle to Learn from Strong Reasoners