Multi-step reasoning grounded in visual information