I'm not sure it's question of prefer, virtual or (synthetic) instruments are really aimed at a different part of the market to bench instruments.
Historically test systems were created using a collection of bench instruments which were equipped with some kind of programmable interface, typically GPIB (HPIB) all bundled into a rack. This is all well and good, but tends to result in a system which has a large foot, issues around long term support and other difficulties, such as keep the firmware on all these differing systems up to date. In addition there's is effectively a lot of repetition of capabilities, in terms of signal conditioning, samplers, detectors etc.
What many people wanted (the DOD for example) was a system that would reduce the total cost of ownership, reduce develop and
deploy time, reduce the physical footprint, reduce the logistics effort in terms of decreased spares, support systems and training, provide greater flexibility through systems that are interoperable among U.S. and allied services and Improve the overall quality of testing. (Agilent Technologies. (2006) Using Synthetic Instruments in Your Test System).
In order to address this, companies like Agilent (formerly Hewlett Packard), National Instruments etc started to produce a range of instruments which leveraged off existing industry standards to create small form factor instruments typically in a card and cage form, VXI, PXI etc. These are not always complete instruments, many are effectively subsystems of an instrument, such that a system can be configured to be for example a spectrum analyser, or a network analyser via software control and switching of the signal paths, virtual instruments (HP's Modular Measuement Sysems is another example).
Now on a bench, in an RD lab, you probably don't really want something like this. You don't want to invest in a card cage and have all your test equipment effectively tied together, you might want your function generator and scope to be in two different places for example, very difficult if there in the same PXI frame. Further more, if you just wanted an Arb or scope, having to buy a mainframe to house it in and power it is a lot of additional expense. LXi, which is a new(ish) standard for programmable instruments avoids this in much the same way as GPIB did, in that it is predominantly a LAN port on a standard bench instrument, so the instrument can be part of a system or not, but that just gets you back to the huge test systems consisting of discrete bench instruments (though you can of course mix discrete and virtual instruments should you so choose).
So as I said at the begining the question isn't about prefer, it's about what are you going to be doing with it. If it's ATE machine on a production line or a support system on a trolley for an aircraft, you might prefer a synthetic or virtual instrument, if it's in the lab, almost certainly not.
Incidentally, some virtual instruments are created by simply (more or less) removing the screen and keypad from an existing bench instrument design and putting it in a smaller box, but curiously even though your getting less, there usually more expensive, go figure?
