In the good old days of Intel Pentium IIIs and AMD Athlons (Before Phenom), lapping the CPU die and heatsink contact patch made a LOT of difference. Thinning the die made for less insulating silicon and greater heat transfer, and the flat and polished contact area meant there was more direct contact and less "Through heat paste" contact which, while good, wasn't as good as a direct transfer. This is why the old methods of heat paste application (razor thin amounts in specific areas) were used - Your heatsink would ideally touch the die directly, with as little paste in the way as possible, and the paste was only there to fill in the slight imperfections in the die and HS base.
These days, with Integrated heat spreaders, there is no appreciable difference. If you are brave enough to de-lid your CPU you might get some bonuses, but that's not what's being asked or tested here. Theoretically the IHS should spread the thermal load out (hence the name) meaning contact area is not nearly as critical (Compare the size of contact area of an IHS with the size of the die underneath). Losing 1 or 2 degrees per square centimeter does next to nothing when you have 5 times the area, and the differences get lost in the margins of error. And generally speaking the "Weak point" in the transfer chain is the IHS material - It doesn't matter how good your contact is, how well performing your paste is, or how perfect your interface is, your thermal transfer performance is limited by that of the IHS. So even if your thermal paste transfers 10 degrees per cubic milimeter, if the IHS material only transfers 2, you're going to see no benefit.
Also, with Intel especially, you can't get the same mounting pressures that you used to, as pin-based sockets have the plastic of the retention mechanism against which to brace and spread the load of down-pressure from the HS, whereas the LGA style puts all that load on the pins themselves.
Of course, I am more than willing to be wrong about this, and if you can go ahead and prove this conjecture wrong, please do so.