Yes, there are lots of schemes that have been thought of to help with quick polling of stations; in theory we can use the silence detector, but in practice we probably won't as it can be a bit of a pain due to the way the tuner is connected and userspace/kernel space latency issues.
Basically it'll probably interleave polls, with weaker AFs being polled less frequently. If it has a good candidate, it'll do a longer poll - long enough to pick up the 16-bit station ID from the RDS stream on the other frequency so it can check that the polled station is actually the same one the user is listening to. If the station being polled is the same, the switch is almost totally inaudible.
The tuner is very fast too, which means that even polls to weak stations are hard to hear.
Hugo