To me it looks like one "throw" shorts the +/- inputs together and the
other attaches the "normal" feedback network.
In the shorted condition you are applying VIN=0, an ideal condition. If
you then "corrected" the amplifier output to null, by some sampled
feedback (summing and closing loop) you could take out all of the
front end error before and including the op amp's input offset. Then
you flip back to "normal" while keeping the null-cal and you have
(potentially) improved the signal chain DC accuracy a lot (at cost of
complexity, additional classes of noise, etc.).