So, let's assume we can formulate a way to store the ray collision positions in a vector 3 format. If we can do this, then we can theoretically create a shadow map with semi-accurate blurring using 2D effects with some added 3D accountability by using blur intensities based on a distance based...
I was actually just suggesting this for implementation in case of lower-end machines so that they aren't murdered mercilessly by extra raycasts. However, if we could store a depth map for shadow-casting objects as well we could potentially have a blur intensity curve based on that. I'll draw up...
Well it doesn't have to be simple, but yes I was planning on having it applied in 2D to give the illusion of soft shadows. This would of course be for those who cannot spare the horse-power for multiple ray samples.
As far I'm aware if the ray traced shadows render to a shadow map we can apply most filters. I sadly don't think that PCF can be applied without the being able to give the distance of the objects from the cast point. At least traditionally.
As far as softening of the shadows go, wouldn't it make more sense to do this post-process instead of taking exceedingly-large samples? This is how it is done in most games.