inter-layer prediction

[seaphone]

I recently study the H.264/AVC extension SVC, and cann't figure out how inter-layer prediction works?
I mean the current MB is predicted from the associated 8x8 MB in the lower resolution frame or is predicted from the reference of the associated 8x8 MB in the lower resolution frame? If it is the later one, where does the residual signal come from? from the upsampling of the one of the lower resolution frame?
as I know for the current MB, it says that it should upsampe the assocaited 8x8 MB in the lower resolution frame, so this upsampled MB replace the current one? or other purpose?

Thanks,
seaphone