I also think that separating the semantic representation of the program from the "user interface" (syntax) is another interesting idea, something that came up again recently via this repost of Moon's thoughts on sexprs. There's another less know bunch of research in this area, done by the folks at the University of St. Andrews in Scotland called hyper-programming, although more recent papers refer to the approach as hypercode. This goes even further trying to accomplish model-view separation for programming languages.

I don't, however, think that serializing all this data as XML is going to prove to be a very workable solution.