Reading the page, it sounds like they're talking about the "torrentfs" mounted filesystem component, where read requests will download just that part of the file.
So theoretically, you could open the archive in a GUI, extract the file you want, and the only part it would download would be the archive header/file listing and the part of the archive corresponding to the file you wanted.
Of course, this wouldn't work for tarballs or solid RARs, but for regular archives it could conceivably work.
