I couldn't find the original paper, but I did find a shorter one, that, too, on Human Power, as an answer to Rohloff. This is what the paper concludes:
Originally Posted by Human Power contributing scientist
To summarize, we are reasonably
confident that the rank order between
transmission efficiencies that we found
would not change appreciably as load is
varied within a normal range. In other
words, transmissions should rank about
the same at either low or high loads. We
feel that the loads we tested under are
typical of the actual conditions under
which hub gears are used and represent a
reasonable average efficiency. In our article
we therefore concluded that hub
gears are about 2% less efficient that derailleur
transmissions under typical field
conditions. We see no reason to change
that conclusion.