This would be an option for users who are above the RAM requirements for SL and have fairly fast processors, SL for me isn't THAT processor intensive really for example (multi-processor support would still rock btw!)
Basically it is as it sounds, when you download an object (or an object's shape update), you compute several different levels of detail for that object (lowest, low, med, high and highest say) and store them all in memory. Now when your frame-rate drops all LOD has to do is swap the appropriate models, rather than doing what all evidence seems to suggest it does atm, and recompute EVERYTHING you can see.
Currently LOD causes more lag for me than it solves, if I were on highest quality all the time then occlussion culling would cope with fine to keep my FPS fairly consistent, currently LOD causes spikes an awful lot.
This would possibly be joined with fewer levels of detail between lowest and highest, but overall more efficient I think.