How to proceed depends on what exactly you want the refinement to do.  If the structure is already approximately correct with no major conformational changes, here are some basic guidelines:

1) Use tight restraints.  The weight optimization doesn't work very well in this resolution range (at least in my hands), so setting very low values for wxc_scale and wxu_scale (which control the weighting of the X-ray term relative to the geometry and B-factor restraints, respectively) is a good place to start.  I would try values of wxc_scale=0.025 and wxu_scale=0.05, but you may need to experiment with this.

2) Use NCS restraints if you have multiple copies in the ASU.  (The torsion restraints are usually sufficient - and much easier to use - but it might be worth trying cartesian restraints too.)

3) Use a reference model if you have a good high-resolution structure available (this can be just part of the structure).

4) Real-space refinement is probably not going to work.

5) Try using a very conservative refinement strategy at first - for instance, rigid-body and group B-factor refinement.  The latter should work better now (in the latest nightly builds, and imminent 1.9 release) since Pavel added restraints on neighboring groups.  You can also try refining TLS for domains.  My own experience was that individual coordinate and B-factor refinement worked best - the latter is somewhat controversial but with tight restraints it ends up not being very different from refining an overall B-factor.  I also found it helpful to reset all B-factors to 20 at the start of refinement to remove bias from the starting model.

6) Be very paranoid about the 2mFo-DFc map - at this resolution it will look almost exactly like your model.  Make lots of omit maps, and if you have any heavy scatterers (metals, etc.) those will be very helpful in confirming that you have the correct structure.

If you actually have to do a lot of rebuilding, it is considerably more difficult; DEN or Rosetta refinement may be helpful.

-Nat