The problem with the current implementation:
The "gradual infill steps" option has a lot of potential to save huge amounts of filament and significantly speed up the print but unfortunately, as it's currently implemented, it's basically useless for all but very simple geometries (completely flat roofs). The problem is that the current algorithm is extremely simple and almost always tries to add new (higher density) infill lines in mid air.
Proposed solution:
The infill line should always be extended (from the currently computed span) in both directions to the object's inside walls or to the nearest previously printed infill lines. This would insure that it never starts or stops in mid-air, something that is currently basically guaranteed to happen.
Result:
Such simple modification would enable printing of large complex shapes that could be almost completely hollow, with next to no plastic or speed loss to infill while still providing extremely solid internal support for any possible roof geometries only in the exact places that it's needed.