Basically, instead of marking live data, then sweeping all unmarked data, you just mark live data, then the next time a slot is requested, you step the allocator over the heap until it finds the next unmarked cell and re-use that. If most cells are dead, this happens very quickly. If you have to sweep over too many live cells (left deliberately vague), you grow the heap.
I have a pretty straightforward implementation of a lazy mark-sweep collector here (in C): https://github.com/silentbicycle/oscar
Basically, instead of marking live data, then sweeping all unmarked data, you just mark live data, then the next time a slot is requested, you step the allocator over the heap until it finds the next unmarked cell and re-use that. If most cells are dead, this happens very quickly. If you have to sweep over too many live cells (left deliberately vague), you grow the heap.