If adaptivity puts more samples inside caustics, it will slow down rays/s (because we are doing expensive caustics lookups more), but the caustics do not get better - it is more samples reusing the same batch of photons, so no net improvements. If this is the case, it should be fairly straightforward to fix