I think the next generational leap will be consistent universal lighting. Whether that's achieved through RT or some other means doesn't matter. This gen, we have moments of lighting about as good as we can expect, but they are static and come with all sorts of shortcomings and inconsistencies.
Or in short, we'll have games where detractors won't be able to cherry-pick rare bad examples of the game looking poop.
For game experience, ML should be the greatest advance, though it may have to wait until next generation. ML might do wonders. And if foveated rendering with eye tracking ever happens, that'll be a game changer beyond anything else.