I've looked at this paper. While the results are indeed pleasing, I find the algorithm cumbersome and non-intuitive. There are other means to derive soft shadows and penumbrae. Im not sold that this method is highly scalable, though i'd need to look at the paper more closely.
I believe standard shadow mapping is somewhat better suited for the next Conitec Engine, with the freedom for the developers to use their own shaders to filter the shadows in whichever way they want.