"complex reasoning is perhaps the most interesting feature of these models right now and it is unfortunately mostly absent from this survey"

Bingo. It is also the hardest to quantify; it's one of those "I know it when I see it" sort of behaviors. It is easy to imagine how one might harness that ability to reason to solve all sorts of problems, including (but certainly not limited to) improving benchmark performances. I think that is what has a lot of people excited.