seems like they should be good at finding solutions which require leveraging partial information, and generating a large quantity of ideas (which can contain quality)

But the optimal thing might be to have a number of more typical/boring people to filter their ideas through, and that might be hard to test experimentally