[ambien_]

Regardless of the model, how can optimizing two metrics ever yield an improvement in the first compared to just optimizing it alone?

Or is the idea that, because the old model sucks, we need to optimize this intermediary metric (diversity) in the interests of the ultimate goal of raising-the-bar?

[lazyasciiart]

It could work better because we are so bad at actually optimizing for the first variable that we would get better results by not optimizing for it.