I haven't dug into QAT deeply, better recovery is my understanding as well, and also that it is out of reach for most people because you have to train a model to back prop errors based on estimated error under quant.

Hopefully more of the lab releases are trained under QAT so we can all benefit.