Comments from Nick:
It's very interesting to see, in the Zhu et al. paper, that the kind of structural correspondence generated depends on the reward function. That's what my framework would predict.
Reply from AG: Hurrah! I think this is a fundamental insight.