That whole structured generation line of work looks promising. I hope someone else takes this and runs evaluations on other benchmarks. Curious to see if the results translate!
In addition to the "prompt consistency" and "thought-control" ideas mentioned in the post, I'm definitely curious how the performance is on more complex structured data (things like codegen).