The novel part of that paper was not merging the responses. The last model can, from the inputs, synthesize higher quality overall responses than any of the individual models.
It’s a bit surprising that works and your take on it is overly reductive largely because you’re wrong in your understanding of what it was doing.
I didn't just look at the implementation, I tried it as well. I was hoping it would work, but the aggregating model mostly either failed to properly synthesize a new response (merely dumping out the previous responses as separate functions) or erratically took bits from each without properly gluing them together. In every case, simply picking the best out of the four responses myself yielded better results.
Interesting, I've seen live demos working fairly well. I've also implemented something adjacent to the work and it works quite well too. I'm not sure why you had a hard time with it.
I am however working in a domain where verification isn't subjective so I know a good response from a bad response fairly easily. Things like this depend quite heavily on the model being used too in my experience.
It’s a bit surprising that works and your take on it is overly reductive largely because you’re wrong in your understanding of what it was doing.