What would be the result if the task was given to multiple models? Instead of alloying them together and switching between models in the same chat, just let the models try to complete the task in their own isolated context, and use the result that completed it successfully?

I would say that that’s at least something the alloying should be benchmarked against, which I didn’t find in the article.