I was just thinking - since GPT-4o and Sonnet are closed models, do we know that this method was not already used to train them? And that Reflection is simply finding a path for greater improvements than they did. Llama 3.1 apparently didn't improve as much. It's just a thought though.