How certain can we be that these improvements aren't just a result of Gemini 3 Pro pre-training on endless internet writeups of where 2.5 has struggled (and almost certainly what a human would have done instead)?
In other words, how much of this improvement is true generalization vs memorization?
In other words, how much of this improvement is true generalization vs memorization?