paper

Imagine you have a really smart robot and a less smart one. The less smart robot is supposed to teach the smart one how to do things.

Normally, you’d think the smart robot can only learn as much as the less smart one knows.

But, surprisingly, the smart robot can sometimes learn to be even smarter than its teacher. However, it can’t learn everything it’s capable of just from the less smart robot. To really help the super smart robot reach its full potential, people need to find better ways to teach it.

This is like trying to teach a really smart student with a book that’s too easy for them; they can learn something, but not everything they could.

While this approach shows that stronger models can indeed be guided and improved using weaker models, it also highlights a limitation. The full capabilities of the stronger models are not entirely realized through this method alone. This indicates that while weak-to-strong generalization is a promising direction, additional techniques and innovations are needed to fully harness and align the capabilities of superhuman AI models.