Language models can reason. They still fail at problems we already know how to solve, and whose solutions are written down.

The gap between what models can do and what they should be able to do is an open problem in the field. We believe we know how to close it.

More soon.