The principles of our approach can be boiled down to a few key ideas:

1. Structure the problem such that it optimally resembles the underlying LLM's training data
2. Enable long-horizon work on the problem so a model can continue to build on its past work
3. Precisely structure the criteria of the solution and give the model a way to measure that

They are sparse on the details but it sounds like their model transforms problems of type Unkown into problems of type Known the best it can, then leverages its training on type Known problems to solve type Unknown problems. This is something that’s done a lot in computation theory (probably other domains of math too).