What's this about?
StudyBench measures whether self-evolving models can learn from a small set of physics textbooks and transfer that knowledge from difficult exercises to Olympiad problems.
What exactly is the “guidance gap”?
Models perform much better when the relevant textbook material is placed directly in context than when self-training tries to absorb the same material into reusable skill. The difference measures failed learning rather than missing information.
Why is this worth noting?
Current methods improve application questions but rarely transfer, and all reach a compute plateau before exhausting the budget. The bottleneck appears methodological rather than simply more data or compute.