Most attempts to make language models better at hard problems either pack more knowledge into the model or hand it relevant passages fetched in advance. A preprint from Shanghai AI Laboratory and university collaborators tries a third approach: teach the model to study. Given a maths exercise and the textbook it came from, with the worked solution removed, the model has to find the right chapters, read the definitions and theorems, and then write a proof. The authors call this self-learning, and it matters because it is closer to how a person tackles a subject they have not met before.
The setup has two parts. BookNavigator is a tool that lets the model look at a table of contents, jump to sections and read them while it reasons. Training then uses reinforcement learning, with rewards for a correct final proof as judged by another language model against the hidden reference solution, plus two finer signals: a bonus for consulting the book that fades over training, and credit for individual tool calls that actually helped. The dataset holds 5,183 proof problems drawn from 158 advanced mathematics textbooks.
The headline numbers are large. Intern-S2-Preview-35B solved 51.5% of problems before training, actually worse than the 55.7% it scored with no book at all. After training it reached 79.5%, 11 points ahead of the same model trained without the textbook and 6.4 points ahead of the best method that simply retrieved the most relevant sections up front. On textbooks and subjects held out from training it scored 85.2% and 86.7%. One finding stands out: training on correctness alone also raised accuracy, to 74.9%, but the model largely stopped opening the book, with tool use falling from 36.9% of attempts to 6.7%. The exploration bonus was needed to keep it reading.
This connects to a wider shift in how agents are built. Coding agents and research assistants already decide what to read before acting, and the results suggest that a model can be trained to get better at that reading step rather than only at answering.
There are reasons for caution. The paper is a preprint and not peer reviewed. The biggest gains are on a task the authors designed, and proof correctness is graded by a language model, not human mathematicians, so some error in the scoring is likely. The improvement on outside benchmarks is much smaller, 1.7 points on BrowseComp-Plus and 2.3 on CL-bench, and every trained version did worse than the base model on one CL-bench category, following procedures step by step. Those modest transfer figures are the fairer guide to what the method offers beyond textbook maths.