TL;DR of the paper, generated automatically.

Paper on self-learning: LLMs read textbooks on the fly to solve proofs, trained with reinforcement learning.

• Method: BookNavigator interface, fine-grained rewards, reference solutions withheld.
• Result: proof accuracy on Intern-S2-Preview-35B rose from 51.5% to 79.5%.
• Other tests: gains of 1.7 and 2.3 points over the base model on BrowseComp-Plus and CL-bench.