Incentivizing Self-Learning in Large Language Models via Reinforcement Learning

Researchers introduced a BookNavigator interface and reinforcement learning method to incentivize self-learning in large language models, enabling them to autonomously retrieve, study, and apply knowledge from long-form reference materials, raising Intern-S2-Preview-35B's proof accuracy from 51.5% to 79.5% on a textbook-based mathematical problem-solving task.