mid-training

This paper investigates how using diverse self-generated data during mid-training improves the effectiveness of Reinforcement Learning in Large Language Models, particularly for reasoning tasks.