Mid-Training with Self-Generated Data Improves Reinforcement Learning in Language Models

A mid-training stage using bootstrapped self-generated diverse reasoning data improves RL fine-tuning effectiveness in LLMs by broadening solution-path coverage before RL begins. The method addresses the known limitation of narrow training distributions degrading RL exploration. Teams running RLHF or GRPO pipelines on reasoning tasks may benefit from inserting this intermediate self-play data generation step.