Why on-policy exploration makes RL fine-tuning of reasoning models work

Original: Reinforcement Learning for Hierarchical Reasoning Rewards: Minimax-Optimal Rates with Transformers

Source: arxiv.org ↗