Darwin Gödel Machine: Self-Improving Agents, Reflect, Retry, Reward: Reflection as a reward signal, Managing entropy in RL, 80/20 rule: High-entropy token for effective RL, Skywork Open Reasoner 1.