natural-language-reflection

GEPA is a prompt optimizer that uses natural language reflection to learn from trial and error, outperforming reinforcement learning methods like GRPO and MIPROv2 with up to 35x fewer rollouts across multiple tasks.