GEPA (2025)

GEPA (Genetic-Pareto) is a prompt optimizer introduced in a 2025 academic paper (accepted to ICLR 2026 as Oral) that uses natural language reflection to learn high-level rules from trial and error. It samples trajectories from AI systems containing one or more LLM prompts, reflects on them in natural language to diagnose problems, propose and test updates, and combine complementary lessons from the Pareto frontier of its attempts. Across six tasks, GEPA outperforms the reinforcement learning method GRPO by 6% on average and up to 20% while using fewer rollouts, and surpasses the leading prompt optimizer MIPROv2 by over 10%. The method also shows promise as an inference-time search strategy for code optimization. Code is released open-source.