GEPA: Reflective Prompt Evolution Can Outperform Reinforcement Learning (arxiv.org)

Researchers have introduced GEPA (Genetic-Pareto), a novel prompt optimization technique that outperforms traditional reinforcement learning methods, notably Group Relative Policy Optimization (GRPO). GEPA utilizes the interpretable nature of language to enhance learning for large language models (LLMs) by sampling and reflecting on various trajectories, including reasoning and outputs, to diagnose issues and propose prompt improvements. This approach requires significantly fewer rollouts—up to 35 times less—while achieving an average performance improvement of 6% over GRPO, and in some cases, up to 20%. GEPA also surpasses MIPROv2, the leading prompt optimizer, by over 10%. This development is significant for the AI/ML community as it challenges the prevailing reliance on RL for adapting LLMs, suggesting that linguistic reflection can be a more efficient and interpretable learning method. By shifting focus from scalar rewards to high-level language reflections, GEPA opens new avenues for optimizing LLMs in various tasks, including code optimization. The release of GEPA's code further fosters collaboration and innovation in the field, setting a new standard for prompt optimization strategies.