[PDG 451] GEPA: Reflective Prompt Evolution Can Outperform RL

Link to article: https://arxiv.org/pdf/2507.19457v1
Title: GEPA: Reflective Prompt Evolution Can Outperform Reinforcement Learning
Content: GEPA (Genetic-Pareto) is a new prompt optimizer for LLMs that leverages natural language reflection to learn from trial and error, rather than relying solely on policy gradients from sparse scalar rewards. The system samples trajectories from AI systems, reflects on them in natural language to diagnose issues and propose prompt updates, then combines complementary lessons from the Pareto frontier of attempts. GEPA achieves superior performance with far greater efficiency - outperforming GRPO by 10% on average (up to 20% on some tasks) while using up to 35x fewer rollouts, and beating the leading optimizer MIPROv2 by over 10%. This demonstrates that the interpretable nature of language can provide a much more effective learning medium for LLMs compared to traditional RL approaches.
Slack link: ml-ka.slack.com, channel: #pdg. Please join us -- if you cannot join, please message us here or to mlpaperdiscussiongroupka@gmail.com.

In the Paper Discussion Group (PDG) we discuss recent and fundamental papers in the area of machine learning on a weekly basis. If you are interested, please read the paper beforehand and join us for the discussion. If you have not fully understood the paper, you can still participate – everyone is welcome! You can join the discussion or simply listen in. The discussion is in German or English depending on the participants.