GEPA: REFLECTIVE PROMPT EVOLUTION CAN OUTPERFORM REINFORCEMENT LEARNING

Summary from the TLDR AI newsletter

Traditional RL approaches need thousands of expensive trial-and-error runs to learn from simple success/failure scores. GEPA uses natural language reflection to diagnose what went wrong and evolve better prompts through a Pareto-based selection strategy. GEPA (Genetic-Pareto) dramatically outperforms both reinforcement learning methods like GRPO and leading prompt optimizer MIPROv2 while requiring up to 35× fewer training attempts.