GEPA: Language-Based Reflection Outperforms RL in AI Prompt Optimization

Researchers introduce GEPA, a novel algorithm for optimizing prompts in complex AI systems. Unlike traditional reinforcement learning (RL), GEPA uses a language-driven evolutionary approach. An LLM analyzes its own performance—reasoning, tool usage, and feedback—to identify and fix errors. GEPA significantly outperforms RL methods, using far fewer system executions while achieving better results across various tasks. This highlights the potential of language-based self-reflection for efficient AI optimization.