GEPA: Reflective Prompt Evolution Can Outperform Reinforcement Learning
GEPA enhances LLM learning through language reflection, outperforming reinforcement learning methods.
GEPA leverages the interpretable nature of language to enhance LLM learning processes. This optimization tool integrates natural language reflection to learn high-level rules through trial and error. In six tasks, GEPA surpasses GRPO by an average of 6%, achieving up to 20% better results while utilizing 35 times fewer rollouts.