Building Reflective Prompt Optimization with GEPA: Multi-Component Prompts, Structured Feedback, and Held-Out Validation

- GEPA is here to fix your LLM’s arithmetic, because apparently, teaching a bot to do math is the new DeFi yield farming. We take a weak prompt, feed it to GEPA, and let it evolve via structured feedback until it stops hallucinating sums. It’s like watching a moonboy finally learn that 2+2 doesn’t equal “to the moon.” The framework optimizes both instructions and output formats, ensuring the model doesn’t just guess but actually calculates. It’s deterministic, rigorous, and far less volatile than your average altcoin portfolio. Stop relying on hype; start relying on held-out validation sets. If you can’t optimize your prompts, you’re just gambling with tokens.