UC Berkeley's GEPA Evolves LLM Prompts Without GPU Fine-Tuning

Researchers at UC Berkeley's Sky Computing Lab have developed GEPA (Genetic-Pareto Evolutionary Prompt Adaptation), an accepted oral paper at ICLR 2026 that optimizes large language model pipelines without requiring GPU resources. The method applies a genetic algorithm approach where an LLM iteratively critiques its own failures and generates improved prompt variants across a population. GEPA selects the best-performing prompts using Pareto-optimal scoring across accuracy, cost, and latency over multiple generations. It is designed to work with any black-box LLM, including GPT-4o, Claude, and Gemini, and is available as a drop-in optimizer called dspy.GEPA for DSPy pipelines. The framework also supports multi-step compound AI systems by tracing execution to identify and target the specific pipeline component responsible for errors.