Berkeley’s GEPA Optimizes AI Prompts Without Weight Updates, Cutting Training Costs 35-Fold

A research team at the University of California, Berkeley has introduced GEPA, a method that teaches AI systems new tasks by optimizing prompts without fine-tuning model weights. The approach can reduce the computing resources and costs required to train large language models, offering a lighter and easier-to-deploy alternative to conventional reinforcement learning.

GEPA has been accepted as an oral paper at ICLR 2026. The researchers had a “reflection LLM” review training logs, analyze errors and repeatedly rewrite prompts. Experiments showed that GEPA outperformed conventional reinforcement-learning methods while requiring only about 1/35 of their training cost—a roughly 35-fold reduction.