Beating RL With Reflection: GEPA and Optimize Anything — Lakshya A. Agrawal, GEPA

Reflective optimization, or GEPA, addresses the sample inefficiency of traditional AI training by leveraging textual feedback to self-improve prompts, agents, and models. Unlike reinforcement learning that relies on binary rewards and gradient updates, GEPA uses an evolutionary loop with Pareto-based candidate selection to iterate on text-based artifacts. This approach extracts diagnostic value from intermediate outputs, tool calls, and error messages, allowing for rapid behavior changes without requiring massive datasets. The "Optimize Anything" API extends this methodology to diverse domains, including code optimization, cloud scheduling, and agent architecture design. By treating any text-expressible parameter as an optimization target, GEPA achieves significant performance gains—often exceeding 2x to 7x improvements—while reducing execution costs and token usage. This paradigm enables models to refine their own instructions and logic, effectively automating complex engineering tasks that previously required extensive manual intervention.