Beating RL With Reflection: GEPA and Optimize Anything — Lakshya A. Agrawal, GEPA
The presentation introduces reflective optimization, a method for self-improving AI agents and models using textual feedback, addressing the critical bottleneck of sample efficiency. The core techniques, GEPA (Generative Prompt Enhancement Algorithm) and Optimize Anything, allow systems to learn from the full trace of an agent's execution (including chains of thought, tool calls, and errors), rather than just a single reward score. This enables significant performance gains (e.g., lifting Gemini Flash on ARC-AGI from 32.5% to 89.5%) by optimizing prompts, agent architectures, and even code/policies, all while maintaining a Pareto pool of candidate solutions to avoid local optima.
Key takeaways
-
GEPA Outperforms Traditional RL
4:50
GEPA achieved twice the performance gains of GRPO (a leading RL technique) using only three data points and one round of reflection, compared to GRPO's 25,000 rollouts. This demonstrates superior sample efficiency. (4:50)
-
Optimization Beyond Prompts: Optimize Anything
10:35
The reflective optimization process can be applied to any system that can be expressed as text and scored, including entire agent harnesses (Python/JavaScript files), CUDA kernels, or cloud scheduling policies. (10:35)
-
Significant Performance Gains Demonstrated
13:35
GEPA was able to optimize an agent for ARC-AGI, boosting Gemini Flash's accuracy from 32.5% to 89.5%. It also improved GPT-5 mini's Go issue resolution from 24% to 93%. (13:35, 15:10)
-
Pareto Pool Strategy Beats Simple Loops
9:00
Maintaining a Pareto pool of candidates ensures a more balanced search process, preventing the model from getting stuck in local optima, which is a major advantage over simple iterative loops. (9:00)