Hi Julian: I totally agree that vast body of literature on evolutionary optimization is relevant; in fact GEPA (arxiv.org/abs/2507.19457) and AlphaEvolve (arxiv.org/abs/2506.13131) are instances of evolutionary optimization. To me, the biggest open question though is that of g…
- GEPA outperforms GRPO by 6% on average and up to 20% across six tasks, while using up to 35x fewer rollouts.
- The same method beats leading prompt optimizer MIPROv2 by more than 10%, including a +12% accuracy gain on AIME-2025.
- Authors argue natural-language reflection on trajectories is a richer learning signal than sparse scalar reward gradients.