GEPA Reflective Optimization Beats GRPO With Just 3 Samples

GEPA, a genetic-Pareto based reflective prompt optimization framework, outperforms GRPO with just 3 samples versus 25,000 rollouts and reaches 89.5% on ARC-AGI. Its author's talk drew wide recommendations from practitioners building agent harnesses.

2026-09-27 ~ 2026-09-27 · 2 related posts

• GEPA reflection beats RL: 3 examples outgain GRPO's 25k rollouts, ARC-AGI to 89.5% — AI Engineer · 2026-09-27
• GEPA talk draws praise: optimize your AI harness prompts with it — CShorten30 · 2026-09-27