GEPA: Reflective Prompt Evolution Outperforms Reinforcement Learning in LLM Optimization
Key Takeaways
- ▸GEPA outperforms GRPO by 6% on average (up to 20%) while requiring 35x fewer rollouts—dramatically improving training efficiency
- ▸Natural language reflection enables more interpretable and effective learning compared to reinforcement learning policy gradients
- ▸Surpasses leading prompt optimizer MIPROv2 by 10%+, with +12% accuracy improvement on AIME-2025 benchmarks
Summary
A new research paper introduces GEPA (Genetic-Pareto), a prompt optimizer that leverages natural language reflection to improve large language models more efficiently than traditional reinforcement learning methods. The system samples trajectories—including reasoning, tool calls, and tool outputs—and reflects on them in natural language to diagnose problems and propose prompt updates, combining complementary lessons from the Pareto frontier of its own attempts.
Across six benchmark tasks, GEPA significantly outperforms GRPO (Group Relative Policy Optimization), a leading reinforcement learning approach, achieving 6% average improvements and up to 20% gains while using up to 35x fewer rollouts. The approach also surpasses MIPROv2, the previous leading prompt optimizer, by over 10%, with notable performance on the AIME-2025 benchmark (+12% accuracy). The method demonstrates additional promise as an inference-time search strategy for code optimization.
The researchers are releasing GEPA's code open-source, making the approach accessible to the broader AI research community. This work suggests that interpretable, natural language-based learning may be a richer medium for LLM adaptation than traditional policy gradients derived from sparse scalar rewards.
- Demonstrates practical value as an inference-time search strategy for code optimization tasks
- Code being released open-source for research community adoption
Editorial Opinion
GEPA represents an important paradigm shift in LLM adaptation, challenging the assumption that reinforcement learning is the optimal path for task-specific optimization. By harnessing natural language—the medium in which LLMs are most fluent—rather than forcing learning through sparse reward signals, this work opens a compelling direction for more interpretable and sample-efficient model tuning. The dramatic reduction in required rollouts (35x fewer) has significant implications for research accessibility and computational efficiency. If these results hold across diverse domains, this could reshape how the community approaches LLM customization.
