Obsidian Metadata
| type | source |
| url | https://arxiv.org/abs/2507.19457 |
| authors | Lakshya A Agrawal, …, Omar Khattab |
| published | 2025-07-25 |
Source: GEPA: Reflective Prompt Evolution Can Outperform Reinforcement Learning
URL: https://arxiv.org/abs/2507.19457 · code: github.com/gepa-ai/gepa
GEPA (Genetic-Pareto) samples trajectories (reasoning, tool calls, outputs), reflects in natural language to diagnose problems, proposes and tests prompt updates, and combines lessons from a Pareto frontier of attempts. It beats GRPO by 6% on average (up to 20%) with up to 35× fewer rollouts, and beats MIPROv2 by >10% (+12% on AIME-2025).
In the FAQ it’s the example prompt-tuning algorithm for judges: it optimizes against the dev set of human labels. Use it only after manually inspecting disagreements, and keep the test set out. Cheap classifiers make it far cheaper. → Prompt Optimization, Judge Validation