GEPA: Reflective Prompt Evolution Can Outperform Reinforcement Learning
arxiv:2507.19457
Post-training · 2025-07-25
Complete cleaned Markdown · Source Markdown including bibliography
Read in parts
Contents
- GEPA: Reflective Prompt Evolution Can Outperform Reinforcement Learning
- Abstract
- 1 Introduction
- 2 Problem Statement
- Sample-Efficient Optimization.
- 3 GEPA: Reflective Prompt Evolution
- 3.1 Pareto-based candidate selection
- 4 Evaluation
- 5 Extended Applications of GEPA
- 5.1 GEPA For Inference-Time Search (Contd.)
- 5.2 GEPA for Adversarial Prompt Search (Contd.)
- 6 Related Work
- 7 Conclusion
- Appendix A Appendix Outline
- Appendix B Usage of Large Language Models
- Appendix C GEPA’s Reflection and Prompt Update Meta Prompt
- Appendix D GEPA Algorithm and Methodology Details
- D.1 Merge: System-aware crossover strategy for Compound AI optimization
- Appendix E Evaluation Setup (Contd.)
- E.1 Benchmarks, Reference compound AI systems, and Feedback Functions
- E.2 Models and Inference Parameters
- E.3 Costs
- E.4 Optimizers
- Appendix F Results and Analysis (Contd.)
- Appendix G Performance vs. Budget (Rollouts) Curves
- Appendix H Generalization Gap
- Appendix I Cost vs. Performance Analysis for optimized systems
- Appendix J GEPA Search Trees
- Appendix K Visualizing the Iterative Refinement achieved by GEPA
- K.1 Prompts at intermediate stages for PUPA
- Appendix L Examples of best prompts for every benchmark
- L.1 HotpotQA, GPT-4.1 Mini
- L.2 HotpotQA, Qwen3 8B
- L.3 IFBench, GPT-4.1 Mini
- L.4 IFBench, Qwen3 8B
- L.5 HoVer, GPT-4.1 Mini
- L.6 HoVer, Qwen3 8B
- L.7 PUPA, GPT-4.1 Mini
- L.8 PUPA, Qwen3 8B
- Appendix M GEPA generated prompts for kernel generation
- M.1 NPUEval: Kernel Code Generation for new hardware architecture
- M.2 KernelBench: CUDA Kernel Code Generation for NVIDIA GPUs
- Appendix N Number of reflection LM calls made by GEPA during optimization
- Instructions for reporting errors