Reproduce “GEPA: Reflective Prompt Evolution Can Outperform Reinforcement Learning” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Lakshya A Agrawal, Shangyin Tan, Dilara Soylu, Noah Ziems, Rishi Khare, Krista Opsahl-Ong, Arnav Singhvi, Herumb Shandilya and 9 more
GEPA adapts language-model workflows by having a reflection LLM inspect execution traces and evaluator feedback, revise prompts, and preserve candidates with complementary strengths through Pareto selection. On IFBench with Qwen3 8B, it scores 38.61% versus GRPO’s 35.88% after 678 versus 24,000 rollouts. GEPA holds model weights fixed; results cover tested tasks and settings, not universal superiority.
Key insights
- Rich, interpretable natural language traces generated during LLM execution provide a far more potent and diagnostic learning signal than sparse scalar rewards used in traditional reinforcement learning.
- Advanced instruction optimization alone, driven by reflective prompt evolution, can be more effective for modern LLMs than methods that jointly optimize instructions and few-shot examples.
- A Pareto-based multi-objective evolutionary search is crucial for robust prompt optimization, enabling the discovery of diverse high-performing solutions and preventing stagnation in local optima.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.