GEPA: Reflective Prompt Evolution Can Outperform Reinforcement Learning

Lakshya A Agrawal, Shangyin Tan, Dilara Soylu, Noah Ziems, Rishi Khare, Krista Opsahl-Ong, Arnav Singhvi, Herumb Shandilya, Michael J Ryan, Meng Jiang, Christopher Potts, Koushik Sen, Alexandros G. Dimakis, Ion Stoica, Dan Klein, Matei Zaharia, Omar Khattab

Read on arXiv

Type: paper
Venue: arXiv / ICLR 2026
Year: 2026
Source: arxiv
Access: free
Language: en
Added: 2026-08-14T19:15:00Z
Verified: 2026-08-14T19:15:00Z

Summary

GEPA (Genetic-Pareto; ICLR 2026 Oral). For a compound LLM system, sample traces, reflect with an LM on textual feedback (compiler errors, missed gold docs, constraint fails), mutate one module’s prompt, keep the candidate if the minibatch improves, and parent-sample from the instance-wise Pareto front instead of always mutating the global best. On Qwen3 8B across HotpotQA/IFBench/HoVer/PUPA/AIME-2025/LiveBench-Math: +9.62 pp vs baseline vs GRPO +3.68 pp at 24k rollouts, using ~4–35× fewer rollouts (aggregate ~3936). Also beats MIPROv2; GPT-4.1 Mini aggregate +12.19 pp. Code MIT gepa-ai/gepa; first-class dspy.GEPA. Discord post is Akshay Pachaar’s April 2026 X article “How to Beat GRPO Without Touching Model Weights” (DSPy cookbook; claims Berkeley beat GRPO by 10 points with 35× fewer rollouts).

Keywords

gepa · dspy · prompt-optimization · grpo · pareto · iclr-2026 · berkeley · x

Topics

prompt optimization, compound AI systems, DSPy

Research notes

Primary: arxiv abs 2507.19457. Discord/X article https://x.com/akshay_pachaar/status/2049916107923034300 via fxtwitter (empty tweet body; article id 2049896113206149120). Code https://github.com/gepa-ai/gepa (MIT). Docs https://dspy.ai/api/optimizers/GEPA/overview/. Artifact https://github.com/gepa-ai/gepa-artifact. Optimizer/code, not a hosted corpus, so no datasets_local row.