gepa-ai/gepa

Optimize prompts, code, and more with AI-powered Reflective Optimization

What it solves

GEPA (Genetic-Pareto) is a framework designed to optimize any system that relies on textual parameters—such as prompts, code, agent architectures, and configurations—against a specific evaluation metric. It addresses the limitation of traditional optimizers that only know if a candidate failed; GEPA identifies why it failed by analyzing execution traces to drive targeted improvements.

How it works

GEPA uses a combination of LLM-based reflection and Pareto-efficient evolutionary search:

  1. Selection: It picks candidates from a Pareto frontier (those performing best on different subsets of tasks).
  2. Execution: Candidates are run on minibatches, and full execution traces (error messages, logs, profiling data) are captured.
  3. Reflection: An LLM analyzes these traces to diagnose the root cause of failures.
  4. Mutation: The LLM generates an improved version of the candidate based on these diagnostic lessons.
  5. Acceptance: Improved candidates are added to the pool and the Pareto front is updated.

It also utilizes "Actionable Side Information" (ASI)—diagnostic feedback that acts as a text-based analogue to a gradient—and can merge the strengths of two Pareto-optimal candidates.

Who it’s for

  • AI Engineers: Who need to optimize complex prompts or agentic workflows without manually guessing changes.
  • Systems Researchers: Looking to optimize non-prompt text artifacts like scheduling policies or code.
  • DSPy Users: Who want to integrate automated optimization into their AI pipelines via dspy.GEPA.
  • Agent Developers: Who want to provide their agents with the ability to self-optimize via the GEPA Agent Skill.

Highlights

  • Efficiency: Claims to be 35x faster than RL (e.g., GRPO) and significantly cheaper than using top-tier models like Claude Opus.
  • Versatility: Can optimize prompts, agent architectures, RAG configurations, and even SVG graphics.
  • Broad Integration: Integrated into major tools like MLflow, Pydantic AI, Comet ML Opik, and Google's Gemini Enterprise Agent Platform.
  • Extensible: Provides a GEPAAdapter interface to plug into any system, with built-in adapters for LangChain, MCP, and RAG.

Related

  • Project
    EvoMap/evolver

    A GEP-powered self-evolution engine for AI agents that transforms runtime logs and experiences into auditable, reusable evolution assets to replace ad hoc prompt tweaking.

  • Project
    microsoft/LMOps

    A research initiative focused on fundamental technologies to improve LLM prompting, inference speed, and context handling for building AI products.

  • Project
    generative-computing/mellea

    A Python library for building predictable AI workflows by turning typed functions into structured LLM calls with automatic validation and retries.

  • Project
    Optima-CityU/LLM4AD_Next

    LLM4AD Next is a Python toolkit that turns a natural‑language problem description into a full, runnable evolutionary‑algorithm search project. An interactive LLM consultant generates the code skeleton, configuration, and evaluator, then built‑in evolutionary methods (IslandGA, Diverse Island GA, EoH, etc.) automatically evolve the generated algorithm. It also offers an advisor to suggest evolution‑ready code blocks, long‑term memory support, Docker images, and an online demo.