What is GEPA?

# Conceptual flow from the paper
def reflective_mutation(current_prompt, failed_examples):
    """Use LLM reflection to intelligently improve prompts."""

    # Step 1: Analyze what went wrong
    reflection = reflection_llm(f"""
    Current prompt: {current_prompt}

    Failed on these examples: {failed_examples}

    Analyze: What specific issues caused these failures?
    What patterns do you see? What should change?
    """)

    # Step 2: Generate improved variant
    new_prompt = reflection_llm(f"""
    Original prompt: {current_prompt}
    Analysis: {reflection}

    Generate an improved prompt addressing these issues.
    """)

    return new_prompt

Performance on different metrics:

Accuracy ↑
    |
    |    C (kept - strong on accuracy)
    |
    |  A (kept - balanced)    D (discarded - dominated)
    |
    |           B (kept - strong on helpfulness)
    |______________________________________→ Helpfulness

Given the fields question, summary 1, produce the fields query.

You will be given two input fields: question and summary 1.
Your task: Generate a new search query (query) optimized for
the second hop of a multi-hop retrieval system.

• The original user question is typically complex and requires
  information from multiple documents to answer.
• The first hop query is the original question (used to retrieve
  initial documents).
• Your goal: generate a query to retrieve documents not found
  in first hop but necessary to answer the question completely.

Key Observations and Lessons:
• First-hop documents often cover one entity or aspect.
• Remaining relevant documents often involve connected or
  higher-level concepts mentioned in summary 1 but not
  explicitly asked in the original question.
• The query should target these missing, but logically linked,
  documents.

[... detailed strategies continue ...]

# How ATLAS calls GEPA (from optimize_teaching.py)
import gepa
from trainers.prompt_adapter import ATLASGEPAAdapter

# Create ATLAS adapter (handles evaluation)
adapter = ATLASGEPAAdapter(
    teacher_model="gpt-5",
    student_model="gpt-4o-mini",
    generation_config={
        "max_tokens": 512,
        "diagnostic_max_tokens": 100,
        "temperature": 0.7
    }
)

# Run GEPA optimization
result = gepa.optimize(
    seed_candidate={
        "teacher_adaptive_template": "You are a teacher...",
        "student_diagnostic_template": "Show your thinking...",
        "student_with_teaching_template": "Apply guidance..."
    },
    trainset=your_examples,
    valset=validation_examples,
    adapter=adapter,  # ATLAS evaluates via full protocol
    reflection_lm=gpt4_reflection,  # LLM for mutations
    max_metric_calls=40,
    candidate_selection_strategy="pareto"
)

optimized_prompts = result.best_candidate

Student baseline:   40 × 10 × gpt-4o-mini  ≈ $0.50
Teacher guidance:   40 × 10 × gpt-5        ≈ $4.00
Student enhanced:   40 × 10 × gpt-4o-mini  ≈ $0.50
Reflection LLM:     40 × gpt-4             ≈ $4.00
RIM scoring:        800 × Gemini Flash     ≈ $1.00
                                    Total  ≈ $10.00

@article{agrawal2025gepa,
  title={GEPA: Reflective Prompt Evolution Can Outperform Reinforcement Learning},
  author={Agrawal, Lakshya A and Tan, Shangyin and Soylu, Dilara and Ziems, Noah and Khare, Rishi and Opsahl-Ong, Krista and Singhvi, Arnav and Shandilya, Herumb and Ryan, Michael J and Jiang, Meng and Potts, Christopher and Sen, Koushik and Dimakis, Alexandros G and Stoica, Ion and Klein, Dan and Zaharia, Matei and Khattab, Omar},
  journal={arXiv preprint arXiv:2507.19457},
  year={2025}
}