Optimize any text parameter — prompts, code, agent architectures, configurations — using LLM-based reflection and Pareto-efficient evolutionary search.
Website |
Quick Start |
Paper |
Blog |
Discord
What is GEPA?
GEPA (Genetic-Pareto) is a framework for optimizing any system with textual parameters against any evaluation metric. Unlike RL or gradient-based methods that collapse execution traces into a single scalar reward, GEPA uses LLMs to read full execution traces — error messages, profiling data, reasoning logs — to diagnose why a candidate failed and propose targeted fixes. Through iterative reflection, mutation, and Pareto-aware selection, GEPA evolves high-performing variants with minimal evaluations.
If you can measure it, you can optimize it: prompts, code, agent architectures, scheduling policies, vector graphics, and more.
Key Results
|
|
| 90x cheaper |
Open-source models + GEPA beat Claude Opus 4.1 at Databricks |
| 35x faster than RL |
100–500 evaluations vs. 5,000–25,000+ for GRPO (paper) |
| 32% → 89% |
ARC-AGI agent accuracy via architecture discovery |
| 40.2% cost savings |
Cloud scheduling policy discovered by GEPA, beating expert heuristics |
| 55% → 82% |
Coding agent resolve rate on Jinja via auto-learned skills |
| 50+ production uses |
Across Shopify, Databricks, Dropbox, OpenAI, Pydantic, MLflow, Comet ML, and more |
“Both DSPy and (especially) GEPA are currently severely under hyped in the AI context engineering world” — Tobi Lutke, CEO, Shopify
Installation
pip install gepa
To install the latest from main:
pip install git+https://github.com/gepa-ai/gepa.git
Python 3.15 is supported by the core package and the dependency-free
gepa[dspy] extra on main; a release containing this support is needed
before installing from PyPI. Optional integrations have their own dependency
requirements. Use Python 3.10–3.14 for gepa[full] and gepa[confidence]
until their dependencies support Python 3.15. Callable models and evaluators
do not require these extras.
Quick Start
Simple Prompt Optimization
Optimize a system prompt for math problems from the AIME benchmark in a few lines of code (full tutorial):
import gepa
trainset, valset, _ = gepa.examples.aime.init_dataset()
seed_prompt = {
"system_prompt": "You are a helpful assistant. Answer the question. "
"Put your final answer in the format '### <answer>'"
}
result = gepa.optimize(
seed_candidate=seed_prompt,
trainset=trainset,
valset=valset,
task_lm="openai/gpt-4.1-mini",
max_metric_calls=150,
reflection_lm="openai/gpt-5",
)
print("Optimized prompt:", result.best_candidate['system_prompt'])
Result: GPT-4.1 Mini goes from 46.6% → 56.6% on AIME 2025 (+10 percentage points).
With DSPy (Recommended for AI Pipelines)
The most powerful way to use GEPA for prompt optimization is within DSPy, where it’s available as dspy.GEPA. See dspy.GEPA tutorials for executable notebooks.
import dspy
optimizer = dspy.GEPA(
metric=your_metric,
max_metric_calls=150,
reflection_lm="openai/gpt-5",
)
optimized_program = optimizer.compile(student=MyProgram(), trainset=trainset, valset=valset)
optimize_anything: Beyond Prompts
The optimize_anything API optimizes any text artifact — code, agent architectures, configurations, SVGs — not just prompts. You provide an evaluator; the system handles the search.
import gepa.optimize_anything as oa
from gepa.optimize_anything import optimize_anything, GEPAConfig, EngineConfig
def evaluate(candidate: str) -> float:
result = run_my_system(candidate)
oa.log(f"Output: {result.output}") # Actionable Side Information
oa.log(f"Error: {result.error}") # feeds back into reflection
return result.score
result = optimize_anything(
seed_candidate="<your initial artifact>",
evaluator=evaluate,
objective="Describe what you want to optimize for.",
config=GEPAConfig(engine=EngineConfig(max_metric_calls=100)),
)
Use GEPA as an Agent Skill
GEPA also ships as an Agent Skill so coding agents can drive optimize_anything for you. In a clone of this repo, Claude Code — and other agents that read .claude/skills/ (Cursor, VS Code/Copilot, Codex, Gemini CLI) — auto-discover it at .claude/skills/gepa-optimize-anything/. To use it in any other project, install the plugin from this repo’s marketplace:
/plugin marketplace add gepa-ai/gepa
/plugin install gepa-optimize-anything@gepa
See the Agent Skill guide for details.
How It Works
Traditional optimizers know that a candidate failed but not why. GEPA takes a different approach:
- Select a candidate from the Pareto frontier (candidates excelling on different task subsets)
- Execute on a minibatch, capturing full execution traces
- Reflect — an LLM reads the traces (error messages, profiler output, reasoning logs) and diagnoses failures
- Mutate — generate an improved candidate informed by accumulated lessons from all ancestors
- Accept — add to the pool if improved, update the Pareto front
GEPA also supports system-aware merge — combining strengths of two Pareto-optimal candidates excelling on different tasks. The key concept is Actionable Side Information (ASI): diagnostic feedback returned by evaluators that serves as the text-optimization analogue of a gradient.
For details, see the paper and the documentation.
Adapters: Plug GEPA into Any System
GEPA connects to your system via the GEPAAdapter interface — implement evaluate and make_reflective_dataset, and GEPA handles the rest.
Built-in adapters:
| Adapter |
Description |
| DefaultAdapter |
System prompt optimization for single-turn LLM tasks |
| ConfidenceAdapter |
Logprob-aware classification optimization — penalizes lucky guesses and feeds confidence diagnostics into reflection. pip install "gepa[confidence]" |
| DSPy Full Program |
Evolves entire DSPy programs (signatures, modules, control flow). 67% → 93% on MATH. |
| Generic RAG |
Vector store-agnostic RAG optimization (ChromaDB, Weaviate, Qdrant, Pinecone) |
| MCP Adapter |
Optimize MCP tool descriptions and system prompts |
| TerminalBench |
Optimize the Terminus terminal-use agent |
| AnyMaths |
Mathematical problem-solving and reasoning tasks |
| LangChain |
Optimize prompts for any LangChain pipeline — chat models, tool-using agents, LangGraph. pip install "gepa[langchain]" |
See the adapters guide for how to build your own, and DSPy’s adapter as a reference.
Integrations
GEPA is integrated into several major frameworks:
Example Optimized Prompts
GEPA can be thought of as precomputing reasoning during optimization to produce a plan for future task instances. Here are examples of the detailed prompts GEPA discovers:
| Example GEPA Prompts |
| HotpotQA (multi-hop QA) Prompt |
AIME Prompt |
Click to view full HotpotQA prompt
[HotpotQA Prompt Begin]
You will be given two input fields: question and summary_1.
Your task is to 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 an initial set of documents. Your goal is to generate a second hop query that retrieves additional relevant documents that were not found in the first hop but are necessary to answer the original question completely.
Detailed task instructions and hints:
-
Input Understanding:
question is the original multi-hop question posed by the user.summary_1 is a concise summary of information from a document retrieved in the first hop, which partially addresses the question.
-
Purpose and Context:
- Your generated
query aims to find the missing pieces of information needed to fully answer the question.
- The multi-hop retrieval system works in stages:
- First hop: The original question returns some documents.
- Second hop: Your query must help retrieve any other relevant documents NOT found in the first hop that hold complementary or broader context necessary for final answer extraction.
-
Key Observations from Examples and Feedback:
- First-hop documents often cover one entity or aspect in the question.
- 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 be formulated to explicitly target these missing, but logically linked, documents.
- Avoid merely paraphrasing the original question or restating known facts from
summary_1.
- Instead, infer what broader or related entities/concepts might provide the crucial missing information.
- For example, if
summary_1 describes a population for a small civil parish, but the question wants total population of the wider region, your query should target that wider region (e.g., “Madeira archipelago population in 2011”).
- Similarly, if
summary_1 covers a song and the question wants the album it came from, but first hop got song-level documents, your query should retrieve documents about the album itself.
-
How to Build the Query:
- Identify the entities or topics mentioned in
summary_1 that appear related but different from first-hop documents.
- Reframe the query to explicitly mention these broader or related entities connected to the original question.
- Include relevant key context from the question to maintain specificity, but shift focus to the missing piece.
- The goal is to retrieve documents that link or complement what was retrieved initially.
-
Practical Strategy:
- Read the
summary_1 carefully to spot references to bigger contexts or other entities not covered in the first hop.
- Ask yourself, “What entity or aspect does this summary hint at that could answer the original question but was not found yet?”
- Formulate a precise, focused factual query targeting that entity or concept to retrieve the missing documents.
-
Output:
- Produce only the field
query as a clear, concise question or keyword phrase designed for efficient retrieval of second-hop documents.
- Ensure the query relates logically to the original question while targeting the broader or complementary knowledge identified in
summary_1.
- Do not include the original question or simply rephrase it.
- Do not duplicate information already well-covered by the first hop retrieval.
By following these principles, you will help the multi-hop retrieval system find all necessary documents to answer the multi-faceted original question completely.
[HotpotQA Prompt End]
|
Click to view full AIME prompt
[AIME Prompt Begin]
You will be given one math problem as plain text under a key like “problem.” Your job is to solve it correctly and return:
- reasoning: a concise, logically ordered solution that uses identities/structure to avoid brute force, ends with a quick verification.
- answer: the final requested number/expression only (no extra words).
Formatting:
- Use exactly two top-level fields named “reasoning” and “answer.”
- Keep reasoning succinct but complete. Bullet points are fine.
- The answer field must contain only the final value requested (e.g., 227, 585, 601).
General problem-solving guidance:
- Parse the problem type (e.g., base representation, intersecting families of subsets, avoiding arithmetic progressions, symmetric sums with constraints, ordered tuples counting).
- Always enforce domain constraints (e.g., base-b digits in 0…b−1; no leading zero for base-10 “three-digit”; ordered vs unordered families; strict increase conditions in sequences).
- Use algebraic identities and modular arithmetic to reduce the search space; prefer structural arguments over naive enumeration.
- For “greatest/least” questions, derive tight bounds and give a construction that attains them.
Domain-specific strategies and pitfalls (learned from typical contest problems and prior feedback):
- Base-conversion/digit rearrangement:
- Translate positional notation correctly: in base b, (a b c)_b = a·b^2 + b·b + c; in base 10: abc = 100a + 10b + c.
- Enforce digit ranges strictly (e.g., in base 9, digits ∈ {0,…,8}; if also a is a base-10 leading digit, then a ∈ {1,…,8}).
- Set up equality and simplify. Use modular constraints to prune:
• Mod 9 often collapses coefficients; e.g., 99a = 71b + 8c ⇒ mod 9 gives b + c ≡ 0 (mod 9).
• Mod 8: 99 ≡ 3, 71 ≡ 7 ⇒ 3a ≡ 7b (mod 8) ⇒ b ≡ −3a (mod 8).
- Solve within digit bounds and verify numerically.
- Palindromes across bases:
- Bound the base length by magnitude (e.g., n < 1000 ⇒ octal has 3–4 digits).
- Characterize palindromes:
• 3-digit octal: (A B A)_8 = 65A + 8B.
• 4-digit octal: (A B B A)_8 = 513A + 72B (with A ≥ 1).
- Enumerate small parameter ranges and test the other-base palindrome constraint. For “greatest”, check candidates in descending order with justification.
- Symmetric sums with a + b + c fixed (ordered triples of nonnegative integers):
- Use identities to compress expressions:
S = ab(a + b) + bc(b + c) + ca(c + a) = (a + b + c)(ab + bc + ca) − 3abc.
- With a + b + c known (e.g., 300), convert the given sum into a relation among ab + bc + ca and abc.
- Use the shift a = A + x etc. to isolate a product like (a−A)(b−A)(c−A) and deduce factorization constraints, enabling clean counting.
- Count ordered solutions carefully; include/exclude symmetric/degenerate cases precisely.
- Intersecting families of subsets (collections from the power set):
- Intersecting means every pair has nonempty intersection. The empty set cannot be included.
- Complement pairs: S and S^c cannot both be present. Use this to structure counts.
- Use size-based pigeonhole facts: In [n], any two subsets of size > n/2 must intersect. For n = 5, any two subsets of size ≥ 3 intersect; thus “all subsets of size ≥ 3” is an intersecting family (size 16).
- Do not assume that “stars” (all subsets containing a fixed element) are the only intersecting families of maximum size. For odd n, both the star and “all subsets of size > n/2” have size 2^{n−1}.
- When counting collections of a fixed size:
• Consider the minimum set size N in the family and do casework on how many 2-element sets are included (for n=5), as these control which 3-sets must be excluded (complements).
• Ensure completeness of cases and avoid double counting by parameterizing canonical patterns (e.g., how many 2-sets, how they overlap, whether they share a common element).
• Remember order of subsets in a collection does not matter; count distinct families.
- Avoiding 4-term arithmetic progressions in a strictly increasing sequence with fixed anchors:
- First bound the variable terms by strict increase (e.g., if fixed terms are 3,4,5,…,30,40,50 then 6 ≤ a < b ≤ 29).
- Pre-eliminate values that cause a 4-term AP with three fixed terms:
• 3,4,5,a forbids a = 6.
• b,30,40,50 forbids b = 20.
• Similarly, a,30,40,50 forbids a = 20.
- Start with the count of pairs from allowed values and then subtract specific pairs that complete APs with two fixed endpoints:
• 3,5,a,b ⇒ (a,b) = (7,9).
• 3,a,b,30 ⇒ (a,b) = (12,21).
• 4,a,b,40 ⇒ (a,b) = (16,28).
• 5,a,b,50 ⇒ (a,b) = (20,35) but may be outside bounds or pre-excluded (e.g., 20 banned).
- Systematically check all endpoint combinations; use the fact that if endpoints differ by Δ, then Δ must be divisible by 3 for a 4-term AP, and solve for integer a,b within bounds.
- Avoid double subtraction; ensure monotonicity and domain constraints are respected.
- Order statistics with sum and absolute-sum constraints (e.g., x_1 ≤ … ≤ x_n, sum |x_i| = 1, sum x_i = 0):
- Total positive mass equals total negative mass: both = 1/2.
- For maximizing x_k (k near the top): if there are T largest terms from k to n (T = n − k + 1), then sum of these T terms ≥ T·x_k. Since the total positive mass ≤ 1/2, we get x_k ≤ (1/2)/T.
- For minimizing x_l (l near the bottom): if there are l smallest terms, sum of these l terms ≤ l·x_l. Since the total negative mass is −1/2, we get x_l ≥ (−1/2)/l.
- To attain these bounds, concentrate masses evenly on exactly those positions: set the smallest l terms equal to −1/(2l), the largest T terms equal to 1/(2T), and the middle to 0 (respecting monotonicity). Verify sums and absolute sums.
- Example: For n=100, maximize x_76 − x_16: T = 25 ⇒ x_76 ≤ 1/50; l = 16 ⇒ x_16 ≥ −1/32; construction with 16 negatives at −1/32, 59 zeros, 25 positives at 1/50 attains 1/50 − (−1/32) = 41/800.
Quality checks:
- Verify digit/base constraints and final equalities numerically if applicable.
- For extremal problems, provide both a tight bound and an explicit construction achieving it.
- For counting, explicitly handle ordered vs unordered, exclude impossible/duplicate cases, and check complements/forbidden pairs.
- For AP-avoidance, confirm integrality and bounds; ensure no missed endpoint combinations.
- For “greatest/least” questions, justify optimality structurally (e.g., convexity/majorization/pigeonhole).
Finally:
- Put the clean final numeric result in the “answer” field only.
[AIME Prompt End]
|
When GEPA Shines
- Expensive rollouts — Scientific simulations, complex agents with tool calls, slow compilation. GEPA needs 100–500 evals vs 10K+ for RL.
- Scarce data — Works with as few as 3 examples. No large training sets required.
- API-only models — No weights access needed. Optimize GPT-5, Claude, Gemini directly through their APIs.
- Interpretability — Human-readable optimization traces show why each prompt changed.
- Complements RL — Use GEPA for rapid initial optimization, then apply RL/fine-tuning for additional gains (BetterTogether).
Further Reading
- Paper: GEPA: Reflective Prompt Evolution Can Outperform Reinforcement Learning (arXiv:2507.19457)
- Experiment reproduction artifact: GEPA Artifact Repository
- Talk Slides: GEPA Talk Slides
- Blog Posts:
- Tutorials & Examples:
- Social and Discussion:
- GEPA Integrations:
Want to use GEPA in other frameworks?
- DSPy Adapter Code (integrates GEPA with DSPy),
- MLflow Prompt Optimization - GEPA is integrated into MLflow’s
mlflow.genai.optimize_prompts() API for automatic prompt improvement using evaluation metrics and training data. Works with any agent framework and supports multi-prompt optimization.
- Pydantic AI - Prompt optimization for Pydantic AI.
- Comet ML Opik - Core optimization algorithm in Opik Agent Optimizer.
- OpenAI Cookbook - Self-evolving agents with GEPA.
- HuggingFace Cookbook - Prompt optimization guide.
- Google ADK / Gemini Enterprise Agent Platform -
adk optimize powered by GEPA, also shipped inside Google Cloud’s Gemini Enterprise Agent Platform Quality Flywheel. ADK docs · Community tutorial.
- Contributed Adapters – see our adapter templates and issue tracker to request new integrations.
- DefaultAdapter - System Prompt Optimization for a single-turn task.
- ConfidenceAdapter - Logprob-aware classification optimization using
llm-structured-confidence. Detects lucky guesses by extracting token-level logprobs from structured JSON outputs with enum constraints, and feeds confidence diagnostics (logprob, probability, top alternatives) into the reflection LLM. Install with pip install "gepa[confidence]". See the guide.
- DSPy Full Program Adapter - Evolves entire DSPy programs including signatures, modules, and control flow. Achieves 93% accuracy on MATH benchmark (vs 67% with basic DSPy ChainOfThought).
- Generic RAG Adapter - Vector store-agnostic RAG optimization supporting ChromaDB, Weaviate, Qdrant, Pinecone, and more. Optimizes query reformulation, context synthesis, answer generation, and document reranking prompts.
- MCP Adapter - Optimize Model Context Protocol (MCP) tool usage. Supports local stdio servers, remote SSE/HTTP servers, and optimizes tool descriptions and system prompts.
- TerminalBench Adapter - Easily integrating GEPA into a Terminus, a sophisticated external agentic pipeline, and optimizing the agents’ system prompt.
- AnyMaths Adapter - Adapter for optimizing mathematical problem-solving and reasoning tasks. Contributed by @egmaminta.
- LangChain Adapter - Optimize prompts for any LangChain pipeline: single-turn chat models, tool-using agents built with
create_agent, custom LangGraph graphs, RAG, and more. Provider-agnostic via LangChain’s init_chat_model. Install with pip install "gepa[langchain]" plus a provider package (e.g. langchain-openai).
- GEPA uses
- Microsoft AI: MAI-Thinking-1 — GEPA / DSPy optimizes the Qwen3-30B LLM-judge prompt used to filter Code pages in the pre-training pipeline (~233B tokens curated from ~2,000 human labels).
- Nubank: Building Customer Support AI Agents at 100M-User Scale — GEPA inside DSPy optimizes LLM-as-a-Judge prompts; lifts E2 eval accuracy 68.88% → 88.89%, drives Cohen’s κ (GPT-4.1 vs GPT-4.1-mini) from 0.00 → 0.745; downstream production wins include +37pp AI transactional NPS and +29pp self-service rate across 5 deployed domains.
- Google: Gemini Enterprise Agent Platform —
adk optimize powered by GEPA
- Nous Research Hermes Agent: evolutionary self-improvement with DSPy + GEPA
- Context Compression using GEPA
- GEPA Integration into SuperOptiX-AI
- GEPA for Observable Javascript
- bandit_dspy
- GEPA in Go Programming Language
- 100% accuracy using GEPA on the clock-hands problem
- Prompt Optimization for Reliable Backdoor Detection in AI-Generated Code
- Attack Selection Reduces Safety in Concentrated AI Control Settings (Pivotal Research + Redwood) — GEPA-optimized red-team prompts outperform handwritten rubric prompts at evading trusted monitoring
- Going recursive: RLM-GEPA on AppWorld (Gabriel Lespérance) — PredictRLM(GPT-5.5 low) hits 0.917 TGC / 0.839 SGC unoptimized (beats leaderboard 0.804 SGC); RLM-GEPA lifts to 0.940 TGC / 0.911 SGC
- DivSkill-SQL: Residual Skill Optimization for Text-to-SQL Ensembles (UC San Diego + Microsoft) — adopts GEPA as inner-loop skill optimizer; +11.1pp on Spider2-Lite Snowflake, +8.3pp on BigQuery, +2.6pp on BIRD-Critic over CHASE-SQL
- DD-GEPA: Dialogue Disentanglement Prompt Optimization (Yokohama National University) — three-component prompt (task instruction / utterance representation / output instruction) optimized with GEPA; Qwen3-30B F1 39.40 → 42.52 on the Kummerfeld benchmark
- Teaching LLMs to Diagnose Production Incidents with ATLAS+GEPA
- DataBricks: Building State-of-the-Art Enterprise Agents 90x Cheaper with GEPA
- comet-ml/opik adds support for GEPA
- Tuning small models (Gemma3-1B) for writing fiction
- Cut OCR Error Rates by upto 38% across model classes (Gemini 2.5 Pro, 2.5 Flash, 2.0 Flash)
- LanceDB: Handwritten Medical-Notes OCR Pipeline with DSPy + GEPA — normalized OCR match 54.1% → 59.4%, avg edit distance 1.01 → 0.87 on Gemini 3.1 Flash Lite, ~15 min run on a laptop
- Optimizing a Data Analysis coding agent with GEPA, using execution-guided feedback on real-world workloads
- Generating Naruto (Anime) style dialogues with GPT-4o-mini using GEPA
- Augmenting RL-tuned models with GEPA: Achieving +142% student performance improvement by augmenting a RL-tuned teacher with GEPA
- DeepResearch Agent Optimized with GEPA
- Boosting Sanskrit QA: Finetuning EmbeddingGemma with 50k GEPA generated synthetic data samples (Tweet), (Code)
- Simulating Realistic Market Research Focus Groups with GEPA-Optimized AI Personas
- Google ADK: Official agent optimization powered by GEPA
- HuggingFace Cookbook on prompt optimization for with DSPy and GEPA
- OpenAI Cookbook showing how to build self-evolving agents using GEPA
- MarkTechPost tutorial: Reflective Prompt Optimization with GEPA — multi-component prompts, structured feedback, held-out validation
- What Do Prompts Reveal About Model Capabilities in Low-Resource Languages? (AfricaNLP 2026)
- Beyond the Answer: Decoding the Behavior of LLMs as Scientific Reasoners (ICLR 2026 Workshop)
- Self-Optimizing Multi-Agent Systems for Deep Research (ECIR 2026 Workshop) — GEPA outperforms TextGrad and expert-crafted prompts
- Prompt Optimisation for Error Detection in Medical Notes (MEDEC) — GPT-5 0.669 → 0.785, Qwen3-32B 0.578 → 0.690
- Automated Risk-of-Bias Assessment of Clinical Trials — GEPA-optimized prompts across 7 RoB domains, 30–40% improvement
- Clinical NER: GEPA vs Bio+ClinicalBERT (IEEE BigData 2025) — up to 12.5% F1 lift in zero-shot clinical NER
- Prompt Triage: Structured Optimization for VLMs on Medical Imaging (Stanford) — median 53% relative improvement on 5 imaging tasks
- Cancer-Myth: GEPA-optimized precautionary prompts for false presuppositions in cancer patient questions
- WER is Unaware (IWSDS 2026) — GEPA-optimized LLM-as-a-Judge for clinical ASR risk (90% accuracy, κ=0.816)
- EvoClinician — GEPA baseline on Med-Inquire multi-turn medical diagnosis
- TRACE — two-phase evolution “inspired by GEPA” over streaming EHRs
- SecureForge (Stanford) — GEPA-based hardening of code-gen LLMs against vulnerabilities; outperforms MIPRO
- OrchMAS — GEPA as MAS prompt-optimization baseline across six QA benchmarks
- REVERE (TCS + Yale) — GEPA as offline baseline on scientific research-coding (SUPER, RCB, ScienceAgentBench)
- Empowering Small Models for GPU Parallelization (OpenACC) — GPT-5 Nano to 100% compilation on PolyBench
- Automated Refinement of Essay Scoring Rubrics (U. Tokyo) — “a simplified version of GEPA”
- Optimized Agentic AI Systems for Asset Pricing (SSRN)
- VeriInteresting — GEPA-style prompt evolution for Verilog HDL code generation
- VeriAct — GEPA as a core part of formal specification synthesis
- Survey on AI-Driven Circuit Verification (ASPDAC 2026, CUHK) — cites GEPA as a promising approach to avoid data scarcity
- FEM-Bench — scientific-reasoning benchmark using GEPA as baseline optimizer
- AssayBench — assay-level virtual cell benchmark; GEPA optimizes pipelines before evaluation
- Reinforced Agent: Inference-Time Feedback for Tool-Calling Agents — reviewer architecture + GEPA prompt optimization
- Databricks Genie: GEPA-optimized table search inside Databricks’ enterprise data agent
Contributions
We welcome adapters, bug fixes, and new use cases. See src/gepa/adapters/ for adapter examples and the contributing guide.
Want to highlight your use case? Reach out to [email protected] or submit via GitHub.
Contributors
Everyone who has committed to this repository, written a post on the GEPA blog, or authored the GEPA or optimize_anything paper. This list is generated by docs/scripts/generate_contributors.py. If you’re missing, or you’d like your full name or GitHub account shown, please open an issue or a PR.
Citation
@misc{agrawal2025gepareflectivepromptevolution,
title={GEPA: Reflective Prompt Evolution Can Outperform Reinforcement Learning},
author={Lakshya A Agrawal and Shangyin Tan and Dilara Soylu and Noah Ziems and Rishi Khare and Krista Opsahl-Ong and Arnav Singhvi and Herumb Shandilya and Michael J Ryan and Meng Jiang and Christopher Potts and Koushik Sen and Alexandros G. Dimakis and Ion Stoica and Dan Klein and Matei Zaharia and Omar Khattab},
year={2025},
eprint={2507.19457},
archivePrefix={arXiv},
primaryClass={cs.CL},
url={https://arxiv.org/abs/2507.19457},
}
评论