GEPA (Genetic-Pareto)
GEPA, short for Genetic-Pareto, is an automatic prompt optimizer that uses natural-language reflection and evolutionary search to improve prompts in an AI system. It samples execution trajectories, diagnoses failures, proposes textual changes, evaluates candidates and retains complementary solutions on a Pareto frontier. GEPA changes prompts rather than updating a model's weights with policy gradients. “Reflective Prompt Evolution” describes the method; it is not the expansion of the acronym.
Origin and context
The originating preprint was submitted on 25 July 2025 and the work was later published at ICLR 2026. The authors positioned natural-language reflection as a richer optimization signal than sparse scalar rewards for tasks where an LLM can inspect trajectories and articulate candidate rules. Their evaluation covered six tasks and compared GEPA with GRPO and MIPROv2. An independent March 2026 arXiv preprint subsequently evaluated GEPA in another prompt-optimization setting and documented failure modes, supporting use of the term beyond its originating team without proving universal superiority.
Why it matters
Many deployed LLM systems encode behavior in prompts, tool instructions and structured program signatures. Improving those artifacts can be cheaper and easier to inspect than fine-tuning model weights, especially when evaluators can return textual diagnoses. GEPA turns that work into an iterative search process and keeps multiple trade-off candidates rather than only one scalar winner. Its practical value depends on the evaluation set: the optimizer can only select changes that its feedback and metrics recognize, so validation and regression checks remain part of the engineering system.
Example
Imagine a retrieval agent whose prompt often cites irrelevant passages. GEPA can run the agent on labeled cases, collect retrieval and answer traces, ask a reflection model to identify a recurring instruction failure, generate revised prompts, and test them on a validation split. A candidate that improves citation precision without sacrificing answer accuracy may remain on the Pareto frontier. Manually rewriting the prompt once is prompt engineering, but it is not GEPA unless the reflective evaluation and evolutionary selection loop is present.
How it differs
Group Relative Policy Optimization (GRPO)
GRPO updates model parameters from relative rewards over groups of completions. GEPA searches over textual prompts using trajectory reflection and Pareto selection. They can be compared as adaptation strategies in a particular experiment, but GEPA is not a GRPO version or a general replacement for reinforcement learning.
Maturity and evidence
Maturity is rated 3. GEPA has a clear algorithm, public implementation, peer-reviewed ICLR publication and an independent empirical preprint that challenges it. That is sufficient for an established research method. The rating remains below 4 because evidence is recent, results vary by task and seed, and broad production adoption across unrelated organizations has not been demonstrated in the reviewed sources.
Limits and open questions
Reflection is generated by models and can be plausible without identifying the real cause of an error. Search can overfit small evaluation sets, consume many model calls or preserve candidates that exploit a weak metric. An independent preprint reports systematic failures from defective seeds and opaque optimization trajectories. Claims such as outperforming GRPO by up to 19 percentage points or using up to 35 times fewer rollouts belong to the originating six-task setup, not to every prompt, model or workload.
Related terms
References
- GEPA: Reflective Prompt Evolution Can Outperform Reinforcement LearningGEPA research team / arXiv · 2025-07-25 · class A
- GEPA: Reflective Prompt Evolution Can Outperform Reinforcement LearningInternational Conference on Learning Representations · 2026 · class A
- Reflection in the Dark: Exposing and Escaping the Black Box in Reflective Prompt OptimizationIndependent researchers / arXiv · 2026-03-19 · class A
Last updated: 2026-09-07
This term is also covered in the Skills Atlas as automated prompt optimization skill.
This term is also covered in the Skills Atlas as prompt engineering skill.