GEPA: Reflective Prompt Evolution Can Outperform Reinforcement Learning

When a team wants an LLM system to get better at a task, one option is reinforcement learning on the weights, for example GRPO, which turns each attempt into a single number and needs many thousands of attempts. Lakshya Agrawal, Omar Khattab, Matei Zaharia and colleagues at UC Berkeley, Stanford, Databricks, MIT and elsewhere argue that this throws information away. An attempt by an LLM system is a readable transcript: the instructions, the reasoning, the tool calls, and often the evaluator's own explanation, like a compiler error or a list of missing documents. A model can read that and work out what to change.

GEPA (Genetic-Pareto) is a prompt optimizer built on that idea. It works on compound systems with several LLM modules, such as a multi-hop retriever with a query writer and a summarizer. Each iteration it picks a candidate set of prompts, runs it on a minibatch of three training examples while recording traces, and calls a feedback function that returns a score plus text. A reflection model reads the current prompt, the traces and the feedback for one module, chosen in turn, and writes an improved prompt. If the new version scores better on the same three examples, it is evaluated on a validation set and added to the pool.

The second idea is how it picks what to improve next. Always mutating the best candidate tends to get stuck on one strategy. GEPA keeps every candidate that is best on at least one validation task, drops those that are dominated, and samples in proportion to how many tasks each one leads. An optional merge step combines modules from lineages that improved different parts of the system.

The prompts it finds are detailed task instructions. On six benchmarks with Qwen3 8B it beat GRPO, trained with 24,000 rollouts, on five, while using 1,839 to 7,051. With GPT-4.1 Mini it beat the prompt optimizers MIPROv2, TextGrad and Trace on every benchmark, and its prompts were up to 9.2 times shorter than MIPROv2's. The paper was accepted as an oral at ICLR 2026.

Most customer deployments use a model you cannot fine-tune, on tasks where every attempt costs a tool call or an API bill. GEPA is designed for exactly that: a few hundred to a few thousand attempts, no weight updates, and an output you can read and review before shipping, because it is a prompt.

Your evaluator is the lever. GEPA learns from the explanation, not just the score. A feedback function that says "the answer cited a document not in the retrieved set" or "the response leaked the customer's account number" gives the reflection model something to fix. If your evaluator returns only pass or fail, you have thrown away most of what makes this method work.

The optimized prompt is an artifact to review. It is a page of instructions. Read it, check it against policy and the customer's terms, and version it like code. Tuning a prompt changes the system's behavior just as a code change does.

Prompt changes transfer. Prompts optimized with Qwen3 8B improved GPT-4.1 Mini by 9 aggregate points without changes, more than the other optimizers achieved working on GPT-4.1 Mini directly. That makes a cheap model a reasonable place to start an optimization.

Keep a held-out test set. GEPA selects on a validation set it scores repeatedly. The paper's numbers are on separate test sets, and yours should be too.