Workshop · Optimize
GEPA
Seed candidates. Run them. Read the traces. Write a reflection on the failures. Mutate new prompts from that reflection. Keep the non-dominated set across accuracy, cost, and other objectives. Repeat. GEPA is also wired as a DSPy optimizer.
What a template cannot do
A single library persona cannot represent trade-offs. The short prompt and the accurate prompt are often different artifacts. Pareto search keeps both until you choose.
- Not a genetic algorithm over model weights.
- Not live evolution in this encyclopedia.
- Not a window into private chain-of-thought; reflections are written artifacts.
How it works
Conceptual workflow
Steppable map of GEPA. Evidence is taken from the spec on this page. Nothing here is a live compiler, eval, decoder, or optimizer run.
vcp · prompts/gepa
run@once
- seed
- execute
- reflect
- mutate
- pareto
Stage 1 / 5 · receive
Seed the archive
Start with more than one candidate. A single clever persona is a fragile population of one.
Artifact · archive.json
User input
three refund-policy drafts
Rule in force
Diversity in the seed beats one costume.
Visible reply
archive size 3
Illustration · not a live model run
Spec evidence
GEPA is reflective prompt evolution with Pareto selection. It is a population optimizer, not a single rewrite, and it is available as a DSPy optimizer as well as a standalone method. Seed an archive with candidate prompts. Diversity in the seed beats a single clever persona. Execute candidates on a minibatch and keep traces: inputs, outputs, metric scores, and failures. The trace is the evidence the mutator will read. Reflect in language on the failures. The reflection is an inspectable note — which instruction was too vague, which example taught the wrong habit — not a hidden chain-of-thought visualization. Mutate from the reflection: rephrase, add constraints, add or drop demonstrations, split instructions. New candidates are children of traces, not of hunches. Pareto-filter the archive on more than one objective: accuracy, brevity, cost, robustness. Non-dominated prompts stay. A longer prompt that does not score better is discarded. GEPA is not running on this page. The workflow is the evolutionary loop: seed, execute, reflect, mutate, select.
- This page is a conceptual map
- The compiler or eval runner is not executed here
Method
Optimize loop · not executed here
Library
A pasted persona belongs in the library
Conceptual workflow · 4.5s / stage · 1/5