Multi-hop claim verification (retrieval)
HoVer
Metric · Retrieval Score
Baseline program
0.47
After CodeEvolver
0.88
+87% relative improvement. Baseline is a robust third-party program running on GPT-5.4 nano; the same model is used after optimization — all gains come from prompt and code changes.
Synthetic knowledge-base QA
PhantomWiki
Metric · Answer Accuracy
Baseline program
0.38
After CodeEvolver
0.69
+80% relative improvement. Synthetic benchmark with fully reproducible evals. This is the run shown in the homepage animation.

Score on validation set shown above. For the holdout test sets, PhantomWiki progressed 41% → 65%; HoVer, 56.9 → 85.8. See our methods below.

How we ran these numbers.

Setup

Each benchmark starts from a third-party program written by researchers — not a strawman baseline. We extract the benchmark dataset and split it into training, validation, and test sets. The test set is touched exactly once, at the very end.

The engine runs with an instruction set combining optimization constraints and seed ideas (architecture changes, retrieval strategies). GPT-5.4 nano is held constant throughout — every point of improvement comes from prompt and code changes, not a bigger model.

The engine

CodeEvolver combines an evolutionary optimizer with a reflection agent that analyzes failures and a coding agent that executes changes. Maintaining a diverse pool of candidates avoids the greedy optimizer problem, where improving a good program gets stuck before reaching a great one.

Every accepted change is preserved with its full experiment history: what was proposed, what it scored, and what it replaced.

Read the research notes.

CodeEvolver Outcomes
Full benchmark results across HoVer and PhantomWiki — with starting and final programs.
The State of Prompt Optimization
Why manual prompting isn't enough, how GEPA works, and where prompt-only optimization hits its ceiling.
Welcome to CodeEvolver
What CodeEvolver is, the messy cycle it replaces, and how the engine combines optimization with coding agents.

Our research paper is in progress.

Put Us to the Test

We'll run it on your benchmark.

Get a Baseline Assessment