What Do Evolutionary Coding Agents Evolve?
Nico Pelleriti Sree Harsha Nelaturu Zhanke Zhou Zongze Li, [0.2em] Max Zimmer Bo Han Sebastian Pokutta, [0.4em] Zuse Institute Berlin · 2026
Read What Do Evolutionary Coding Agents Evolve? by Nico Pelleriti Sree Harsha Nelaturu Zhanke Zhou Zongze Li, [0.2em] Max Zimmer Bo Han Sebastian Pokutta, [0.4em] Zuse Institute Berlin (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2605-20086.txt.
Description
Nico Pelleriti and six co-workers ask what coding agents actually evolve, beyond their benchmark scores. A score can reflect new algorithmic structure, retuning, reusing ideas already in the model, or overfitting to the evaluator. EvoTrace gathers coding traces across four frameworks, both kinds of models, and sixteen tasks. EvoReplay replays search states behind top scores, testing changes to constants, removal of parts and model swaps. Most gains come from a small set of edit types, and about thirty percent of added lines bring back lines deleted earlier.
Catalog record
| CREATOR | Nico Pelleriti Sree Harsha Nelaturu Zhanke Zhou Zongze Li, [0.2em] Max Zimmer Bo Han Sebastian Pokutta, [0.4em] Zuse Institute Berlin |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 65848 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2605.20086