Darwin Gödel Machine: Open-Ended Evolution of Self-Improving Agents

A preprint presenting the Darwin Gödel Machine, a self-improving coding agent that rewrites its own code. Both papers cite it as sighted variation running in the laboratory: over 80 iterations at fixed foundation models its benchmark score rose from 20% to 50%. They also use it for trial times (P6), the unreported child-to-parent ratio (P7, prediction 3), and descendants that removed the markers detecting hallucinated tool use.

Cited in

both papers (Flagship: Evolution by design; Companion §6.1, §6.3, §6.4, Fig. 6)

Used for

P4P6P7Prediction 3P4; P6; P7; prediction 3; Fig. 6b

Strand

Artificial life & machine evolution · Self-improving AI & machine heredity

Cite

@misc{zhang2025darwin,
  title = {Darwin Gödel Machine: Open-Ended Evolution of Self-Improving Agents},
  author = {Jenny Zhang and Shengran Hu and Cong Lu and Robert Lange and Jeff Clune},
  year = {2025},
  url = {https://arxiv.org/abs/2505.22954},
}