AI & Agent Evaluation
2,151total visitsadmin

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

peer-reviewed paper · source date 2025-05-30 · added 2026-08-07 15:52:43 · updated 2026-08-07 15:52:43 · Open original blog

Problems / challenges / motivations

  • The classical Gödel machine requires a formal proof that a self-modification is an improvement, which is infeasible for real systems.
  • Agent design is currently done by hand, one scaffold change at a time.
  • Greedy self-modification collapses into a local optimum and loses variants that would have paid off later.

Key ideas

  • A self-referential coding agent that rewrites its own Python codebase and validates each self-modification empirically on coding benchmarks instead of proving it correct.
  • Maintains an ever-growing archive of variants so open-ended evolutionary search can revisit branches rather than only descending the current best.
  • Reported improvements: SWE-bench 20.0% to 50.0%, Polyglot 14.2% to 30.7%.
  • Sakana AI + UBC (Clune group); ICLR 2026, with open-source code.

Why it matters for AI engineering

  • DGM essentially defined the current recursive-self-improvement wave; nearly every in-window self-improvement paper (Huxley, Red Queen, POLARIS, Live-SWE-agent) builds on or benchmarks against it.
  • The transferable lesson is the archive: keeping the losers is what makes the search open-ended rather than greedy.
  • Caveat: out-of-window (May 2025) and included as the foundational anchor. Improvements are demonstrated within coding benchmarks, and whether self-improvement transfers beyond coding is exactly what its successors target. The offline evolution is compute-intensive relative to runtime self-evolution alternatives.

Comments

No comments yet.