Created yes · retained yes · inherited yes · invoked yes
Mechanism and system boundary must be read with the associated study; task gains alone do not prove better future improvement.
preprint · 2025-05-29
Darwin Gödel Machine: Open-Ended Evolution of Self-Improving Agents
Jenny Zhang, Shengran Hu, Cong Lu, Robert Lange, Jeff Clune
Why it matters here
Agents modify their own scaffolds and reuse descendants in further search while foundation-model weights stay fixed.
Recursive improvement evidence
What to keep in mind
- Historical v1; subsequent revisions exist. Coding-task improvement does not prove open-ended accelerating AI R&D.
Recursive-improvement study
Evidence for scaffold-level recursion with fixed underlying model weights; bounded coding evaluations.
| What changed | scaffold, tools |
|---|---|
| What stayed fixed | Foundation model weights; Outer archive-selection experiment |
| Improved system used as optimizer later | yes |
| Generations attempted / accepted | 80 / not reported |
| Held-out transfer | Full Polyglot and cross-model/language transfer assessed by authors. |
| Resource accounting | 80 generated-agent iterations; 2 parallel SWE-bench and 4 parallel Polyglot iterations. No new experiments run by this tracker. |
| Human contributions | Human-designed evaluation, initial agent, safety boundaries and experiment setup. |
| Author claims | Authors describe a self-improving agent with reusable descendants. |
Recursive-improvement evidence
tracker evidence synthesisAssessment by RSI Tracker (Codex evidence synthesis) · 2026-09-26 · extraction review agent checked.
Budget matched: unknown · independent evaluation: unknown
The revised agent produces descendants, but gains in task performance do not isolate improved search effectiveness under a matched meta-level comparison.
Not established by this particular source; no universal negative claim.
Resource budget:unknown · See study resource accounting; evaluation constraints are distinct from cumulative search cost.·Independence:known · Source-author report; no independent replication recorded.
Adaptivity
qualitativeBest-of-search coding gains over 80 search iterations, not one continuous lineage.
Retention
not reportedNot recorded in this dataset.
Transfer
qualitativeFull Polyglot and cross-model/language checks.
Efficiency
not reportedNot recorded in this dataset.
Stability
qualitativeEvaluation exploitation and harmful candidates remain concerns.
Meta-recursion
qualitativeModified agents participate in later self-modification; archive and parent selection remain fixed.