ASI-LIB-015 technical code available

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

Jenny Zhang, Shengran Hu, Cong Lu, Robert Lange, Jeff Clune

DGM is one of the clearest empirical steps toward recursive self-improvement: the system edits its own codebase and validates those edits against software engineering benchmarks.

Results

The reported run improved SWE-bench performance from 20.0% to 50.0% and Polyglot from 14.2% to 30.7%, while maintaining an archive of many generated agents instead of climbing a single hill.

ASI relevance

Open-ended archives matter. ASI progress may depend less on one perfect update and more on preserving diverse stepping stones, including changes that look locally odd but unlock later improvements.