The Darwin Gödel Machine (DGM), a research paper from Sakana AI and the University of British Columbia, proposes a self-improving AI system inspired by biological evolution. Unlike the original Gödel Machine which required formal mathematical proofs for self-modification, DGM uses open-ended evolution: an archive of coding agents where each agent can rewrite its own code to produce new versions. Selection favors better-performing agents but maintains diversity to avoid local optima. Crucially, agents evolve both their coding skills and their self-improvement capabilities. Evaluated on SWE-bench Verified and Polyglot benchmarks using Claude 3.5 as the base model over 80 iterations, DGM improved from 20% to ~50% on SWE-bench and from 14.2% to 30.7% on Polyglot, outperforming open-source baselines and demonstrating the value of open-ended exploration over greedy single-path refinement.
Sort: