Mendel Gödel Machine: Recursive Self-Improving Coding Agents via Comparative Evolution
The paper’s core move is to let self-improving coding agents learn from archived comparisons, not just one failed run.
Mendel Gödel Machine adds two self-modification paths: reaction-norm mutation across multiple task trajectories, and cross-lineage hybridization using another agent’s trajectory on the same task. The authors argue, with a fitness-landscape proof and surrogate simulations, that these strategies converge faster and better than single-trajectory baselines. They report consistent gains on SWE-bench and Polyglot in performance, efficiency, and generalization. HF Daily Papers' note
Mendel Gödel Machine adds two self-modification paths: reaction-norm mutation across multiple task trajectories, and cross-lineage hybridization using another agent’s trajectory on the same task. The authors argue, with a fitness-landscape proof and surrogate simulations, that these strategies converge faster and better than single-trajectory baselines. They report consistent gains on SWE-bench and Polyglot in performance, efficiency, and generalization. HF Daily Papers' note
score 5