Mendel G\"odel Machine Boosts Self-Improving Coding Agents

Self-improving coding agents, which can iteratively modify their own code, have shown impressive capabilities in programming tasks. However, current methods typically learn from one failure at a time, missing out on valuable comparative data from an agent's history of attempts. Inspired by Mendelian principles, researchers have developed the Mendel G\"odel Machine (MGM). This framework introduces two novel self-modification strategies beyond the standard single-trajectory "clonal mutation." The "reaction-norm mutation" allows an agent to be edited based on its performance across multiple tasks simultaneously, while "cross-lineage hybridization" enables an agent to learn from a reference agent's trajectory on the same task but from a different evolutionary lineage. Theoretical proofs and simulations indicate that these new strategies lead to faster and more effective convergence compared to single-trajectory baselines. Experiments conducted on benchmarks like SWE-bench and Polyglot confirm that MGM consistently improves performance, efficiency, and generalizability for coding agents.