Mendel Gödel Machine Adds Biological Mutation Operators to Self-Rewriting Coding Agents — Learning From Many Trajectories, Not One Failure

arXiv / HuggingFace Daily Papers·medium signalXBlueskyLinkedInCopy link

A paper from Changzhi Liu, Volker Tresp, Yunpu Ma and colleagues (2608.07645, submitted August 7) argues that current self-improving coding agents learn from a single failure trajectory at a time, and adds two mutation strategies borrowed from biology: reaction-norm mutation, which modifies the agent based on multiple task trajectories simultaneously, and cross-lineage hybridization, which imports trajectories from agents in a different evolutionary lineage. Under an additive fitness model they prove convergence is faster and better than single-trajectory baselines, and validate consistent gains in performance, efficiency and generalizability on SWE-bench and Polyglot. Code is released.