Self-improving coding agents that rewrite their own source code have shown promise, but existing systems typically derive self-modification from a single failure trajectory, ignoring the rich comparative signals in the agent's archive of past attempts. A new paper from Hugging Face researchers introduces the Mendel Gödel Machine (MGM), a framework inspired by Mendelian genetics to make evolution more efficient and effective.
MGM introduces two novel self-modification strategies beyond the standard single-trajectory clonal mutation: reaction-norm mutation, which edits an agent based on its performance across multiple tasks simultaneously, and cross-lineage hybridization, which transfers knowledge from a reference agent in another lineage tackling the same task. These methods aim to exploit comparative evidence and enable information sharing across evolutionary branches.
The researchers provide theoretical proof under an additive fitness landscape model and validate their approach through controlled simulations, demonstrating faster and better convergence than single-trajectory baselines. Experiments on the SWE-bench and Polyglot benchmarks confirm MGM's consistent improvements in performance, efficiency, and generalizability.
This work shifts the focus from merely asking whether agents can evolve to how they can evolve with higher quality, ultimately scaling self-evolution. The paper is available on Hugging Face's research platform.