Mendel Gödel Machine: Recursive Self-Improving Coding Agents via Comparative Evolution

2026-08-07

AI Summary

This work addresses the limitation of existing self-improving coding agents that rely solely on single failed trajectories for code rewriting, thereby failing to leverage richer comparative information from historical attempts. Inspired by Mendelian inheritance principles, we propose a recursive self-improvement mechanism incorporating two novel strategies: reaction norm mutation and cross-lineage hybridization, which transcend the constraints of single-trajectory evolution. Built upon an additive fitness landscape model, our approach integrates clonal mutation, multitask reaction norm editing, and cross-lineage trajectory fusion to significantly enhance convergence speed and generalization capability. Experiments on SWE-bench and Polyglot benchmarks demonstrate consistent superiority over current single-trajectory baselines in performance, efficiency, and generalization, with both theoretical analysis and empirical simulations corroborating the method's effectiveness.

Abstract

Self-improving coding agents that iteratively rewrite their own source code have demonstrated impressive performance on coding tasks. However, existing solutions generally derive self-modification from a single failure trajectory at a time, overlooking rich comparative signals available in the agent's expanding archive of past attempts. According to Mendelian principles of controlled inheritance, we introduce Mendel Gödel Machine (MGM). In addition to the general single-trajectory clonal mutation, MGM includes two new types of self-modification that better utilizes evidences accumulated: the reaction-norm mutation edits an agent based on its trajectories on multiple tasks simultaneously, and the cross-lineage hybridization edits an agent using the trajectory of a reference agent from another lineage on the same task. Under an additive fitness landscape model, we prove theoretically and demonstrate via controlled surrogate simulation that the new strategies facilitate a faster and better convergence over single-trajectory baselines. Experiments on SWE-bench and Polyglot confirm MGM's consistent improvement in performance, efficiency, and generalizability.

Problem

• self-improving agents
• code rewriting
• comparative signals
• trajectory analysis
• inheritance in AI

Innovation

• self-improving agents
• comparative evolution
• reaction-norm mutation
• cross-lineage hybridization
• recursive code rewriting