跳转至

文章背景与核心概要

近年来，能够迭代重写自身源代码的自我改进型编码智能体在各项代码基准测试中展现出了卓越的性能。然而，现有的框架通常每次只依赖于单一的失败轨迹，忽略了智能体不断增长的历史尝试档案中所蕴含的丰富比较信号。

为了弥补这一空白，作者受到了孟德尔受控遗传原理的启发，引入了孟德尔哥德尔机（Mendel Gödel Machine, MGM）。除了传统的单轨迹克隆突变外，MGM 还引入了两种新颖的自我修改策略：1. 反应范数突变：根据智能体在多个任务上的跨任务性能轨迹对其进行编辑；2. 跨谱系杂交：使用来自完全不同谱系的参考智能体在同一任务上的轨迹来编辑智能体。

在加性适应度地形模型下，作者从理论上证明了——并通过受控替代模拟进行了演示——与标准的单轨迹基线相比，这些新的比较策略能够实现更快、更优的收敛。在 SWE-bench 和 Polyglot 上的实验进一步证实了其在整体性能、运行效率和泛化能力上的持续提升。

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

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

arXiv ID: 2608.07645
Subjects: Artificial Intelligence (cs.AI); Machine Learning (cs.LG)
Submitted: August 7, 2026
Authors: Changzhi Liu, Yilun Liu, Sikuan Yan, Volker Tresp, Yunpu Ma

arXiv ID: 2608.07645
Subjects: Artificial Intelligence (cs.AI); Machine Learning (cs.LG)
Submitted: August 7, 2026
Authors: Changzhi Liu, Yilun Liu, Sikuan Yan, Volker Tresp, Yunpu Ma

摘要概要

自我改进型编码智能体通过迭代重写其自身的源代码，在各种编码基准测试中展现出卓越的性能。然而，现有的框架通常一次仅依赖单一的失败轨迹，忽视了智能体不断积累的历史尝试档案中蕴含的丰富比较信号。

Self-improving coding agents that iteratively rewrite their own source code have shown remarkable performance on various coding benchmarks. However, existing frameworks typically rely on a single failure trajectory at a time, ignoring the rich comparative signals found within an agent's growing archive of past attempts.

为了填补这一空白，作者受孟德尔受控遗传原理的启发，推出了孟德尔哥德尔机（Mendel Gödel Machine, MGM）。除了传统的单轨迹克隆突变之外，MGM 还引入了两种新颖的自我修改策略： 1. 反应范数突变（Reaction-Norm Mutation）：基于智能体在多个任务上的综合性能轨迹对其进行编辑。 2. 跨谱系杂交（Cross-Lineage Hybridization）：利用来自完全不同谱系的参考智能体在同一任务上的轨迹来编辑智能体。

To bridge this gap, the authors introduce the Mendel Gödel Machine (MGM), inspired by Mendelian principles of controlled inheritance. In addition to traditional single-trajectory clonal mutations, MGM introduces two novel self-modification strategies: 1. Reaction-Norm Mutation: Edits an agent based on its performance trajectories across multiple tasks simultaneously. 2. Cross-Lineage Hybridization: Edits an agent using the trajectory of a reference agent from a completely different lineage on the same task.

在加性适应度地形模型下，作者从理论上证明——并通过受控代理模拟进行验证——这些新的比较策略比标准的单轨迹基线实现了更快、更优的收敛。在 SWE-bench 和 Polyglot 上的实验进一步证实了其在整体性能、运行效率和泛化能力方面的持续提升。

Under an additive fitness landscape model, the authors prove theoretically—and demonstrate via controlled surrogate simulations—that these new comparative strategies achieve faster and superior convergence compared to standard single-trajectory baselines. Experiments on SWE-bench and Polyglot further confirm consistent improvements in overall performance, operational efficiency, and generalizability.

关键链接与资源

引用 (BibTeX)

Citation (BibTeX)

@misc{liu2026mendelgodelmachinerecursive,
      title={Mendel Gödel Machine: Recursive Self-Improving Coding Agents via Comparative Evolution}, 
      author={Changzhi Liu and Yilun Liu and Sikuan Yan and Volker Tresp and Yunpu Ma},
      year={2026},
      eprint={2608.07645},
      archivePrefix={arXiv},
      primaryClass={cs.AI}
}