文章背景与核心概要
当前的基于大语言模型（LLM）的自动化算法设计通常将算法视为单一、不可分割的整体程序。这种方法会产生“耦合”问题，即有用的局部逻辑往往会随着宿主程序的淘汰而被一同丢弃，从而阻碍了各个算法组件的重复使用与评估。
PACE（Primitive-Aware Code Evolution，原语感知代码演进）通过将局部逻辑解耦为被称为“可执行算法原语（EAPs）”的持久化单元，有效解决了这一问题。通过维护一个动态的EAP库并利用原语感知演进算子，该系统确保了有价值的代码片段得以保留并在不同程序之间传递。该方法利用基于亲本相对改进的汤普森采样（Thompson sampling）来优化性能，从而在无需外部评估数据集的情况下实现高效选择。
PACE: Primitive-Aware Code Evolution for Automated Algorithm Design
Authors: Zhuoliang Xie, Ruihao Zheng, Xiang Xu, Genghui Li, Zhengkun Wang
Date: August 7, 2026
Primary Subject: Software Engineering (cs.SE)
Secondary Subject: Artificial Intelligence (cs.AI)
Identifier: arXiv:2608.07395
Summary
Current Large Language Model (LLM)-based automated algorithm design typically treats algorithms as monolithic, indivisible programs. This approach creates a "coupling" problem where useful local logic is discarded along with the host program, preventing the reuse and evaluation of individual algorithmic components.
PACE (Primitive-Aware Code Evolution) solves this by decoupling local logic into persistent units called Executable Algorithmic Primitives (EAPs). By maintaining a dynamic library of these EAPs and utilizing primitive-aware evolution operators, the system ensures that valuable code snippets are retained and transferred across different programs. Performance is optimized using Thompson sampling based on parent-relative improvements, allowing for effective selection without the need for external evaluation datasets.
当前的基于大语言模型（LLM）的自动化算法设计通常将算法视为单一、不可分割的整体程序。这种方法会产生“耦合”问题，即有用的局部逻辑往往会随着宿主程序的淘汰而被一同丢弃，从而阻碍了各个算法组件的重复使用与评估。
PACE（Primitive-Aware Code Evolution）通过将局部逻辑解耦为被称为可执行算法原语（Executable Algorithmic Primitives, EAPs）的持久化单元，有效解决了这一问题。通过维护一个动态的EAP库并利用原语感知演进算子，该系统确保了有价值的代码片段得以保留并在不同程序之间传递。该方法利用基于亲本相对改进的汤普森采样来优化性能，从而在无需外部评估数据集的情况下实现高效选择。
Key Features
- Decoupled Architecture: Transforms monolithic code into modular, reusable EAPs.
- Structural Preservation: Evolution operators are designed to guarantee that high-performing algorithmic components are preserved and transferred.
- Efficient Selection: Employs Thompson sampling to identify the most effective primitives, bypassing the need for additional, resource-heavy evaluation datasets.
- Proven Efficacy: Experimental results across four distinct tasks confirm that PACE successfully discovers competitive algorithms while maintaining the integrity of valuable code components.
核心特性
- 解耦架构： 将整体式代码转变为模块化、可重用的EAP。
- 结构保留： 演进算子旨在确保高性能的算法组件得以保留和传递。
- 高效选择： 采用汤普森采样来识别最有效的原语，无需依赖额外且消耗大量资源的评估数据集。
- 经验证的效能： 在四个不同任务上的实验结果证实，PACE成功发现了具有竞争力的算法，同时保持了有价值代码组件的完整性。
Access & Resources
获取与资源
Citation Information
If you are using this paper in your research, you can cite it via the following platforms: * NASA ADS * Google Scholar * Semantic Scholar
引用信息
如果您在研究中使用了本文，可以通过以下平台对其进行引用： * NASA ADS * Google Scholar * Semantic Scholar