一句话：提出「绿色架构重构」愿景：用 LLM 逆向工程遗留架构，识别能耗反模式以指导可持续重构。
随着 ICT 行业碳足迹持续增长，可持续性正成为软件工程中的关键非功能需求。现有研究多聚焦于 Green Code，即在实现层面做优化，但这类微观改进往往会被架构层面的低效所抵消。本文作为一篇愿景论文（vision paper），提出「绿色架构重构」概念，主张从以性能为中心的代码优化转向以能耗为中心的架构推理。为处理实现层面无法解决的结构性低效，作者提出一种由 LLM 驱动的方法，对遗留软件架构进行逆向工程，并支持围绕能耗相关设计权衡的架构推理；具体探讨了 AI 辅助的架构蒸馏如何识别架构级能耗反模式（如资源过度供给或通信过于频繁的微服务），并据此指导面向更可持续设计的重构决策。论文还给出研究路线图与架构能耗分析的实验框架，为定义架构能耗度量、指导未来实证验证提供概念基础，将能效确立为一等的架构关注点。
As the carbon footprint of the Information and Communication Technology (ICT) sector continues to grow, Sustainability is increasingly recognized as a critical non-functional requirement in software engineering. While existing research predominantly focuses on Green Code, which targets implementation-level optimizations, such micro-level improvements can be undermined by architectural inefficiencies. This vision paper articulates the concept of Green Architectural Reconstruction, advocating a shift from performance-centric code optimization to energy-centric architectural reasoning. To address structural inefficiencies that cannot be resolved at the implementation level, we propose an LLM-driven method to reverse-engineer legacy software architectures and support architectural reasoning about energy-related design trade-offs. Specifically, we explore how AI-assisted architectural distillation can identify architectural energy anti-patterns (e.g., over-provisioned or overly chatty microservices) and inform refactoring decisions toward more sustainable designs. We outline a research roadmap and an experimental framework for architectural energy analysis, providing a conceptual basis for defining architectural energy metrics and guiding future empirical validation. By positioning energy efficiency as a first-class architectural concern, this work contributes a foundation for future research on sustainability-aware software architecture.
@inproceedings{paperbot11906,
title = {Green Architectural Reconstruction: Bridging the Gap Between Code Efficiency and Sustainable Design},
author = {Esther Hahyeon Kim and Antheas Kapenekakis and Bent Thomsen and Dragos Ionescu and Michele Albano and Søren Kejser Jensen},
booktitle = {ICSE},
year = {2026},
doi = {10.1145/3774748.3795873}
}