一句话：提出可解释性、冗余与可持续性之间的权衡原则，并用带类型上下文记忆框架指导冗余感知系统设计。
论文指出，AI 密集型系统为满足可解释性会积累嵌入、检查点、归因向量和日志等冗余产物，带来存储、能耗与维护成本，但软件工程缺乏量化与平衡这一成本的原则、抽象和指标。作者将冗余视为可解释性的运行阴影，提出把可解释性、冗余与可持续性视为相互依赖的软件质量属性，并定义冗余成本函数，将解释保真度与系统资源消耗联系起来，以指导冗余感知的可解释系统参考架构。设计通过带类型的上下文记忆框架说明：产物被标注解释价值和冗余贡献，并由生命周期控制进行提升、降级或退役。该工作为构建大规模可解释、可复现且可持续的 AI 系统建立新的软件工程方法。
Explainability has become a first-class requirement for AI-intensive systems, ensuring transparency, reproducibility, and accountability. However, explainability introduces a hidden burden. Pipelines accumulate redundant artifacts, including embeddings, checkpoints, attribution vectors, and logs. These artifacts support interpretability, but inflate storage, energy, and maintenance costs. Software engineering lacks principles, abstractions, and metrics to systematically quantify this redundancy cost or balance it against sustainability goals. In this paper, we argue that redundancy is not an isolated concern but the operational shadow of explainability. We propose a trade-off principle that frames explainability, redundancy, and sustainability as interdependent software quality attributes. Our methodology defines a redundancy cost function that links explanatory fidelity to system resource consumption, and uses this function to guide a reference architecture for redundancy-aware explainable systems. The design is illustrated through a typed context memory framework, where artifacts are annotated with explanatory value and redundancy contribution, and managed through lifecycle controls that promote, demote, or retire artifacts. This work establishes new software engineering methods for building AI systems that remain interpretable, reproducible, and sustainable at scale.
@inproceedings{paperbot12013,
title = {Redundancy as the Shadow of Explainability: A Trade-Off Principle for AI-Intensive Systems},
author = {Yan Liu and Jun Huang and Abdelwahab Hamou-Lhadj and Zheng Li},
booktitle = {ICSE},
year = {2026},
doi = {10.1145/3786582.3786829}
}