一句话：提出SE4FM框架，用DataOps、ExperimentOps、EvalOps、FieldOps四支柱规范基础模型构建与运行。
基础模型已成为最复杂的软件系统之一，但其工程实践仍依赖口口相传的经验和临时流程，难以应对季度级发布节奏以及模型规模、上下文长度和多模态能力的快速增长。本文以90分钟技术简报形式提出SE4FM，即面向基础模型的软件工程框架，围绕四大支柱组织：DataOps负责数据决策、整理、优化与治理；ExperimentOps强调假设驱动且考虑缩放规律的实验，涵盖RLHF/RLAIF/RLVR流水线；EvalOps关注系统化测试、优先级排序与调试；FieldOps覆盖发布、监控与合规。该工作旨在为软件工程研究者提供基础模型规模化构建与运行的内部视角，并提炼亟需新方法、工具与理论的SE挑战与机会。
Foundation Models (FMs) are among humanity’s most complex software systems. They are currently engineered through tribal knowledge and ad‑hoc processes, despite quarterly release cadences and rapidly expanding model sizes, context lengths, and modalities. This 90‑minute technical briefing presents SE4FM—a disciplined software engineering framing for FMs—organized around four pillars: DataOps (deciding, curating, optimizing, and governing data), ExperimentOps (hypothesis‑driven, scaling‑law‑aware experimentation from small to large; including RLHF/RLAIF/RLVR pipelines), EvalOps (systematic testing, prioritization, and debugging), and FieldOps (releases, monitoring, and compliance). The goal is to give software engineering researchers a unique, insider view of how FMs are really built and run at scale, and to surface concrete SE challenges and opportunities where new SE-centric methods, tools, and theory are urgently needed.
@inproceedings{paperbot11799,
title = {Software Engineering for Foundation Models (SE4FM)},
author = {Boyuan Chen and Dayi Lin and Arthur Leung and Zhilong Chen and Gopi Krishnan Rajbahadur and Gustavo Oliva and Yihao Chen and Xiaoshuang Liu and Chun Yong Chong and Ahmed E. Hassan},
booktitle = {ICSE},
year = {2026},
doi = {10.1145/3774748.3776600}
}