一句话：提出LLM原生应用概念，以LLM为编排核心，并给出参考架构、生命周期模型及文献综述copilot案例。
当前多数软件系统要么以 LLM 为核心包裹既有系统（LLM-based），要么将 LLM 嵌入既有工作流（LLM-augmented），虽能快速获益，却脆弱、难以复现且治理不足。本文提出“LLM 原生应用”概念：系统从设计之初就以 LLM 作为编排核心，统一协调工具、记忆、知识集成与治理。作者给出参考架构，并通过一个用于文献综述的 LLM 原生 copilot 展示可行性；同时提出生命周期模型，说明当 LLM 成为一等编排者时开发过程如何变化。该工作为 LLM 时代的软件工程新范式提供了概念框架与初步案例。
Large Language Models (LLMs) are increasingly embedded into software systems, yet most current solutions are either LLM-based, wrapping an existing system around an LLM, or LLM-augmented, embedding LLMs into existing workflows. These approaches yield quick gains but remain brittle, difficult to reproduce, and poorly governed. We introduce the notion of LLM-native applications: systems conceived from the outset with an LLM as the orchestration core, coordinating tools, memory, knowledge integration, and governance. We present a reference architecture and illustrate feasibility through an LLM-native copilot for literature reviews. We further propose a lifecycle model that highlights how development processes change when LLMs become first-class orchestrators.
@inproceedings{paperbot11821,
title = {LLM-Native Applications: Toward a New Paradigm in Software Engineering},
author = {Alan Portela Bandeira and Gabriel Luiz Barros De Oliveira and Paulo Henrique Mendes Maia and Carlos Aderaldo and Nabor Mendonca},
booktitle = {ICSE},
year = {2026},
doi = {10.1145/3774748.3795680}
}