一句话：提出任务无关评估框架，用元模型从可复用原则自动合成 LLM 嵌入式应用的评估。
LLM 嵌入式应用在不同上下文中表现出的非确定性行为，使其评估与传统软件测试存在本质差异：现有基准难以覆盖领域特定需求，而实践者只能采用耗时且不一致的临时策略。本文提出一种任务无关的评估框架，不再要求为每个任务定制配置，而是从可复用的评估原则自动合成评估方案。其核心是一个捕获任务本质属性的元模型，借此在客观任务与主观任务上均可自动生成评估。作者通过系统文献综述、灰色文献分析与从业者验证，建立了面向多种 LLM 任务类型的评估选择指南；并通过行业案例研究检验框架的实际效果，从泛化性、性能、时间效率与工作流契合度等方面进行评估。该研究将 LLM 评估从任务特定设计推进到基于可复用原则的系统化合成，奠定了相应的方法论基础。
Evaluating Large Language Model (LLM) embedded applications presents fundamental challenges distinct from traditional software testing due to non-deterministic behaviour across diverse contexts. Existing benchmarks fail to capture domain-specific requirements, while practitioners resort to ad-hoc strategies that are time-consuming and inconsistent. This research proposes a task-agnostic evaluation framework that automatically synthesises evaluations from reusable principles rather than requiring custom configurations. The framework employs a meta-model capturing essential task properties to enable automated evaluation generation across objective and subjective tasks. Through systematic literature review, grey literature analysis, and practitioner validation, we establish guidelines for evaluation selection across diverse LLM task types. Industry case studies validate the framework’s practical effectiveness, measuring generalizability, performance, time efficiency, and workflow alignment. This research establishes methodological foundations that shift LLM evaluation from task-specific design to systematic synthesis from reusable principles.