OSVBench
收藏资源简介:
OSVBench是一个用于评估大型语言模型(LLMs)在生成操作系统内核验证任务相关完整规范代码方面的基准。该基准将规范生成问题定义为在语法和语义的限定范围内进行程序合成问题,并为LLMs提供编程模型。LLMs需要理解提供的验证假设和潜在的语法和语义空间,然后在操作系统的高级功能描述的指导下,为可能存在错误的操作系统代码实现生成完整的规范。该基准建立在真实的操作系统内核Hyperkernel之上,总共包含245个复杂的规范生成任务,每个任务大约包含20k到30k个token。我们对12个LLMs的综合评估表明,当前LLMs在操作系统验证的规范生成任务上的性能有限。它们在基准测试上的性能差异显著,突出了它们处理长上下文代码生成任务的能力差异。评估工具包和基准测试可在https://github.com/lishangyuhkust/OSVBench获得。
OSVBench is a benchmark for evaluating large language models (LLMs) on the task of generating complete formal specification code related to operating system kernel verification. This benchmark defines the specification generation problem as a program synthesis task within the bounds of syntax and semantics, and provides programming models for LLMs. LLMs are required to understand the provided verification hypotheses and the underlying syntactic and semantic spaces, then generate complete specifications for bug-prone operating system code implementations under the guidance of high-level functional descriptions of the operating system. This benchmark is built on the real operating system kernel Hyperkernel, and contains a total of 245 complex specification generation tasks, each containing approximately 20k to 30k tokens. Our comprehensive evaluation of 12 LLMs shows that the current performance of LLMs on operating system verification specification generation tasks is limited. Their performance on this benchmark varies significantly, highlighting the differences in their ability to handle long-context code generation tasks. The evaluation toolkit and benchmark are available at https://github.com/lishangyuhkust/OSVBench.

- 1OSVBench: Benchmarking LLMs on Specification Generation Tasks for Operating System Verification香港科技大学 · 2025年



