一句话：提出 LASIK，首个自顶向下多智能体 LLM 框架，融合按需仓库级上下文与运行时执行证据检测 MISRA C 违规。
MISRA C 合规性对嵌入式软件的安全与可靠性至关重要，但工业界使用的静态规则检查器漏报与误报率都居高不下，现有基于 LLM 的方法在大型复杂程序上也表现欠佳。为此，作者提出 LASIK——首个面向 MISRA C 的自顶向下多智能体 LLM 合规检查框架，它将按需的仓库级上下文提供与运行时执行证据相结合，从而在大型程序上做出更准确的违规判定。在 10 个嵌入式项目上的评估表明，相较最先进的静态分析器和朴素 LLM 方法，LASIK 检出的真实违规数多 1.3–2.7 倍，同时误报率最低（26.7%）。在 VPlusLab 公司的工业软件上，其真实违规检出量最高为基线的 2.1 倍，误报率仅 10.8%。
Ensuring MISRA C compliance is crucial for the safety and reliability of embedded software. For this purpose, industries utilize static rule checkers but suffer high ratios of both missed violations and false alarms. Although companies also try to utilize LLMs for this purpose, existing LLM-based approaches have suboptimal performance for large, complex programs. We introduce LASIK, the world-first top-down multi-agent LLM compliance-checking framework for MISRA C, which combines on-demand repository-level context provision with runtime execution evidence. This design enables LASIK to report more true violations than state-of-the-art static analyzers and a naive vanilla LLM approach, while maintaining a low false positive ratio. Our evaluation on 10 embedded projects shows that LASIK detects 1.3-2.7 times more true violations than the static analyzers and the vanilla LLM approach, while achieving the lowest false alarm ratio (26.7%). Furthermore, on an industrial software developed by VPlusLab Inc., LASIK detects up to 2.1 times more true violations with a low false alarm rate (i.e., 10.8%) than the baseline techniques.
@inproceedings{paperbot12357,
title = {Top-Down Multi-agent LLM System with Runtime Data for MISRA C Violation Detection},
author = {Ahcheong Lee and Yeongbin Kang and Sechang Jang and Heechan Yang and Moonzoo Kim},
booktitle = {ASE},
year = {2026},
doi = {10.1145/3832783.3834497}
}