一句话：系统探究七种先进大语言模型提示生成漏洞样本的潜力与不足，并据此提出改进方法。
数据驱动的漏洞分析依赖高质量、大规模的训练数据，但现有专用漏洞数据生成技术存在两点根本局限：一是只能处理简单的（单行注入式）漏洞，二是容易对种子样本过拟合。大语言模型在生成任务上表现优异，有望突破这些限制，但其在漏洞样本生成任务上的实际表现尚不明确。本文通过设计提示（prompting）策略，对七种当前最先进的大语言模型开展系统探索，考察它们在生成真实世界漏洞数据方面的潜力与差距，并在此基础上提出改进方案。论文标题表明作者不仅停留在评估层面，还针对发现的不足给出了改进的数据生成途径。该工作为利用大语言模型缓解漏洞分析中的数据稀缺与分布单一问题提供了实证依据。
Data-driven approaches have proven to be promising for vulnerability analysis, contingent on quality and sizable training data being available. Several dedicated vulnerability data generation techniques have demonstrated strong merits, yet they are limited to simple (single-line injection induced) vulnerabilities only and suffer from overfitting to seed samples. Large language models (LLMs) may overcome these fundamental limitations as they are known to be effective at generative tasks. However, it remains unclear how they would perform on the task of vulnerable sample generation. In this paper, we explore the potential and gaps of seven state-of-the-art (SOTA) LLMs for that task via prompting.
@inproceedings{paperbot11696,
title = {Exploring and Improving Real-World Vulnerability Data Generation via Prompting Large Language Models},
author = {Guangbei Yi and Yu Nong and Minzhang Li and Haipeng Cai},
booktitle = {ICSE},
year = {2026},
doi = {10.1145/3744916.3773176}
}