一句话：提出 INTENTFIX，用 LLM 将开发者意图形式化为模型，差分定位语义缺口并自动合成补丁修复逻辑漏洞。
逻辑漏洞源于开发者意图与代码实现之间的语义鸿沟：这类缺陷能通过传统测试，却携带可导致严重安全事件的关键缺陷。论文提出 INTENTFIX，一个以意图为中心的安全分析框架，用于自动修复逻辑漏洞。它首先利用大语言模型系统性地抽取开发者的隐含意图，并将其形式化为结构化的意图模型；随后在该意图模型与实际实现之间进行差分分析，以精确识别二者之间的语义缺口；最后通过多视角、LLM 驱动的推理过程合成并迭代精化补丁。摘要未报告具体数据集与量化实验结果。
Logic vulnerabilities, which arise from semantic gaps between a developer’s intent and the actual code, represent a critical and growing challenge in software security. Unlike syntactic bugs, these vulnerabilities pass traditional testing while harboring critical security flaws that can lead to severe breaches. We introduce INTENTFIX, a novel framework that automatically repairs logic vulnerabilities through intent-centric security analysis. INTENTFIX first leverages a Large Language Model (LLM) to systematically extract and formalize the developer’s implicit intent into a structured model. It then performs a differential analysis between this intent model and the implementation to precisely identify semantic gaps. Finally, it synthesizes and refines a patch through a multi-aspect, LLM-driven reasoning process.
@inproceedings{paperbot11632,
title = {INTENTFIX: Automated Logic Vulnerability Repair via LLM-Driven Intent Modeling},
author = {Jinseok Heo and Dongwook Choi and Jinyoung Kim and Misoo Kim and Eunseok Lee},
booktitle = {ICSE},
year = {2026},
doi = {10.1145/3744916.3773260}
}