| Abstract [eng] |
Symbolic execution is a key program analysis technique for ensuring the quality, security and safety of embedded, cyber–physical and electronic systems, but it is limited by path explosion and costly constraint solving. This study systematically reviews how artificial intelligence (AI) is used to guide symbolic execution. Following the PRISMA 2020 guidelines, five bibliographic sources were searched, yielding 679 records; after removing 170 duplicates and screening 509 titles and abstracts, 54 full texts were assessed and 37 studies met the inclusion and quality criteria. We propose a taxonomy that classifies these studies by the stage of symbolic execution they support, mainly path selection, constraint solving, state management and environment modeling, and by how AI is integrated, covering reinforcement learning, graph neural networks, neural constraint solving and large language models. The evidence is strongest for learned path selection and learned solver strategies, whereas learned state management, neural constraint approximation and learned function summarization remain largely unevaluated, and only four studies were evaluated on embedded, IoT or safety-critical software. We discuss applications in firmware and IoT security, automotive and aerospace verification and smart contract auditing, and derive a research agenda covering explainable AI guidance, cross-domain generalization and real-time adaptability for dependable embedded systems. |
