Neurosymbolic code auditing

Abstract

Code auditing examines whether an implementation satisfies target properties such as reliability, security, performance, and design intent. Classical static analysis provides rigorous foundations for automated auditing, but it remains difficult to apply broadly in real-world software. The emergence of large language models (LLMs) and agents creates new opportunities for leveraging semantic cues that are embedded in source code, documentation, and other software artifacts yet difficult for classical static analyzers to derive. This survey presents a comprehensive study of neurosymbolic code auditing, an emerging synthesis that integrates LLM-driven reasoning with classical static analysis and other structured program reasoning techniques. By formalizing the cognitive processes underlying manual code auditing, this study conceptualizes an agent-centric framework that serves as a unifying foundation for neurosymbolic analysis. Specifically, this study categorizes LLM-enabled primitive analyses over both code and non-code artifacts, addressing the fundamental reasoning tasks routinely performed by human auditors. Furthermore, the authors review a broad spectrum of neurosymbolic architectures that combine symbolic tools with LLM-based reasoning, illustrating how these systems replicate and extend the practical strategies employed by human auditors. Lastly, the authors synthesize the survey around three recurring bottlenecks, namely, observation, verification and memory, then outline future directions and the persistent technical challenges.