Beyond NL2Code: A Structured Survey of Multimodal Code Intelligence

The article presents a comprehensive survey on Multimodal Code Intelligence, highlighting the need for models that integrate visual inputs—such as screenshots and charts—with code generation and execution. It categorizes existing systems and benchmarks into four domains: Graphical User Interface, Scientific Visualization, Structured Graphics, and Frontier Tasks and Frameworks, emphasizing the importance of correctness in multimodal contexts. The authors propose future research directions focused on verification methods to enhance the reliability and functionality of multimodal code systems, which could significantly improve the development of evidence-grounded executable applications.