AI-Native V2X and Internet of Vehicles Systems: Architectures, Learning Paradigms, and Open Challenges

Artificial intelligence is increasingly embedded into vehicular communication systems to support advanced applications such as autonomous driving, cooperative perception, and intelligent traffic management. While existing solutions often treat AI as an auxiliary component deployed on top of predefined architectures, emerging vehicular systems require a more fundamental integration of intelligence at the architectural level. This paper presents a comprehensive survey of AI-native V2X and Internet of Vehicles systems, where learning and intelligence are considered core design principles rather than add-on functionalities. The survey first reviews the evolution from traditional VANETs and IoV architectures toward AI-native V2X systems. It then introduces a structured taxonomy of AI-native architectures, highlighting how intelligence is distributed across vehicles, roadside units, edge servers, and cloud platforms. Major learning paradigms applicable to AI-native vehicular systems are analyzed, including centralized, distributed, federated, split, online, and multi-agent learning. The paper further reviews key enabling technologies that support AI-native V2X deployments, such as advanced V2X communications, edge computing, intelligent infrastructure, and sensing–communication integration. In addition, the survey systematically discusses performance trade-offs related to latency, accuracy, communication overhead, energy consumption, scalability, and robustness, as well as security, privacy, and trust challenges inherent to distributed vehicular intelligence. Finally, open research challenges and future directions are identified to guide the design of scalable, secure, and robust AI-native intelligent transportation systems. This work aims to provide a unified system-level perspective and serve as a reference framework for future research on AI-native vehicular networks.