From LLMs to Agentic and Embodied AI for Next-Generation Intelligent Vehicular Communications: A Comprehensive Survey

Abstract

The internet of vehicles (IoV) is rapidly evolving into an artificial intelligence (AI)-driven, multi-agent collaborative ecosystem, yet traditional communication paradigms struggle to address the challenges posed by dynamic network topologies, stringent resource constraints, and heterogeneous service demands. To bridge this gap, this survey systematically reviews the transformative potential of three emerging AI paradigms, including large language models (LLMs), Agentic AI, and Embodied AI, in reshaping future vehicular communications. We first introduce the concept of IoV and cellar vehicle-to-everything (C-V2X), while highlighting challenges IoV faces. Subsequently, we detail the fundamentals of these AI paradigms: architectural innovations, training methodologies, and prompt engineering for LLMs, and the core modules of Agentic AI and Embodied AI systems. The survey then analyzes domain-specific adaptation strategies, such as model compression techniques, specialized dataset construction, and cloud-edge-vehicle collaborative deployment for LLMs; distributed multi-agent collaboration frameworks for Agentic AI; as well as reconstruction of perception modules, world models, and executors in Embodied AI systems. Furthermore, we investigate practical applications across critical vehicular communication scenarios, covering beamforming, resource allocation, semantic communication, network optimization, and multi-agent collaboration. To validate these theoretical frameworks, we present a representative case study on Embodied AI-enhanced vehicular networks. Finally, we identify key challenges and outlines future research directions.