Abstract: Vision–Language–Action (VLA) models constitute a critical pathway toward embodied intelligence, with their core function being the seamless transformation of multimodal perception and understanding into concrete actions in the physical world. Action representation and generation strategies, serving as the pivotal bridge between perception and execution, face significant challenges stemming from high-dimensional continuous spaces, the diversity of action modalities, and the stringent demands of real-time robotic control. This survey provides a systematic review of the evolution, key methodologies, and future directions of action representation and generation in VLA models. We analyze discrete and continuous representations in depth, and examine autoregressive, non-autoregressive, and hybrid generation strategies, highlighting their inherent trade-offs in terms of precision, diversity, and efficiency. In addition, we cover emerging high-efficiency approaches designed for real-time control, such as hybrid generation architectures. Finally, we present a comparative synthesis of the current technological landscape and outline frontier challenges and research opportunities, including integration with world models and the development of generalizable representations across heterogeneous robotic embodiments. This work aims to provide a comprehensive reference for advancing more general, efficient, and reliable embodied agents.