Reinforcement Learning in Generative Multimodal AI: A Survey

Generative multimodal AI has seen remarkable progress, fueled by large-scale pre-training and the rise of powerful foundation models. However, these models are predominantly trained with supervised objectives, which often fall short of capturing task-specific nuances and human preferences. Consequently, Reinforcement Learning (RL) has emerged as a vital training framework for enhancing the alignment, controllability, and performance of multimodal models. This survey provides a structured, comprehensive, and in-depth review of RL research with the generative multimodal landscape. We first provide a formal formulation of RL in two dominant generative paradigms: autoregressive and diffusion models. We then systematically organize existing literature across the RL pipeline, focusing on four core pillars: data, rewards, algorithms, and applications. This taxonomy highlights key design choices, methodological tradeoffs, and emerging trends across different paradigms. Finally, we discuss open challenges and future research directions to guide the development of more reliable and aligned generative multimodal systems. Our curated resources are available at https://github.com/hu-zijing/Awesome-Multimodal-RL.