Parameter-efficient fine-tuning for multimodal fusion: A survey of injection-driven integration mechanisms

Multimodal information fusion has become a fundamental paradigm for enabling comprehensive understanding and generation across heterogeneous data sources. Recent advances in multimodal foundation models (MFMs) have significantly improved cross-modal learning capabilities; however, their massive parameter scale poses critical challenges for efficient adaptation, including prohibitive computational costs and the risk of disrupting pre-trained cross-modal alignments. In this survey, we present a comprehensive review of parameter-efficient fine-tuning (PEFT) for multimodal models from the perspective of information fusion. Specifically, we view PEFT as a set of parameterized fusion mechanisms that regulate how multimodal information is aligned, interacted, and integrated within pre-trained architectures. Based on this view, we systematically categorize existing approaches into four paradigms: additive injection for explicit feature-level fusion, reparameterization for implicit cross-modal coupling, input-level parameterization for semantic-guided fusion, and advanced collaborative methods for adaptive and context-aware fusion. Furthermore, we analyze these paradigms from multiple perspectives, including task adaptability, parameter efficiency, and structural flexibility, highlighting the trade-offs in designing multimodal fusion strategies. We also discuss key challenges in multimodal fusion, such as mitigating modal disparity and enhancing cross-modal alignment under constrained adaptation. Finally, we outline promising future directions toward scalable, adaptive, and intelligent multimodal fusion systems. This survey provides a unified perspective on multimodal PEFT through the lens of information fusion, offering practical insights for developing efficient and robust multimodal learning systems.