Abstract: Large-scale pre-trained models have demonstrated remarkable capabilities in fields such as natural language processing. Fine-tuning these models is a common approach to adapt them to downstream tasks. However, full fine-tuning of large models faces severe challenges, including high computational costs and substantial storage requirements. Parameter-efficient fine-tuning (PEFT) has emerged as a key technical paradigm to address these issues by introducing or selecting a minimal number of trainable parameters, significantly reducing computation and storage overhead while effectively preserving the core capabilities of models. This paper provides a systematic review of the mainstream methodologies, key technological advances, and development trends within the PEFT field. First, we categorize existing methods into four major paradigms: Additive, selective, reparameterized, and hybrid fine-tuning, offering an in-depth analysis of their core mechanisms, performance characteristics, application scenarios, and strategic advantages. Furthermore, we focus on the technical evolution of PEFT, analyzing the intrinsic principles behind its development. We summarize a clear technical trajectory: The field is moving from isolated method innovations toward a sophisticated tri-balance of storage, computation, and performance, and further advancing into unified frameworks emphasizing automation, intelligence, and hardware-software co-design. Furthermore, we conduct a systematic quantitative comparison of representative methods from various PEFT categories, evaluating their performance and parameter efficiency on uniform models and datasets. In addition, this survey also covers the extended applications of PEFT technology in fields such as vision, speech, and cross-modal models, demonstrating its broad applicability. Finally, we discuss promising future research directions to facilitate the development of more efficient and adaptable fine-tuning techniques for large-scale models across diverse tasks.