Abstract: With the rapid development of deep learning and large model technologies, Brain-Computer Interface (BCI) systems are gradually entering a critical transition period from traditional task-specific command decoding to generalized intent understanding. Long constrained by the limitations of handcrafted feature engineering and the inherent characteristics of electroencephalogram (EEG) signals—such as high noise, non-stationarity, and significant individual differences—brain signal decoding models designed and trained for conventional BCI systems are usually only applicable to specific tasks, making it difficult to achieve efficient and natural interaction in open and generalized scenarios. The rise of deep learning, especially the Transformer architecture, has given birth to general-purpose brain signal large models pre-trained on large-scale unlabeled neural data, providing a novel approach to address the aforementioned bottlenecks. This paper systematically reviews the latest research advances of large model technologies in the field of BCI, introduces four core modeling paradigms: the autoregressive generative paradigm, masked signal modeling paradigm, cross-modal semantic alignment paradigm, and automated analytical agent paradigm, and conducts an in-depth analysis of the mathematical assumptions, core mechanisms, and engineering implementations underlying different paradigms. Finally, this paper explores how brain signal large models can upgrade BCI from a single-purpose speller to a "secondary brain" with multi-modal interaction, agent-based execution, and neural adaptive capabilities, thereby constructing a new ecosystem of human-machine symbiosis.