Representation, Alignment, and Generation: A Comprehensive Survey of Foundation Models for Non-Invasive Brain Decoding

The ability to decode human thoughts, intentions, and perceptions directly from non-invasive brain recordings holds transformative potential for healthcare, communication, and human-computer interaction. However, translating the safety and scalability of methods like fMRI, EEG, and MEG into real-world utility has traditionally been hindered by low signal-to-noise ratios, limited spatial-temporal resolution, and the difficulty to collect large-scale high-quality data from an individual user. Recently, the emergence of Foundation Models (FMs)--large-scale, pre-trained artificial intelligence architectures--has catalyzed a paradigm shift in overcoming these obstacles. This survey provides a comprehensive overview of how FMs are redefining the boundaries of non-invasive brain decoding. We propose a unified methodological framework that synthesizes recent advancements into a coherent process: extracting robust, transferable representations from noisy neural signals; aligning these signals with the rich semantic spaces of pre-trained vision and language models; and leveraging powerful conditional generative priors to reconstruct high-fidelity outputs. We systematically review state-of-the-art applications across three key domains: visual reconstruction, language and speech decoding, and auditory processing. Furthermore, we critically examine the persisting challenges of computational efficiency, cross-subject generalization, and privacy governance. By mapping the current landscape and identifying key gaps, this work outlines a strategic research agenda aimed at transitioning FM-driven neurotechnology from laboratory proofs-of-concept to reliable, real-world applications.