Continuous-Latent Predictive Modeling with Semantic Alignment for EEG-Language Foundation Models

Summary

EEG foundation models have shown that large-scale pretraining can support generalizable neural decoding across subjects, recording environments, and datasets, but the authors argue that the two dominant pretraining paradigms each have a structural flaw. Masked autoencoding tends to prioritize low-level signal reconstruction over task-relevant semantics, spending capacity on detail that downstream tasks do not need. Autoregressive modeling creates a mismatch between continuous neural dynamics and a discrete token space.