CoCoT-EEG: Contrastive pretraining beats masked reconstruction for EEG decoding

Researchers Gabriel Mahuas and colleagues have developed CoCoT-EEG, a novel contrastive-pretrained model that combines multiscale temporal convolution input layers with Transformer encoder blocks for decoding electroencephalogram (EEG) signals. Unlike many recent large-scale EEG foundation models that rely on tokenizing raw signals followed by masked reconstruction pretraining, CoCoT uses contrastive learning—a strategy that the authors argue is better suited for EEG data, which suffers from high noise amplitude and information confined to narrow frequency bands. Their systematic ablations show that contrastive pretraining yields models that match or exceed the performance of larger reconstruction-pretrained models across extensive benchmark datasets with heterogeneous electrode configurations. Notably, even when trained from scratch, CoCoT outperforms previous single-task decoding models and rivals the performance of many pretrained models, highlighting its architecture's flexibility and data efficiency.

The paper, submitted to arXiv on July 10, 2026, provides two key insights. First, contrastive learning offers a viable alternative to masked reconstruction for building EEG foundation models, particularly given the domain-specific noise characteristics. Second, the multiscale temporal convolution front-end is critical for capturing both fine-grained and broad temporal features, which enables the model to handle diverse electrode layouts without retraining. The authors suggest that these findings prompt further investigation into alternative large-scale pretraining strategies for EEG. For professionals in brain-computer interfaces and neuroscience, CoCoT-EEG represents a significant step toward more practical, data-efficient models that can decode neural signals with higher accuracy and lower computational cost.