EmBrace: A Collective Knowledge Fusion Framework Toward Unified EEG Foundation Models

Electroencephalography (EEG) foundation models (EFMs) have achieved strong performance across a wide range of downstream EEG tasks via pretraining and fine-tuning. Through empirical analysis, we observe that (i) no single EFM consistently dominates all tasks, yet identifying the task-specific optimal model by fine-tuning all EFMs introduces substantial computational overhead; and (ii) models with inferior task-level performance still exhibit strengths at the sample level as distinct architectures induce diverse inductive biases. These observations motivate EmBrace, a representation-centric framework for sample-aware knowledge fusion that avoids the constraints of parameter-level or output-level alignment. EmBrace synchronizes discriminative intermediate representations into a unified manifold and adaptively weights multiple EFMs at the sample level while selecting the most compatible model as the carrier. Extensive experiments across multiple EEG benchmarks demonstrate that EmBrace consistently improves over SOTA EFMs and generalizes effectively under cross-task settings.

EEG records the brain’s electrical activity and is used in many applications, from sleep analysis to brain-computer interfaces. Recent large EEG models can be pre-trained on broad brain-signal data and then adapted to new EEG tasks, but no single model works best for every task. A researcher could try every model and pick the best one, but this is costly and slow. We also found that a model that looks weaker on average can still make useful judgments on particular samples, because different model designs notice different patterns in the same brain signal. To use this hidden diversity, we introduce EmBrace, a method that combines the internal representations learned by multiple EEG foundation models. Instead of forcing models to share parameters or only merging their final predictions, EmBrace aligns what they have learned into a common space and decides, sample by sample, which models should contribute more. It also selects the most suitable model to represent the combined knowledge for each case. Across several EEG benchmarks, EmBrace improves performance over strong existing EEG foundation models and remains effective when tested on new task settings. This makes it easier to build accurate and efficient EEG systems without repeatedly retraining every available model. To support future research, we have made the code publicly available.