New probes test what EEG foundation models actually use

An October 21 preprint introduces tests for spatial and spectral grounding in EEG foundation models. The authors perturb channel identity and signal characteristics, then examine both internal representations and downstream predictions across five models.

In one result, changing LaBraM's positional codes substantially shifted its embeddings while reducing motor-imagery classification accuracy by only 1.2 percentage points. Information can therefore be represented inside a model without being important to its final decision.

The study proposes evaluation methods rather than a clinical device. Its practical point is that accuracy alone can leave important questions unanswered about what a decoder relies on when conditions change.