We hypothesize that applying SAE-based mechanistic interpretability to EEG foundation models can expose representational failures and thereby improve clinical trust.

Entities in the same semantic neighborhood but without a typed relation to this one — candidates for new edges or unrecognized duplicates.

• EEG foundation models achieve state-of-the-art clinical performance yet their internal computations remain opaque, constituting a barrier to clinical trust.claim0.836
  Motivating claim for the entire paper
• Clinical concepts in EEG foundation models fall into three operational regimes: selectively steerable, encoded but entangled, and non-encoded.claim0.812
  Interpretive claim summarizing the spectrum of concept steerability discovered.
• SAE features can be grounded in clinical taxonomy (abnormality, age, sex, medication) to benchmark monosemanticity and entanglement.claim0.800
  Claim that feature grounding enables interpretability metrics.
• SAE-based mechanistic interpretability will be superseded by manifold-based analysis for understanding semantic concepts within 24 months.prediction0.788
• Age and pathology are clinically entangled in EEG foundation model representations such that suppressing one concept inevitably corrupts the other.claim0.785
  A specific representational failure with direct clinical safety implications
• How are clinical concepts represented and steerable in EEG foundation models?question0.782
  Core research question driving the mechanistic investigation.
• A single SAE hyperparameter procedure driven by an intrinsic dictionary health audit transfers robustly across all three EEG transformer architectures.claim0.776
  Key methodological contribution claim about architecture-agnostic SAE tuning
• Circuits could act as an epistemic foundation for interpretability by breaking down model behavior into falsifiable statements about small subgraphs.claim0.768
  Normative vision for how the circuits agenda could resolve the pre-paradigmatic state of interpretability