How would attention stability across different random initializations distinguish true resonances from noise artifacts?

Your question asks about a methodological principle—using consistency across random seeds to validate physical discoveries—but it's framed in terms of machine learning attention mechanisms applied to resonance detection.

The provided preprints don't directly address this idea. However, I can explain why it's a sound principle using what they *do* discuss.

The closest match is [AG-2025.01-521], which shows how attention maps in gravitational wave analysis can reveal whether a model is learning "physically meaningful information" versus noise artifacts like glitches. The key insight: when a Vision Transformer's attention focuses on real signal features, those features should be robust; when it focuses on glitches, predictions become biased [AG-2025.01-521].

Your suggestion extends this logic: if you retrain the same architecture with different random weight initializations, a *true resonance* (a real physical feature) should consistently capture the model's attention across runs, while a *noise artifact* would be attended to unpredictably—sometimes heavily, sometimes not—because its salience depends on random quirks of optimization, not physics.

This mirrors a principle from [AG-2025.09-124], which finds that ML detection sensitivity "fluctuates due to dataset variability" [AG-2025.09-124]. Real signals should be stably detected; spurious patterns should vary with initialization.

The archive preprints don't explicitly evaluate attention stability across random seeds as a signal-validation technique, so I cannot cite direct experimental support for this specific proposal.