How much BSM morphological diversity in training is needed before adversarial robustness stops hurting rare signal detection?

Your question touches on a genuine tension in machine learning for rare-signal detection, but the provided preprints don't directly address the specific tradeoff you're asking about — namely, *how much* Beyond-Standard-Model morphological diversity in training is needed before adversarial robustness stops *hurting* sensitivity to new signals.

Let me unpack what the archive does say about related pieces:

The core tension: Anomaly detection for BSM physics faces a dilemma. Training on only Standard Model backgrounds leaves classifiers vulnerable to systematic errors and Monte Carlo artifacts [AG-2024.11-1226]. Yet adversarial robustification can make classifiers *less* sensitive to rare deviations — the classifier learns to be robust to perturbations rather than maximally responsive to signal morphology.

What helps generalize to unseen signals:

Signal-aware contrastive learning trained on a *diverse set* of hypothesized BSM signals substantially improves generalization, even to signal topologies absent from training [AG-2026.03-1712]. The critical claim: "For signals represented in the contrastive training data, the method can elevate discovery sensitivity from previously inaccessible levels to the discovery regime. Critically, the approach retains sensitivity to BSM models not present during training."

Multi-background representation learning, rather than training on a single dominant background, improves robustness *and* detection fidelity [AG-2024.01-1031]. Decorrelation via optimal transport maintains signal sensitivity while removing classifier dependence on background misspecification [AG-2024.09-1051].

What's missing: None of these works systematically quantify the threshold — the morphological coverage needed before robustness regularization stops degrading sensitivity to truly novel signals. That's an open experimental question.