Does adversarial hardening on SM background alone narrow sensitivity compared to signal-aware contrastive training?

Your question asks about a fairly specific technical comparison in machine learning for particle physics, but the preprints provided don't directly address whether adversarial hardening *on SM background alone* produces narrower sensitivity than signal-aware contrastive training.

That said, I can sketch what the literature suggests:

Signal-aware contrastive training [AG-2026.03-1712] constructs latent spaces by training simultaneously on Standard Model backgrounds *and* a diverse set of hypothesized Beyond the Standard Model signals. The key advantage is that the resulting low-dimensional space becomes "signal-sensitive"—it can discover BSM models both represented in training and, crucially, those not seen during training through interpolation and extrapolation.

Adversarial hardening on SM background alone [AG-2024.11-1226] improves generalization by reducing sharp minima in the loss landscape using white-box adversarial attacks. This makes models more robust to Monte Carlo artifacts, but the training objective doesn't explicitly encode signal structure—it only hardens the model against perturbations in the feature space.

The practical implication: contrastive training actively shapes the feature space to be sensitive to signal diversity, while background-only adversarial hardening passively improves robustness without building in signal awareness. You'd expect contrastive training to retain better sensitivity to a broad range of BSM scenarios, though the preprints don't directly benchmark these two approaches head-to-head.

The provided preprints don't contain a direct experimental comparison of these two training strategies on the same task.