Could contrastive learning with unseen signal topologies accidentally miss real new physics that doesn't interpolate well?

# On Contrastive Learning's Blind Spots for Exotic Physics

Your worry touches a real tension in machine learning for physics discovery: can a method trained on a fixed set of "hypothesized" new physics accidentally make *unknown* new physics invisible?

The short answer from the archive is nuanced. Contrastive learning does retain some sensitivity to unseen topologies—but with important caveats about what "unseen" means and how far that generalization stretches.

What the Archive Shows

The strongest reassurance comes from [AG-2026.03-1712], which explicitly tests this scenario. The authors train a signal-aware contrastive latent space on simulated Standard Model backgrounds plus "a diverse set of hypothesized Beyond the Standard Model signals"—then deliberately test models *not* in the training set. Their key claim: "the approach retains sensitivity to BSM models not present during training: interpolation and extrapolation to unseen signal topologies yield substantial improvements." This matters because it suggests the learned latent space isn't just memorizing the training signals, but capturing something more general about what "interesting" physics looks like in the data.

However—and this is crucial—that same paper focuses on *interpolation and extrapolation within a broadly similar class of signals* (supersymmetry, extended Higgs, resonances, FCNCs). All of these live in roughly similar kinematic/topological neighborhoods. The method hasn't been tested against, say, long-lived particles, exotic decay chains with many soft objects, or resonances that couple in completely unfamiliar ways.

The Harder Problem

A deeper issue emerges from [AG-2026.04-565], which studies how machine learning handles "tacit" and "implicit" structure. That work finds that models "demonstrate near-ceiling performance on explicit derivations within stable conceptual frames, but systematic degradation when tasks require reconstruction of omitted reasoning steps or reorganization of representations under global consistency constraints." Translate that to anomaly detection: if new physics violates the *implicit assumptions baked into the contrastive training* (e.g., which particle combinations are "natural" to cluster together), the model might not just miss it—it might actively suppress it.

[AG-2026.03-1363] hints at this too. Comparing explicit symmetry encoding against implicit data-driven learning, it finds they achieve "comparable performance"—suggesting that *either* way, you're relying on some framework. Neither automatically captures physics that doesn't fit the frame.

The Practical Safeguard

The most honest guidance comes from [AG-2025.06-1496], which uses *unsupervised* autoencoders for anomaly detection instead of supervised contrastive learning. By not training on *any* signal hypotheses, it sidesteps the problem—but at a cost: sensitivity to known signals is lower. The lesson: trading away some sensitivity to known signals might be the price of not accidentally erasing unknown ones.

---

Bottom line: The archive shows contrastive methods can generalize beyond their training signals *within a similar class*, but doesn't demonstrate robustness to truly alien topologies—and there's no free lunch. If you want insurance against missing exotic new physics, you may need to accept lower sensitivity to conventional BSM scenarios, or run multiple complementary methods (supervised contrastive + unsupervised anomaly detection).