Could ML surrogates trained on LHC data generalize to beyond-Standard-Model particles they've never seen?

# ML Surrogates and Generalization Beyond the Standard Model

This is a sharp question that cuts to a real tension in collider physics: ML models trained on known processes may not recognize truly novel physics.

The honest answer is: it depends strongly on what you mean by "generalization," and current work shows both encouraging and cautionary results.

The Optimistic Case: Domain-Agnostic Learning

There's evidence that well-designed ML surrogates can extrapolate beyond their training data in limited ways. The key insight is that modern approaches try to learn *underlying structure* rather than memorize. For instance, [AG-2024.07-1501] demonstrates that different BSM theories can be mapped into a shared "latent space" where models cluster by their LHC phenomenology — meaning the same learned representation captures relationships between distinct BSM scenarios. This suggests the model is learning something fundamental about *how* new physics manifests, not just fitting the training set.

Similarly, [AG-2024.11-1157] found that graph neural networks, which explicitly encode particle relationships rather than just feature vectors, achieved >94% AUC on BSM signal detection. The point is that networks capturing *physics structure* (which particles interact with which) have better inductive bias for novel events than those operating on flat feature lists.

The Critical Limitation: Distribution Shift

But here's the catch. [AG-2025.11-1399] introduces the SURF method to validate whether generative surrogates actually capture the true data distribution — and finds that some models (like autoregressive GPT variants) *unphysically exaggerate* the discriminating power encoded in their training data. In other words, they hallucinate separation that isn't real. This is a cautionary tale: a model can seem to perform well on validation data drawn from the same distribution, yet fail catastrophically on something genuinely new.

The practical workaround comes from [AG-2024.02-1165], which proposes "Classifier Surrogates" trained only on public, high-level features with explicit Bayesian uncertainty quantification. The idea is to flag when the model encounters inputs that fall outside its learned regime — effectively admitting ignorance rather than confidently guessing.

Narrow Successes

One encouraging case is [AG-2024.09-1418], which showed ML-based tracking can reconstruct "quirky tracks" (curved, oscillating trajectories from hypothetical quirk particles) that traditional helical-trajectory code would ignore entirely. The technique isn't specific to quirks, suggesting the model learned general principles of track geometry. But this is still a relatively *constrained* new-physics scenario — the tracks still obey known physics, just in an unusual regime.

The Bottom Line

ML surrogates trained on LHC data can generalize to *variations* of BSM physics within their learned hypothesis space [AG-2024.07-1501], and when built on physical structure (graphs, conserved quantities), they have better extrapolation potential [AG-2024.11-1157]. But they are not general detectors of "anything beyond the Standard Model" — they risk confidently mispredicting when confronted with truly exotic signatures. The honest approach is uncertainty quantification [AG-2024.02-1165] and explicit validation that the surrogate captures the true data distribution [AG-2025.11-1399], not blind trust in held-out test accuracy.