How much do HEP foundation models share? Measuring representational alignment on LHC data.
by
BSP 626
EPFL
A growing body of evidence suggests that independently trained neural networks converge towards related internal representations, even when their latent spaces come from different architectures, training data and objectives. Recent work traces this convergence to local neighbourhood similarities that persist across models. We test whether the same holds for foundation models in high-energy physics, using two models trained on the JetClass dataset and applied to real CMS collision data. If this hypothesis holds, the learned representations carry semantic value and offer a new basis for physics analysis. Latent-space methods have already been used in model-independent searches, but because the latent space was uncharacterised, these analyses were limited to distance- or density-based anomaly scores. Evidence that a shared latent space is semantically structured would make it possible to navigate that space directly in the search for new physics, potentially guided by its geometry.
Irene Bachiller Perea, Ismet Siral