Anomaly detection aims to identify collision events that deviate from Standard Model expectations and is a key model‑agnostic strategy for new‑physics searches at the Large Hadron Collider. Searching for rare events in learned low‑dimensional representations avoids the need for hand‑crafted observables but sacrifices their immediate physical interpretation. For anomaly detection to enter a discovery pipeline, flagged events must be diagnosed locally to separate potential new physics from modeling or detector effects and to pinpoint observables for targeted follow‑up searches.
We evaluate supervised and self‑supervised neural embedding strategies on the COLLIDE‑2V dataset (750 million events simulated under realistic High‑Luminosity LHC conditions). The anomaly score is the reconstruction error of an autoencoder trained in the learned latent space. Across configurations and benchmark signals, supervised objectives that exploit simulation process labels generally outperform self‑supervised counterparts and, in their best settings, improve upon an autoencoder trained directly on kinematic features, raising signal efficiency at 10 % background acceptance for all held‑out processes.
We further introduce a local diagnostic pipeline: a Gaussian Mixture Model (GMM) is fitted to the Standard Model latent space, each component is profiled through high‑level observables without using process labels, flagged events are localized, and observables are ranked by their local discriminating power. The ranking recovers the defining kinematics of the held‑out processes; correlating the anomaly score with the same observables quantifies how closely the score tracks them, revealing cases where the most discriminating observable is not the one the score follows most closely.
Review: This study demonstrates that combining supervised learning with a local statistical model on large‑scale HL‑LHC simulations not only boosts anomaly‑detection performance but also provides an interpretable diagnostic framework, paving the way for practical physics analyses and targeted new‑physics searches.