HL-LHC: explaining abnormalities detected by neural networks locally.

A study of 750 million simulated collisions links latent anomalies to physical observables. The challenge is to make the warnings usable by physicists.

Neural networks can condense collisions into representations where unusual events are sought. However, this compression makes the physical interpretation less immediate. The study compares supervised and self-supervised objectives on COLLIDE-2V, under simulated conditions of the future high-luminosity LHC.

A local procedure groups the regions of the learned space and then classifies the observables capable of distinguishing the reported events. It also shows that a discriminant variable does not necessarily govern the anomaly score. The results concern simulations and reference signals, without announcing the discovery of a new particle.

Preprint filed on arXiv on October 9, 2026; results are those reported by the authors and remain to be independently confirmed.