Scientists have developed VyPER, a novel geometric learning framework for event reconstruction in particle collider experiments. This system addresses the challenge of inferring the kinematics of short-lived particles from the stable final states recorded by detectors. Reconstruction is decomposed into two primary tasks: assigning measured jets and charged leptons to their parent particles, and predicting the kinematics of unmeasured neutrinos.

VyPER represents collider events as hypergraphs with a physics-inspired topology. It combines supervised classification of hyperedges for particle assignment with a diffusion model for predicting neutrino kinematics. Both tasks are unified through a joint loss function, optimizing reconstruction within a single framework. This approach enables more precise inference of the properties of ephemeral particles.

VyPER's performance has been demonstrated across various proton-proton collision processes, outperforming existing analytical and machine-learning-based reconstruction techniques. This advancement facilitates accurate event reconstruction across a diverse range of Standard Model physics processes, opening new avenues for precision measurements in the Higgs boson, electroweak, and top-quark sectors.