Topology-Agnostic Event Reconstruction with Hierarchical Graph Neural Networks

We address the task of reconstructing unknown hierarchical structures from unordered sets of observations. To this end, we propose a hierarchical graph network that assembles such structures without relying on any topological priors. The model operates in stages: it first identifies low-level components and then infers higher-level assemblies. It also supports simultaneous set-level classification. Evaluated on a challenging particle physics benchmark, our method is the first in the field to be fully topology-agnostic, yet it matches the efficiency and achieves higher reconstruction purity than models constrained by predefined topologies.