TIGER: A Topology-Agnostic, Hierarchical Graph Network for Event Reconstruction

TL;DR

TIGER tackles the LHC event-reconstruction bottleneck by introducing a topology-agnostic, hierarchical graph network that exploits the common sequential two-body decay structure to reconstruct complex events without assuming a fixed topology. The model encodes particle objects, applies a diffusion transformer, and uses a two-stage hierarchical graph learning scheme with meta-nodes to assemble final-state particles, optionally followed by an event-level classifier in a multi-task setup. Compared with specialized baselines, TIGER achieves similar reconstruction efficiency but substantially higher purity across fully hadronic ttbar and semileptonic ttbarH topologies, and it outperforms SPANet in ttH vs tt+bb classification. This topology-agnostic framework offers robust generalization to mixed event topologies and can be extended with additional physics information, enabling more flexible analyses at the LHC. The work demonstrates the practicality of a unified, data-driven approach for event reconstruction and classification, with code publicly available for community use.

Abstract

Event reconstruction at the LHC, the task of assigning observed physics objects to their true origins, is a central challenge for precision measurements and searches. Many existing machine learning approaches address this problem but rely on a single event topology, restricting their applicability to realistic analyses where multiple signal and background processes with different structures are present. To overcome this, we present TIGER, a novel hierarchical graph network that is fundamentally topology-agnostic. By incorporating only the common underlying structure of sequential two-body decays, our model can reconstruct complex events without process-specific assumptions. This flexible architecture supports multi-task learning, enabling simultaneous event reconstruction and classification. TIGER thus provides a powerful and generalizable tool for physics analysis at the LHC.

Figures (3)

• Figure 1: Overview of the TIGER architecture. After an initial transformer-based encoding, a two-stage hierarchical learning module reconstructs the event. The first stage identifies intermediate particles ($W,H$) as meta-nodes, while the second stage combines them with other nodes to form top quarks. An optional final block pools features from the graph for event classification.
• Figure 2: Two event displays from the $t\bar{t}H$ test dataset. The left and middle matrices show the classification probabilities from the first stage of the hierarchical graph, while the right matrix depicts the second-stage incidence matrix, showing only the $W$ candidates selected by the first stage. The selected pairs are determined by the algorithm described in Section \ref{['sec:algorithm']}. The top panel shows a correctly reconstructed event, whereas the bottom panel illustrates a misassignment between a $b$-jet from the Higgs and the top.
• Figure 3: Receiver Operating Characteristic (ROC) curves for the $t\bar{t}H$ vs. $t\bar{t}+bb$ classification, comparing TIGER with SPANet baselines (results taken from spanet2).