PanopTag: Simultaneously Tagging All Jets in a Particle Collision Event

TL;DR

PanopTag tackles the limitation of independent, single-jet tagging by proposing an event-level, multi-jet tagging framework. It employs a DETR-inspired encoder–decoder architecture where particle flow object embeddings are processed once per event by an Event Encoder, and jet kinematics are used as queries in a Jet Query Decoder to produce per-jet classifications via cross-attention. The approach yields marked improvements in heavy-flavor tagging over state-of-the-art single-jet baselines and demonstrates robustness to topology shifts not seen during training, highlighting the benefit of shared event context. This work suggests a path toward unified, foundation-model-like modeling for collider jet analyses, with potential speedups and broad applicability to tasks such as pileup jet ID and jet calibration.

Abstract

Jet tagging, identifying the origin of jets produced in particle collisions, is a critical classification task in high-energy physics. Despite the revolutionary impact of deep learning on jet tagging over the past decade, the paradigm has remained unchanged. In particular, jets are classified independently, one at a time. This single-jet approach ignores correlations, overlaps, and wider event context between jets. We introduce PanopTag, a new paradigm for jet tagging that departs from traditional single-jet tagging approaches. Rather than classifying jets independently, PanopTag simultaneously tags all jets by employing an encoder-decoder architecture that uses jet kinematics as queries to cross-attend to particle flow object embeddings. We evaluate PanopTag on heavy-flavor $(b/c)$-tagging and demonstrate remarkable performance improvements over state-of-the-art single-jet baselines that are only accessible by exploiting event-level features and correlations between jets.

Figures (6)

• Figure 1: A high-level overview of the PanopTag architecture. PFOs are processed through the Event Encoder to produce embeddings, which serve as keys and values in a cross-attention mechanism. Simultaneously, jet kinematics are passed through an MLP to produce jet embeddings that act as queries, with the outputs after classification heads predicting class labels.
• Figure 2: Illustration of the Event Encoder block. PFOs are first embedded using an MLP, then processed through two stages: (i) local feature extraction using EdgeConv layers for neighborhood aggregation, and (ii) global feature extraction using ISAB layers for set-based attention. Skip connections are added to preserve information flow across both stages.
• Figure 3: The background rejections ${\text{Rej}_{X}}$ as a function of the $b$-jet (left) and $c$-jet (right) tagging efficiencies $X$ for PanopTag (blue), ParT (orange), and ParticleNet (green) models. The ratio of the PanopTag background rejection to the baselines is shown in the bottom panel of each plot.
• Figure 4: The background rejections ${\text{Rej}_{X}}$ as a function of the $b$-jet efficiency comparing the PanopTag (solid) and ParticleNet (dashed) models for the $t\overline{t}$ (blue) and $ZH$ (orange) processes. The ratio of the PanopTag background rejection to ParticleNet is shown in the bottom panel for both processes.
• Figure 5: Representative event display in the $(\eta,\phi)$ plane. Markers show reconstructed PFOs with size proportional to $p_\mathrm{T}$. Dashed circles (of $\Delta R=0.4$) indicate jets, colored by flavor label. For each jet, we select the top-10 PFOs with the largest attention weights and draw line segments from the jet axis to these particles, visualizing the constituents most emphasized by the model. The width of line segments are proportional to the attention weights.