Comprehensive Reconstruction of Collider Events with Hypergraph Representation Learning and Graph-Conditioned Diffusion
Synopsis
The work presents VyPER, a geometric learning framework that represents collider events as hypergraphs with a physics-inspired topology, combining supervised classification of hyperedges for assigning measured jets and charged leptons to parent particles with a diffusion model for predicting unmeasured neutrino kinematics, optimized jointly through a joint loss function within a unified framework, and compares its performance to existing analytical and machine-learning-based reconstruction techniques across several proton-proton collision processes, indicating that accurate event reconstruction is achievable across a diverse range of Standard Model physics processes.
Figure 6: Migration matrices for m t t ¯ m_{t\bar{t}} and cos ϕ \cos\phi observables for the four reconstruction approaches in the t t ¯ t\bar{t} (2L) process. The total count of each row is normalized to unity by convention, and the fractional per-row yield is annotated in each bin. Bins with under 5% of the row’s yield are not annotated. The binning for the m t t ¯ m_{t\bar{t}} matrices is not equal, with finer binning concentrated around the bulk of the distribution, but presented with equal spacing for readability.
arXivInterpretation
Event reconstruction is explicitly decomposed into two subtasks: assigning measured jets and charged leptons to parent particles, and predicting unmeasured neutrino kinematics. Compared with treating reconstruction as a single monolithic task, this provides a clear task decomposition so that two different kinds of missing information can be modeled separately. The decomposition is stated explicitly in the abstract as a framework-level design claim; the abstract provides no ablation or comparison data for the decomposition itself.
VyPER is proposed, representing events as hypergraphs with a physics-inspired topology and using supervised hyperedge classification for particle assignment. Compared with graph- or set-based representations, a hyperedge can connect multiple final-state objects at once, more naturally matching the physical structure in which several final states originate from the same parent particle. The abstract states the representation has a 'physics-inspired topology' and that supervised hyperedge classification is used; network architecture and hyperedge construction details are not expanded in the provided text.
A graph-conditioned diffusion model predicts neutrino kinematics and is optimized jointly with the classification task through a joint loss function. Compared with recovering neutrino momentum via analytical constraints or a separate regression, this uses generative diffusion modeling and shares an optimization objective with the assignment task, so the two reconstructions cooperate within one framework. The abstract explicitly mentions a 'diffusion model' and a 'joint loss function'; no specific metrics such as neutrino momentum resolution are reported in the abstract.
Comparison with existing analytical and machine-learning-based reconstruction methods across several proton-proton collision processes indicates accurate reconstruction is achievable across diverse Standard Model processes. Compared with reconstruction methods validated on a single process, this emphasizes cross-process applicability and points toward precision measurements in the Higgs boson, electroweak, and top-quark sectors. The abstract states 'across several proton-proton collision processes' and reports a comparison, but gives no specific process list, sample sizes, or numerical performance gaps.
Perspective
The framework targets event reconstruction in proton-proton collisions, aiming to map stable final states recorded by detectors back to the kinematics of short-lived particles in the hard scatter while handling both parent-particle assignment and missing neutrino kinematics; the abstract positions it as usable across several Standard Model processes and points toward precision measurements in the Higgs boson, electroweak, and top-quark sectors. Its applicability is bounded by the process range and comparison baselines described in the abstract.
The provided text contains only the title, authors, abstract, and bibliographic information, without the main body, figures, or numerical results, so it is not possible here to confirm per-process performance, the size of gaps relative to baselines, implementation details of hypergraph construction and diffusion sampling, or the weighting of the two tasks in the joint loss; these are points a reader should verify in the original. In addition, the abstract's conclusions rest on the authors' own comparison and await independent reproduction and testing on a broader range of processes.