One pretrained backbone for collider events: 500 million simulated collisions, 7.6σ on CMS open data
Synopsis
Collider analyses have trained a separate model per task; EVENET instead represents each collision as an unordered particle cloud and pretrains one backbone on 500 million simulated events with self-supervised plus physics-informed objectives, then adapts it through interchangeable task heads. In an anti-isolated ϒ→μ+μ− search on CMS Open Data, the fully pretrained variant reaches a median local significance of 7.6σ after calibration, above the published 6.4σ, and across a heavy-resonance mass grid its maximum SIC averages about 50% above training from scratch.
Interpretation
A reusable shared backbone is emerging for collider analysis: one pretrained model feeds classification, object-to-resonance assignment, final-state segmentation and diffusion-based reconstruction heads, spanning new-physics searches, open-data anomaly detection and a top-quark-pair quantum-correlation measurement. The prevailing practice designs a dedicated model for each analysis, trains it on purpose-built labelled simulation and validates it independently; earlier collider foundation models mostly operated on single jets or a few downstream tasks. On the (mX, mY) mass grid of the X→YHSM heavy-resonance search, individually fine-tuned EVENET–Full achieves the highest sensitivity across most of the grid and stays comparable elsewhere; averaged over the grid its maximum SIC improves by about 50% over EVENET–Scratch, 20% over XGBoost and 10% over TabPFN.
The pretraining benefit is tested on real collision data, not simulation alone. Earlier collider foundation models were mostly validated on simulation, with little transfer testing on real experimental data. In the anti-isolated ϒ→μ+μ− search on the 2016 CMS Open Data dimuon sample, calibrated EVENET–Full reaches a median local significance of 7.6σ (interval 7.6-0.4+0.4σ), above the published 6.4σ; EVENET–SSL reaches 7.4σ, while scratch training reaches only 1.5σ and needs a 74.4% calibration correction versus 11.6% for the full pretrained model.
Each event is treated as an unordered particle cloud, and the backbone is a point–edge transformer that adds k-nearest-neighbour message passing to attention to use local geometric structure. Pretraining runs in two stages: self-supervised masked-particle reconstruction first, then joint optimization of self-supervised generation, supervised generation and process classification, with the masking probability raised from 30% to 100%. The pretraining corpus holds about 5×108 simulated standard-model events retained after preselection, covering ten process categories including QCD, t tbar, W/Z+jets, dibosons and HSM(WW*), without cross-section weighting; model training took about 105 GPU-hours on 512 NVIDIA A100 GPUs, while downstream fine-tuning typically needs a few to tens of GPU-hours.
Perspective
The demonstrated scope is four downstream settings: the X→YHSM mass-grid search, anti-isolated ϒ rediscovery in CMS Open Data, the HSM→aa→4b exotic Higgs search and the dileptonic top-quark-pair quantum-correlation measurement; the pretraining samples contain no pile-up, so the mass-grid benchmark also tests domain shifts from pile-up and detector-response differences.
The authors state that EVENET is not yet a fully uncertainty-aware framework, so explicit modelling and propagation of detector and theoretical uncertainties remain necessary; predictions from different heads are not explicitly constrained to be mutually consistent; systematic stress tests under substantial simulation-to-data mismodelling have not been performed; and latent-space organization and interpretability are deferred to future work.