What it is
Collider experiments usually optimise a separate model for each analysis. EVENET is instead an event-level foundation model, pretrained on 500 million simulated collision events with self-supervised learning and physics-informed supervision, that builds one shared representation from the variable-length set of physics objects in each event and is then fine-tuned for each task. Across searches for heavy resonances and exotic Higgs decays it outperformed established and state-of-the-art baselines, particularly when training data were scarce. In an exotic Higgs search it reached a sensitivity score of 4.1, against 1.6 for the same network trained from scratch and 1.4 for SPANet, and with only 5% of the reference training set it already beat the best results both baselines reached with 150%. On real data it rediscovered the Upsilon meson in CMS Open Data, with a median local significance of 7.6 sigma against the 6.4 sigma of a published earlier search. In simulated top-quark-pair events it also extracted a quantum-correlation observable to a precision near 3.33%, about 37% better than a 5.26% benchmark, and stable under the tested systematic variations.
Why it matters
Collider analyses usually train a new model for every search; this one backbone, pretrained once, adapted to tasks it never saw in pretraining and to real CMS data. Across a grid of 121 mass points for a heavy new particle decaying to a lighter one and a Higgs boson, it gave the highest sensitivity over most of the grid and matched the tabular baselines elsewhere, improving the best sensitivity on average by about 50% over the same network trained from scratch, 20% over XGBoost and 10% over the tabular foundation model TabPFN. The authors present it as a reusable backbone for diverse collider analyses, paid for once: pretraining took of the order of 100,000 GPU-hours on 512 A100 GPUs, before counting the generation of the simulated events. The model, its weights and its pretraining data are public.
Underlined numbers link to their source. Every metric and quoted figure is listed under Sources and data below.
Filed underParticle physics theoretical and experimental studies, High-Energy Particle Collisions Research, Particle Detector Development and Performance
