Speaker
Description
We introduce the Global Particle Transformer (GloParT), a large-scale pretrained model developed within the CMS experiment for boosted-jet tasks. Initiated in mid-2022, GloParT is designed as a “foundation model” to support a wide range of tagging and regression applications. It is pretrained on jets from over 300 categories, learning a rich and generalizable representation of jet substructure. We demonstrate effective fine-tuning of GloParT for multiple downstream tasks, including tagging Standard Model top, W, and Z jets, as well as a “scouting GloParT” variant adapted to the scouting data stream for analysis and HLT applications. The model and its fine-tuned variants are now deployed within CMS. GloParT outperforms previous approaches across a range of simulation benchmarks and has already been used in several physics analyses, including the recent public HH→4b result. We conclude by discussing the broader potential of universal pretraining in high-energy physics and directions for future development.