Simulation-to-data transfer of scaling gains in flavor tagging

Determine whether the performance gains obtained by scaling supervised transformer-based flavor-tagging models and their simulated training datasets translate into corresponding gains on collision data.

Background

ATLAS and CMS are scaling flavor-tagging models to larger architectures and training samples, with studies suggesting substantial additional gains in simulated light-jet rejection. Yet the thesis notes that the rejection measured in data has not improved as favorably as the simulated rejection. This unresolved simulation-to-data transfer question could determine whether continued scaling of supervised cross-entropy training remains sufficient or whether self-supervised and data-driven objectives will be required.