Jet constituents provide a finer description of the radiation pattern inside a jet than global observables. In simulations for ATLAS Run-2 data (2015-2018), transformer-based taggers trained on low-level inputs significantly outperformed traditional approaches using high-level variables with conventional neural networks for quark–gluon discrimination. As the High-Luminosity LHC era approaches, with increased luminosity and center-of-mass energy, the ATLAS detector will undergo thorough/major upgrades, including an extended inner tracker that enhances coverage in the forward region, previously uncovered by a tracking detector. This study evaluates how these improvements enhance the accuracy and robustness of jet taggers, which is essential for key analyses such as vector boson fusion, vector boson scattering, and supersymmetry searches, where precise jet identification is critical for suppressing backgrounds.