Speaker
Description
Hadronically decaying tau leptons are essential signatures in many ATLAS measurements and searches, including studies of the Higgs boson and potential new physics. Their identification is particularly challenging due to the large background from quark and gluon initiated jets in proton-proton collisions. This poster presents an overview of hadronic tau reconstruction and identification in the ATLAS experiment, with a focus on the newly developed GNTau algorithm, a transformer-based neural network that combines information from charged-particle tracks, calorimeter energy deposits, and high-level observables. Public Run-3 performance studies demonstrate that GNTau significantly improves the rejection of misidentified jets compared to the previous recurrent neural network (RNN) based approach while maintaining the same signal efficiency across a wide range of transverse momentum, pseudorapidity, and pileup conditions. These advances enhance the sensitivity of ATLAS analyses involving tau leptons and represent an important step toward precision measurements and future discoveries at the High-Luminosity LHC.