Speaker
Description
Hadronically decaying tau leptons are key signatures in precision measurements and searches for new physics at the LHC, including Higgs boson studies and scenarios beyond the Standard Model. Their reconstruction is challenging due to complex decay modes and large backgrounds from quark- and gluon-initiated jets.
This talk reviews recent advances in hadronic tau reconstruction and identification at the CMS experiment, with a focus on modern machine learning techniques. Significant progress has been achieved in offline tau identification using deep learning, particularly the latest DeepTau algorithm, which employs convolutional neural networks with domain adaptation to improve discrimination power and reduce data–simulation discrepancies. Alternative approaches based on jet reconstruction with graph neural networks and particle transformers are also discussed, and their Run 3 performance is compared.
At the trigger level, new machine-learning-based algorithms have been deployed in the high-level trigger to efficiently select hadronic tau signatures under the challenging Run 3 conditions of high pileup and strict latency constraints. Dedicated strategies for non-standard tau topologies are presented, including boosted, displaced, and low-transverse-momentum taus reconstructed in the Scouting data stream. Finally, recent machine-learning-based, data-driven methods for estimating jet→tau backgrounds are discussed, demonstrating improved accuracy and robustness across phase space.
| I read the instructions above | Yes |
|---|