Recent advancements in the tau reconstruction and identification techniques in CMS

TL;DR

The CMS tau program tackles the challenge of reconstructing hadronic tau decays ($\tau_\mathrm{h}$) across standard and nonstandard topologies. It surveys advances including the DeepTau v2.5 CNN with domain adaptation to improve data/MC agreement, and alternative unified jet-tagging approaches using graph neural networks and transformer architectures. It also details specialized algorithms for displaced and boosted taus (DisTau, Boosted DeepTau) and low-$p_T$ contexts (TauNet) to extend coverage, along with online improvements at the High Level Trigger (L2TauNNTag and online DeepTau). Together, these developments enhance tau-identification efficiency and expand the physics reach for Higgs and Beyond-Standard-Model searches at Run 3 and beyond, including trigger-level gains.

Abstract

Tau leptons play a crucial role in studies of the Higgs boson and searches for Beyond the Standard Model physics at the present LHC and in its high luminosity upgrade. This talk presents the latest advancements in the reconstruction and identification of hadronic decays of tau leptons at the CMS experiment, both at the online and offline levels. The tau identification algorithm deployed for the early Run 3 data-taking period, based on a deep convolutional neural network with domain adaptation, showcases significantly improved discrimination of genuine hadronic tau decays against mis-identified quark and gluon jets, electrons, and muons. During live data-taking, a simplified version of the algorithm is used to select events with tau leptons at the High Level Trigger (HLT). The performance and calibration of both algorithms using early Run 3 data are presented. Many CMS physics analyses involving tau leptons are expected to benefit from these improvements. Alternative approaches to identify hadronic taus combined with jet flavour, based on graph neural networks and particle transformers, are also covered. Additionally, the dedicated techniques used to reconstruct and identify displaced tau leptons originating from long-lived particle decays using graph neural networks are discussed.

Figures (5)

• Figure 1: Distribution of the DeepTau discriminator against quark and gluon jets before (left) and after (right) domain adaptation, for the dataset used to train the domain adaptation CMS:2025kgf.
• Figure 2: Visible mass distribution for a muon and $\tau_\mathrm{h}$ system for 2024 (left) and 2025 (right) data CMS-DP-2025-074. For 2024, data correspond to an integrated luminosity of 109$\,\mathrm{fb}^{-1}$. For 2025, data correspond to an integrated luminosity of 26.5$\,\mathrm{fb}^{-1}$, collected up to the 23$^\mathrm{rd}$ of July 2025.
• Figure 3: ROC curves illustrating the performance of the DeepTau, PNet, and UParT algorithms to discriminate quark and gluon jets (left) and electrons (right) misidentified as $\tau_\mathrm{h}$ candidates CMS-DP-2025-073. The efficiencies are evaluated taking into account the preceding reconstruction and pileup mitigation algorithms used, labeled HPS, CHS, and PUPPI CMS:2018jrdCMS:2020eboBertolini:2014bbaCMS-DP-2024-043.
• Figure 4: Left and middle: Jet rejection vs $\tau_\mathrm{h}$ identification efficiency for the DisTau graph network at high displacement CMS-DP-2024-053, and for the Boosted DeepTau CNN algorithm at high $p_{\mathrm{T}}$CMS-DP-2025-047, respectively. Right: calibration of the low-$p_{\mathrm{T}}$$\tau_\mathrm{h}$ reconstruction algorithm using the $\Upsilon\rightarrow\tau_\mu$$\tau_\mathrm{h}$ resonance Collaboration:2905110.
• Figure 5: Efficiency as a function of the $\tau_\mathrm{h}$ candidate $p_{\mathrm{T}}$ evaluated on the di-$\tau_\mathrm{h}$ trigger algorithm at the L1+L2 stage (left) and with the full L1+HLT system (right) CMS-PAS-TAU-24-002.