The DeepTau identification algorithm, based on a Deep Neural Network model, has been developed to reduce the fraction of jets, muons and electrons misidentified as hadronically decaying tau leptons by the hadron-plus-strip algorithm in CMS.
Its recently deployed version for the Run 3 of LHC, DeepTau v2.5, has brought several improvements to the existing algorithm, most importantly the inclusion of domain adaptation techniques specifically designed to reduce simulation-to-data discrepancies in the high-score region of the tagger. In this work, the main novelties of DeepTau v2.5 are briefly discussed and its improved performance are presented. The new model delivers a reduced jet fake rate by $ \thicksim$50$\%$ across the regions of interest and, thus, sets a new improved baseline for the tau identification task. Lastly, the calibration of the tagger is performed and proves that, as sought, the new version is less sensitive to simulation mismodelling.