Hadronically decaying tau leptons are a challenging signature to study given it can be mimicked by quark and gluon jets, electrons, or muons. The identification of this signature via a convolutional neural network performed by CMS during the LHC Run 2 brought a massive improvement with respect to previous strategies. To further improve the identification and reconstruction of hadronic decays of tau leptons, CMS has deployed, as of the start of Run 3, a new algorithm: DeepTau v2.5. These proceedings present the performance resulting from these improvements in the network architecture using Run 2 data and showcases the use of domain adaptation techniques to improve the modelling of data.