A transformer-based encoder is proposed for the offline reconstruction of tau-lepton decays, with the aim of improving the estimation of the di-tau invariant mass. The model performs a regression of the transverse momentum of each tau candidate, effectively recovering part of the full kinematics, including the contribution from undetected neutrinos. The self-attention mechanism enables the network to combine information from the visible decay products together with event-level observables when inferring the tau kinematics. The performance is evaluated on Standard Model and beyond the Standard Model $\tau\tau$ samples and compared with the FastMTT algorithm, commonly used as a reference for di-tau mass reconstruction. The transformer-based method provides improved mass resolution and higher signal efficiency across a broad kinematic range, indicating its suitability as an alternative strategy for offline di-tau mass reconstruction at the CMS experiment.