<p>Reverberation, non-line-of-sight propagation, and temporal bias in enclosed spaces frequently impair the efficacy of acoustic indoor localization, which is attractive for infrastructure-light location. In this paper, an event-token sequence modeling framework for 2D localization is presented and a two-base-station ultrasonic chirp scenario is studied. Each chirp event is represented by residual offset-corrected time-difference-of-arrival (TDoA), Doppler-inspired relative-speed difference, confidence, and inter-event timing following band-limited preprocessing and matched-event creation. A compact scenario- and field-aware Transformer regressor processes these heterogeneous descriptors after they have been discretized into fixed-length token sequences. To separate no-refit generalization from same-trajectory adaptation, we use an event-disjoint edge-block protocol in which held-out edge blocks share no underlying event indices with the training or validation windows. Under this protocol, the event-token Transformer achieves a mean-of-fold localization error of 1.521 m and a P90 error of 1.602 m, outperforming the scalar hyperbolic TDoA grid baseline (2.671 m), Kalman filter (2.263 m), EKF TDoA model (2.583 m), and Particle Filter TDoA baseline (2.063 m). The clean continuous-input Ridge diagnostic achieves the lowest mean error of 1.198 m, indicating that continuous event descriptors remain strong in this small-data setting. Thus, the Transformer is interpreted as a compact and competitive sequence model, not as evidence that tokenization is universally superior. The same-split refit result is retained only as an oracle same-trajectory adaptation upper bound. Overall, the findings show that compact event-token sequence modeling is a useful and efficient method for Two-BS acoustic indoor localization. The results are interpreted cautiously and bolstered by bootstrap-based uncertainty analysis due to the limited held-out test set.</p>