An unsupervised machine learning framework is developed for identifying anomalous events in proton–proton collisions at √𝑠 = 13 TeV, trained on the ATLAS Run-2 open dataset collected during 2015–2016. The analysis utilize a model independent approach based on TAB-transformer autoencoder trained to characterize Standard Model event topologies and identify deviations. The anomaly scores show clear separation between background and MC generated beyond-Standard Model and rare benchmark processes. Clustering of high-scoring events reveals distinct kinematic structures consistent with signatures of rare and BSM dynamics. The result demonstrates the
sensitivity of the Tab-Transformer-based anomaly detection framework to diverse rare and BSM signatures, enabling data-driven exploration beyond traditional search strategies.