Total: 1
We introduce optimal transport (OT) as a physics-based intermediate event representation for weakly supervised anomaly detection. With only $0.5\%$ injection of resonant signals in the LHC Olympics benchmark datasets, the OT-augmented feature set achieves nearly twice the significance improvement of standard high-level observables, while end-to-end deep learning on low-level four-momenta struggles in the low-signal regime. The gains persist across signal types and classifiers, underscoring the value of structured representations in machine learning for anomaly detection.