In this project we will investigate a wide range of deep-learning paradigms and algorithms for anomaly searches at the LHC. Semi-supervised, weakly-supervised, or unsupervised training can be used to achieve sensitivity to weak or complex New Physics signals (anomalies) in a largely model-independent way. These concepts will be developed for QCD jets, i.e., for analysis objects that are present in huge quantities at the LHC and whose dynamics is well understood within the Standard Model. We will systematically explore various machine learning architectures, such as autoencoders, normalising flows and transformer networks, trained using different jet-data formats, ranging from images to point clouds and higher-level observables such as energy flow polynomials.