Speakers
Description
Most machine learning at colliders operates downstream of a theory prediction: reconstruction, simulation, inference, and search all take a Standard Model expectation as input. This talk is about the other side — how AI is changing the calculations themselves. In scattering amplitudes, transformers and symbolic regression now assist bootstrap calculations, integral reduction, and the extraction of analytic structure. This is the setting where a learned answer can be checked exactly.Precision QCD faces a different obstacle: fixed-order and resummed calculations reach high accuracy for a limited set of observables, while comparison to data requires complete event distributions, supplied by parton showers at lower formal accuracy. Here I discuss information-theoretic approaches, including maximum-entropy reweighting constrained by moments of high-precision observables. In BSM phenomenology, machine learning enters both in searching the space of models, where agents now propose candidates on their own (neutrino flavor being furthest along), and in extracting more from the same data through learned likelihood ratios in EFT fits.