Can machine learning replace the parton shower?

Nested-GPT uses a hierarchical Transformer to generate variable-multiplicity parton-shower sequences—the cascading quark splits that happen in high-energy collisions. The model learns the ordered, branching structure of emissions directly and decides when to terminate the shower dynamically, unlike flow-matching baselines that need the final particle count specified upfront. Tests on non-global-logarithm resummation show the method reproduces reference simulations within statistical noise, opening a path toward AI-accelerated shower generators.