By E. Bothmann et al. · arXiv:2608.21338
The first fully GPU-resident event-generation chain: normalizing-flow phase-space proposals trained online, exchanging points and target densities with the Pepper parton-level generator directly in device memory — no host round-trips. Deployed across all subprocesses of full pp collisions with many jets.
Why it matters · This is the direction of travel for generator computing — and directly adjacent to our own GPU event-generation program.
Caveat: Parton level: showering, hadronization and detector simulation still live outside the GPU loop.
By C. Jesús-Valls · arXiv:2608.25107
A neutrino interaction generator with exact derivatives propagated through the nuclear ground state, hard scattering and the stochastic intranuclear cascade — making event weights and their gradients first-class outputs rather than afterthoughts obtained by reweighting.
Why it matters · Differentiable simulation unlocks gradient-based tuning and inference; expect this pattern to migrate to collider generators.
Caveat: Physics choices follow ACHILLES; demonstrated for neutrino scattering, not yet a general-purpose generator.