Speaker
Description
Reconstruction and simulation at a collider experiment are long chains of specialised algorithms, each tuned to a single step. We ask whether one model can serve many of those steps at once, while still producing the intermediate objects (tracks, calorimeter cells, clusters, particles and jets) that make the chain interpretable. We represent every object in an event in a single shared token vocabulary built with a residual vector-quantised autoencoder, and train one model to map any subset of these modalities to any other. A task is then only a choice of which modalities to provide and which to request: particle flow, detector simulation, unfolding and calibration are all directions through the same set of weights. We train over all valid directions, with a decoder emitting tokens either in parallel or autoregressively. With tokenisation, both architectures train stably with little tuning. To judge fidelity beyond one-dimensional distributions we use classifier two-sample tests conditioned on the event's inputs to ask whether a prediction is high-quality per-event. Under this test the autoregressive decoder's detector simulation is conditionally indistinguishable from \textsc{Geant}4, while the single-pass decoder is not. Additionally, state-of-the-art reconstruction that looks excellent by every standard metric is seen to be noticeably worse than the autoregressive model introduced in this work. The model also generalises to modality combinations never seen in training, and improves when a task is decomposed into a physically motivated chain of steps.