AlphaHits extends tracking to whole events: no transverse-momentum cut, a TrackML score of 0.959
Synopsis
Track reconstruction at the High-Luminosity LHC usually keeps only charged particles above a transverse-momentum threshold, and that cut discards more than ninety percent of the particles. AlphaHits-TrackFormer instead processes all pixel and strip hits in complete TrackML events without a transverse-momentum selection on target tracks, reaching a TrackML score of 0.959 on 820 test events. The same framework, pretrained on tracking and then jointly fine-tuned, uses a shared hit-level representation to output tracks, jets, jet flavour and event-level observables together, with a Higgs classification AUC of 0.944.
Interpretation
AlphaHits-TrackFormer widens charged-particle tracking from a high-transverse-momentum subset to all reconstructable particles in a complete event, taking every pixel and strip hit as input. Earlier tracking studies often imposed a transverse-momentum threshold near 1 GeV on target tracks, a cut that removes more than ninety percent of charged particles; end-to-end alternatives such as HEPTV2 used pixel hits only and stayed at higher momentum. On 820 TrackML test events, evaluated inclusively over all reconstructable tracks, it reaches a TrackML score of 0.959, and 0.981 for tracks above 1 GeV.
Reconstruction runs in two levels: a global Transformer predicts discrete track parameters for every hit, and a local refiner resolves the remaining track-identity ambiguities. The global stage bins track parameters in inverse transverse momentum, azimuth and pseudorapidity, so hits sharing a token form a candidate track, but fragments in neighbouring bins and particles sharing a token still leave ambiguities. The 18-layer Tracker alone scores 0.844, charge-aware preconditioning raises this to 0.880, and the local Refiner applied to all candidate centres reaches 0.959; a learned Gate that refines only 56 percent of centres still holds 0.959.
AlphaHits trains nine task heads jointly on a tracking-pretrained backbone, emitting tracks, jets with flavour, missing transverse momentum, leptons and the event class in one pass from the same hit-level representation. The earlier HiggsFormer performed single-task classification from inner-tracker hits only; AlphaHits extends the same backbone to multi-task reconstruction and classification. On hard-scatter events without pileup, Higgs classification reaches an AUC of 0.932 from silicon hits alone and 0.944 with calorimeter towers; on the same events a Particle Transformer baseline on Delphes objects reaches 0.877 with tracks alone and 0.924 with tracks and calorimeter.
Perspective
The demonstrated scope is full charged-particle tracking on simulated TrackML events and multi-task reconstruction and classification on hard-scatter events without pileup; the multi-task model has not been evaluated on high-pileup events.
The multi-task AlphaHits is currently evaluated only on hard-scatter events without pileup, and extending joint reconstruction and classification to complete high-pileup events remains a next step; learning curves show the models are still limited by training-data statistics, and ablations of the auxiliary objectives and inference-latency optimisation are left for later work.