This paper measures how well a new machine‑learning tagger, called GN2, finds jets that come from b‑hadrons (so‑called b‑jets) in data collected by the ATLAS detector at the Large Hadron Collider. The study uses proton–proton collisions recorded in 2022 and 2023 at a collision energy of 13.6 TeV and an integrated luminosity of 56.0 ± 1.1 fb⁻¹. The main output is a set of calibration factors that tell analysts how to correct simulation to match what is seen in real data.

The researchers selected events enriched in top‑quark pairs where both W bosons decay to leptons (the “dileptonic” tt̄ channel). That selection requires exactly two isolated leptons of different flavours and opposite charge, and two jets. These events are a rich source of b‑jets because each top quark almost always decays to a W boson and a b quark. The b‑jet efficiency was measured as a function of jet transverse momentum (pT) from 20 to 400 GeV. The team reported results in six ranges of the tagger’s cumulative efficiency, which were defined using simulated tt̄ events: [100%, 90%], [90%, 85%], [85%, 77%], [77%, 70%], [70%, 65%], and [65%, 0%]. Data were collected with single‑lepton triggers that required electrons (pT > 26 GeV) or muons (pT > 24 GeV).

GN2 is a transformer‑based machine‑learning model. In plain terms, a transformer is a modern neural network that can look at many inputs together and learn which parts matter most. GN2 predicts the probability that a jet is a b‑jet, a c‑jet (from charm hadrons), a τ‑jet (from hadronic τ decays), or a light‑flavour jet. These probabilities are combined into a single score, DGN2, using fixed coefficients (fc = 0.2 and fτ = 0.01) chosen to maximise background rejection in simulated tt̄ events. GN2 is also trained on related tasks such as predicting the position of decay vertices, which helps it learn faster and improves performance.