Physics Objects in CMS Run 3

In these proceedings we review the physics objects used by the CMS experiment during LHC Run 3 at 13.6 TeV, including charged leptons, photons, jets, and missing transverse momentum. Their performance and calibration is critical for physics analysis. In particular, the algorithms need to be resilient against the high pileup conditions in Run 3 collisions. Furthermore, transformer-based algorithms are deployed for the identification of heavy-flavor jets and boosted resonances.

💡 Research Summary

The paper provides a comprehensive overview of the reconstruction, calibration, and identification strategies employed by the CMS experiment during LHC Run 3 at a center‑of‑mass energy of 13.6 TeV. It begins by noting the unprecedented data set—over 500 fb⁻¹ accumulated across Runs 1‑3 and an average pile‑up of about 60 interactions per bunch crossing—which demands robust, pile‑up‑resistant algorithms to preserve sensitivity to rare processes and precision measurements.

The traditional Particle‑Flow (PF) reconstruction, which combines tracker, calorimeter, and muon‑system information to build a global event description, is augmented by a new Machine‑Learned PF (MLPF) algorithm. MLPF uses a transformer architecture to infer particle content directly from raw tracks and clusters, achieving comparable physics performance while delivering a factor‑two speed‑up on heterogeneous CPU‑GPU platforms.

For charged leptons, muon reconstruction links silicon tracks with muon‑system hits, and the single‑muon trigger threshold remains at 24 GeV. Isolation requirements were relaxed in 2024 to improve efficiency. Both cut‑based and multivariate muon ID are employed, with dedicated low‑pT and high‑pT algorithms. Tag‑and‑probe studies on J/ψ and Z→μμ resonances monitor efficiencies, and momentum scale corrections (multiplicative for magnetic field modeling, additive for tracker alignment) achieve a standard precision of ~0.05 %, while dedicated W‑mass analyses reach 0.006 % through custom track fits.

Electron reconstruction uses a Gaussian‑Sum‑Filter (GSF) combined with “mustache” super‑clusters to mitigate bremsstrahlung losses. Energy regression based on shower shape and pile‑up density further improves resolution. Efficiencies are derived from Z→ee events, and a novel unbinned multivariate scale‑factor method yields high‑dimensional efficiency maps. Photon identification relies on shower‑shape and isolation variables; an MVA‑based ID has been retuned for pile‑up stability, with energy scale calibrated on Z→ee and validated with Z→μμγ.

Jets are clustered from PF candidates using the anti‑kₜ algorithm with radii R = 0.4 and 0.8. The calibration chain includes L1 pile‑up offset subtraction, simulation‑based response corrections, and residual data‑driven corrections derived from dijet η‑balance and γ/Z+jet absolute pₜ studies. Run 2 jet energy scale (JES) uncertainties were at the 1 % level; Run 3 adopts the Pileup Per Particle Identification (PUPPI) algorithm as the default pile‑up mitigation technique, replacing Charged Hadron Subtraction (CHS). PUPPI rescales particle four‑momenta using local shape information, effectively eliminating the need for a large L1 offset and strongly suppressing pile‑up jets within the tracker acceptance. JES and JER calibrations have been derived for 2022‑2025 data, with early 2025 results showing improved end‑cap performance due to refined calorimeter calibrations.

Heavy‑flavour and boosted object tagging have transitioned from the DeepJet recurrent neural network (used in Run 2) to transformer‑based architectures. ParticleNet treats jet constituents as an unordered particle cloud, enabling superior exploitation of low‑level information. The Unified Particle Transformer (UParT) incorporates pairwise interaction features and adversarial training to achieve robustness against simulation mismodeling, and uniquely enables strange‑jet tagging and quark/gluon discrimination. Both models are commissioned in Run 3 and also perform pₜ regression as an auxiliary task, improving jet energy resolution by up to 20 % after MC corrections. Tau leptons are seeded with the hadrons‑plus‑strips algorithm and identified by DeepTau (v2.5), which reduces jet mis‑identification by 30‑50 % relative to the previous version. Boosted top tagging employs ParticleNet and ParT on large‑radius AK8 jets, delivering up to tenfold background rejection improvement over the mass‑decorrelated DeepAK8 tagger and better top‑mass resolution when combined with soft‑drop grooming.

Missing transverse momentum (pₜ^miss) reconstruction benefits from the switch to PUPPI. While PF‑based pₜ^miss without pile‑up subtraction was standard in Run 2, PUPPI‑based pₜ^miss is now recommended for Run 3 due to its reduced pile‑up dependence and improved resolution. The DeepMET deep neural network further refines pₜ^miss by learning optimal per‑candidate weights, yielding a 10‑30 % resolution gain.

On the Monte‑Carlo side, the standard tt̄ simulation for Run 2 and 3 uses POWHEG‑hvq interfaced with PYTHIA 8 (CP5 tune). Improvements include more accurate ISR/FSR modeling, top‑pₜ spectrum tuning, and multiple colour‑reconnection (CR) models tuned to CMS data, including a QCD‑inspired model that better describes ATLAS colour‑flow measurements. Future upgrades anticipate MiNNLO generators for NNLO+PS accuracy.

In summary, the paper demonstrates that CMS has successfully commissioned a suite of advanced reconstruction and identification tools—MLPF, PUPPI, transformer‑based heavy‑flavour and boosted object taggers, and DeepMET—that together maintain high efficiency and precision in the challenging high‑pile‑up environment of Run 3. Coupled with precise luminosity determination and refined MC modeling, these developments position CMS to continue delivering precision top‑quark physics and to explore new phenomena in the coming years.