Pith. sign in

REVIEW 4 major objections 4 minor 38 references

Under standardized training with a controlled feature hierarchy, symmetry-constrained graph and transformer models (LorentzNet, Particle Transformer) beat flat tabular models for ttH multilepton signal–background classification at 14 TeV.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-01 16:18 UTC pith:TTMY56B2

load-bearing objection A careful benchmark that makes a credible case for symmetry-aware architectures on simulated ttH multilepton data — with a real caveat about the fake-lepton background. the 4 major comments →

arxiv 2607.18022 v1 pith:TTMY56B2 submitted 2026-07-20 hep-ph

Benchmarking Machine Learning Architectures for ttH Multilepton Signal Sensitivity

Luk\'a\v{s} Vicen\'ik , Andr\'e Sopczak , Oleksandr Shekhovtsov This is my paper
classification hep-ph
keywords ttH multileptonsignal-background classificationsymmetry-constrained architecturesgraph neural networksparticle transformerLorentzNetprofile likelihood sensitivityHL-LHC
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

Classifier comparisons in high-energy physics are often muddied by different features, training recipes, and metrics. This paper removes those confounds: six architectures, from boosted trees and multilayer perceptrons to graph networks and symmetry-aware transformers, are trained on the same ttH multilepton simulated events under identical hyperparameter optimization and a four-level feature hierarchy. The central result is that architectures with built-in physical symmetries win: Particle Transformer and LorentzNet reach the highest ROC AUC (0.799 each), LorentzNet yields the best signal-strength uncertainty (ϱμ = 0.784 vs 0.789 for ParT at 3000 fb−1), and even a cheap azimuthal-symmetry encoding (rotary position encoding in a standard message-passing network) consistently beats unconstrained baselines. The paper also shows ROC AUC tracks the analysis-relevant profile-likelihood metric under realistic weighting, that joint multi-channel training helps most in data-scarce channels, and releases the simulated dataset. If the conclusions hold, event-level searches should move toward symmetry-constrained graph/transformer models and shared multi-channel training.

Core claim

On the paper's own terms, the discovery is that architecture and inductive bias matter as much as feature engineering: under a standardized pipeline, LorentzNet (which enforces full Lorentz-group invariance by construction) achieves the best profile-likelihood sensitivity, σ(μ)/σ_ref = 0.784 at 3000 fb−1, and ties Particle Transformer for the best ROC AUC (0.799), while both clearly outrank XGBoost (0.771 AUC, 0.841 ϱμ), MLP, and the unconstrained MPNN. A controlled ablation isolates the driver: encoding only the detector's exact azimuthal symmetry (global rotation about the beam axis) into an otherwise standard message-passing network, via rotary position encoding, yields consistent gains a

What carries the argument

The load-bearing mechanism is a controlled hierarchy of input features combined with exact symmetry constraints. Features are added in four levels: raw four-momenta; per-object charge or b-tag score; detector coordinates (pT, η, ϕ, m); and global event quantities (MET, HT, multiplicities). Against this common ladder, the paper contrasts flat models, a permutation-invariant message-passing network, and three symmetry-constrained variants: MPNN+RoPE, which makes azimuthal rotation about the beam axis an exact symmetry by rotating attention query/key vectors by each particle's ϕ; Particle Transformer, which injects pairwise physics features (invariant masses, angular separations, charge product

Load-bearing premise

The simulation's description of non-prompt (fake) leptons is assumed to be appropriate, even though a full analysis would determine that rate with data-driven methods.

What would settle it

Re-run the six-architecture comparison after replacing the simulated non-prompt lepton background in the 2ℓSS1τ, 1ℓ2τ, and 2ℓ2τ channels with a data-driven estimate (or with a full detector simulation and data closure); if LorentzNet and Particle Transformer no longer lead, or the per-channel ranking changes, the superiority claim is an artifact of the fake-lepton simulation rather than a general property of the architectures.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Event-level searches and measurements at 14 TeV should treat symmetry-aware graph/transformer models, not boosted trees, as the default starting point; the expected gain is ΔAUC≈0.036 and Δϱμ≈0.067 between best and worst architecture.
  • ROC AUC can serve as a fast proxy for signal-strength sensitivity during model development, provided training uses physically realistic event weights, since the two rankings agree here.
  • Joint multi-channel training is preferable to dedicated per-channel training for almost every architecture and channel, so a single shared model across analysis categories is a viable path.
  • None of the architectures saturates with the available simulated data, so larger simulated samples should continue to improve all models, with symmetry-constrained models retaining their edge at low statistics.
  • The released simulated dataset (593,390 events across six channels) gives other analyses a common ground for fair architectural comparisons.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Because the benchmark fixes the pipeline so carefully, its protocol (feature hierarchy, standardized hyperparameter optimization, closed-form ϱμ reference) could be transferred to other final states; the authors do not make that claim explicitly, but the released dataset and configuration make it possible.
  • If the simulation's non-prompt lepton modeling is wrong in the tau channels, the per-channel rankings there may not transfer to real data; a data-driven closure test would be needed before relying on the per-channel conclusions.
  • A single-frequency rotary encoding appears to capture most of the symmetry benefit; this suggests existing message-passing analyses could gain by adding RoPE without adopting a full transformer or Lorentz-equivariant architecture, a cheap upgrade worth testing on other processes.
  • LorentzNet's advantage is largest when data are scarce, while Particle Transformer scales more steeply; with larger simulated samples, ParT may overtake LorentzNet, but the authors only observe the crossover, not prove it.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

4 major / 4 minor

Summary. The paper presents a controlled benchmark of six ML architectures (XGBoost, MLP, MPNN, MPNN+RoPE, ParT, LorentzNet) for ttH multilepton signal-background classification at sqrt(s)=14 TeV, using a newly generated 593,390-event Delphes fast-simulation dataset with ttH, ttW, and ttZ processes across six analysis channels. Standardized training, a four-level feature hierarchy, per-architecture HPO, five-seed statistics, and a profile-likelihood sensitivity metric rho_mu are used. The main findings are that symmetry-constrained graph/transformer models (ParT, LorentzNet) outperform flat and generic graph models on both ROC AUC and rho_mu, that RoPE gives consistent gains over the unconstrained MPNN, that joint multi-channel training helps particularly in data-scarce channels, and that performance has not saturated at the available sample size. The dataset and configuration files are planned for public release.

Significance. If the central claim holds, the paper provides a useful, reproducible benchmark suggesting that event-level HEP classification should shift from tabular GBDTs to symmetry-aware graph/transformer architectures. Concrete strengths include the public dataset release, the controlled feature hierarchy, the standardized HPO and train/validation/test protocol, five-seed statistical reporting, explicit azimuthal-rotation invariance tests, and robustness scans over luminosity and TransfoD binning. The closed-form reference uncertainty in Appendix D is also a nice contribution. These strengths make the benchmark a valuable reference point for the community, provided the modeling caveats below are adequately addressed or explicitly scoped.

major comments (4)
  1. [Section 3.2 / Table 1 / Section 8] The manuscript contains an internal inconsistency about the dominant background. Section 3.2 states that "non-prompt lepton contributions are not specified separately because they are included in the simulation," but Table 1 lists only ttH, ttW, and ttZ with no fake-lepton process. Section 8 then acknowledges that in the 2lSS1tau, 1l2tau, and 2l2tau channels the dominant background in a full analysis arises from non-prompt (fake) leptons, and that the simulation merely "assumes" these are described appropriately. Since the per-channel rankings in Figs. 8 and 9 and the joint-vs-dedicated conclusions for these channels are computed against a background mixture that a real ttH multilepton analysis does not face, the transferability of the central claim to data is not established. Please add a dedicated fake-enriched sample or, minimally, a robustness test varying the fake fraction and kinem
  2. [Section 8 / Appendix D.3, Eq. (D3)-(D6)] The primary metric rho_mu is evaluated using only the 10% held-out test split, with f_test=0.10 explicitly entering the per-event weights in Eq. (D3). This makes the Monte-Carlo statistical floor in Eq. (D6) ten times larger than if the full 593,390 selected events were used for the statistical model, as a real analysis would. The reported rho_mu values are therefore not the expected statistical uncertainties for the stated luminosity, and the ranking across architectures could depend on this inflated MC-stat term, especially in the signal-rich high-score bins. Please compute the profile-likelihood quantities from the full selected sample while keeping the test split only for classifier evaluation, or scan f_test and demonstrate ranking invariance.
  3. [Section 9.1 / Table C1 / Conclusion] The headline claim that "ParT and LorentzNet are essentially tied for the highest ROC AUC (0.799 for both)" compares ParT at the full feature level with LorentzNet at the object level, the highest level each architecture can use. This is not an architecture-only comparison. At the common object level, Table C1 gives LorentzNet 0.7991 versus ParT 0.7788, a much larger difference; for rho_mu the object-level gap is 0.7844 versus 0.8135. Please compare both at the same feature level or explicitly rephrase the abstract and conclusion as "best under each model's applicable feature set," avoiding the word "tied."
  4. [Section 9.1 / Appendix B.2] For the object feature level, all architectures use hyperparameters optimized on the core subset, not on the object features themselves. Since object is LorentzNet's peak configuration and is used in the headline comparison, this HPO reuse could disadvantage or advantage specific models at a load-bearing point. The paper should either re-run HPO on the object level for at least ParT and LorentzNet, or provide evidence that core-optimal hyperparameters are adequate for object-level features. Without this, the reported margin at object is less clean than the standardized-HPO framing suggests.
minor comments (4)
  1. [Throughout] There are several typographical artifacts: "T able 1", "T able A1", etc.; the symbol for the sensitivity ratio is inconsistently written as rho_mu, "rho(mu)", and "rho(mu)" across the text. Please normalize notation.
  2. [Figure 5] Figure 5 is extremely dense and difficult to parse, especially the architecture-specific annotations. Consider splitting it into a per-architecture panel or moving the detailed annotations to a table.
  3. [Section 3.2] The sentence "Non-prompt lepton contributions are not specified separately because they are included in the simulation" should be clarified: either state which generated process produces them, or explicitly say that no dedicated fake-lepton sample is generated and that this is a limitation.
  4. [Section 8] The statistical model fixes k_ttW = k_ttZ = 1 and includes no experimental systematics; this is fine for a pure architecture comparison, but the text should state more explicitly that rho_mu is an idealized statistical-only sensitivity, not the expected measurement uncertainty.

Circularity Check

0 steps flagged

No significant circularity: the paper is an empirical benchmark with held-out evaluation; no prediction reduces to a fitted input or self-citation.

full rationale

This paper is an empirical benchmarking study, not a derivation. The central claims—that ParT and LorentzNet outperform tabular baselines under controlled training and feature hierarchies—are supported by held-out test evaluations (ROC AUC and profile-likelihood ϱμ from an Asimov fit), with hyperparameters selected on validation AUC and the test split used only once. No fitted parameter is renamed as a prediction: LorentzNet's evaluation at the object feature level is an openly stated configuration choice ('LorentzNet is evaluated at object instead of full, as its architecture cannot accommodate the global event-level features introduced at the detector and full levels, and object is where it achieves peak performance'), not a circular derivation. The assumption about non-prompt leptons in Section 8 ('We assume that our simulation describes the non-prompt leptons appropriately, being aware that in a full analysis, the non-prompt rate is determined with a data-driven method') is a genuine limitation and an external-validity risk, but it is not a circularity: the benchmark compares classifiers on the same simulated sample, and the assumption does not define any architecture's score in terms of the claimed outcome. References to ATLAS [12] and to the architecture papers [7,8,26] are external and independent; there is no load-bearing self-citation chain. The claim that ROC AUC correlates with ϱμ ranking is an empirical observation on the benchmark, not an identity imposed by construction. Overall, the derivation chain is self-contained and no circular step can be exhibited.

Axiom & Free-Parameter Ledger

5 free parameters · 6 axioms · 0 invented entities

The benchmark's conclusions rest on simulation fidelity and statistical modeling choices, not on new physics entities or formal derivations. The main hand-tuned/fitted quantities are HPO hyperparameters and fixed evaluation settings; the paper reports robustness to some (luminosity, binning) but not to the simulation assumptions.

free parameters (5)
  • Per-architecture HPO hyperparameters = Optuna-selected (final values not reported; search spaces in Table B2)
    Determines final model ranking; validation ROC AUC is the HPO objective while the headline metric is ρμ.
  • TransfoD binning parameter a = a = z_b = z_s = 10
    Controls classifier-score binning in the profile-likelihood fit; robustness scan shows ranking unchanged, but absolute ρμ values depend on it.
  • RoPE frequency omega = omega = 1
    Chosen as the single natural scale for phi in [0,2pi); not optimized.
  • ParT pair-embedding dimensions = [64,64,64]
    Fixed to reduce HPO search space; affects ParT's pairwise bias capacity.
  • LorentzNet coordinate scale c = HPO-searched in [1e-4,1e-1], final not reported
    Controls the equivariant coordinate update; influences LorentzNet convergence and performance.
axioms (6)
  • domain assumption Delphes 3.5 fast simulation with the ATLAS card is a sufficient approximation of full ATLAS detector response for ranking classifiers.
    Used for all objects/events; Section 3.1, 3.3. If fast-sim response differs from full Geant4, per-channel and architecture rankings could shift.
  • domain assumption Non-prompt (fake) lepton backgrounds are adequately modeled by the simulation; no data-driven fake estimation is applied.
    Explicitly acknowledged in Section 8; dominant background in 2ℓSS1τ, 1ℓ2τ, 2ℓ2τ; a load-bearing simplification for those channels.
  • standard math Asymptotic profile-likelihood formulas (Wald regime) describe the expected signal-strength uncertainty at the tested luminosities.
    Used for σ(μ) and σ_ref (Section 8, Appendix D.3); standard in HEP but an approximation.
  • domain assumption The six analysis channels are statistically independent with no shared systematic uncertainties.
    Used to combine σ_ref via inverse-variance weighting (Eq. 15); real analyses include shared systematics, but here only MC-statistical nuisances are modeled.
  • domain assumption Validation ROC AUC is a sufficient proxy for the final ρμ objective during HPO.
    HPO uses validation ROC AUC (Section 7); paper shows correlation but not identity (Section 9), so optimal models for ρμ may be slightly different.
  • domain assumption Event weights from Eq. (2) recover the physical mixture without per-event kinematic reweighting.
    Assumes generation-level weights are constant per process; no NLO/PS acceptance corrections per event.

pith-pipeline@v1.3.0-alltime-deepseek · 25514 in / 11611 out tokens · 118716 ms · 2026-08-01T16:18:21.916785+00:00 · methodology

0 comments
read the original abstract

Statistical testing for signal discovery and signal-strength estimation in high-energy physics increasingly relies on machine-learning models trained on simulated data. We present a synthetic dataset for $t\bar t H$ multilepton signal--background classification and perform a systematic evaluation of machine-learning models ranging from the widely used XGBoost for tabular data to LorentzNet, which processes events with built-in Lorentz symmetry. Existing studies often differ in feature definitions, training procedures, and evaluation metrics, making it difficult to isolate the impact of model architecture on performance. To address this, we apply standardized training and hyperparameter-optimization procedures, construct a controlled hierarchy of feature sets, and perform a comprehensive comparison of the models. In addition to the analysis-driven metric of signal-strength uncertainty, we report the commonly used ROC AUC metric and show that under realistically weighted training, it correlates well with the uncertainty-based ranking of models. We further investigate model performance as a function of input feature set, training-set size, and the choice between channel-specific and unified multi-channel training. Our results establish the superiority of Particle Transformer and LorentzNet within the considered setup. We also identify potential avenues for further improvement, including more expressive architectures, joint training across additional analysis channels, and larger simulated datasets. Although this study focuses on a specific Higgs-boson analysis, we expect the main conclusions to generalize to a broader class of searches and measurements at the LHC and HL-LHC.

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

38 extracted references · 2 canonical work pages

  1. [1]

    Choudhury, A., Mondal, A., Sarkar, S.: Searches for the BSM scenarios at the LHC using decision tree based machine learning algorithms: a comparative study and review of Random Forest, AdaBoost, XGBoost and LightGBM frameworks. Eur. Phys. J. Spec. Top.233, 2571–2604 (2024) https://doi.org/10.1140/epjs/ s11734-024-01308-x arXiv:2405.06040 [hep-ph]

    Pith/arXiv arXiv 2024

  2. [2]

    Liu, J., Gurrola, A.: Probing freeze-in dark matter via a spin-2 portal at the lhc with vector boson fusion and machine learning (2026) arXiv:2604.02604 [hep-ph]

    Pith/arXiv arXiv 2026

  3. [3]

    Benbrik, R., et al.: Probing heavy neutral higgs bosons via single vector-like bottom quark production at the HL-LHC (2026) arXiv:2603.29076 [hep-ph]

    Pith/arXiv arXiv 2026

  4. [4]

    ATLAS Collaboration: Observation ofw +w−γproduction inppcollisions at √s= 13 TeV with the ATLAS detector and constraints on anomalous quartic gauge- boson couplings (2025) arXiv:2509.14070 [hep-ex]

    Pith/arXiv arXiv 2025

  5. [5]

    Bronstein, M.M., et al.: Geometric deep learning: Grids, groups, graphs, geodesics, and gauges (2021) arXiv:2104.13478 [cs.LG]

    Pith/arXiv arXiv 2021

  6. [6]

    Qu, H., Gouskos, L.: ParticleNet: Jet Tagging via Particle Clouds. Phys. Rev. D101(5), 056019 (2020) https://doi.org/10.1103/PhysRevD.101.056019 arXiv:1902.08570 [hep-ph]

    Pith/arXiv arXiv 2020

  7. [7]

    Qu, H., Li, C., Qian, S.: Particle Transformer for jet tagging. Proc. Int. Conf. Machine Learning (ICML)162, 18281–18292 (2022) arXiv:2202.03772 [hep-ph]

    Pith/arXiv arXiv 2022

  8. [8]

    JHEP07, 030 (2022) https://doi.org/10.1007/JHEP07(2022)030 arXiv:2201.08187 [hep-ph]

    Gong, S.,et al.: An efficient Lorentz equivariant graph neural network for jet tagging. JHEP07, 030 (2022) https://doi.org/10.1007/JHEP07(2022)030 arXiv:2201.08187 [hep-ph]

    Pith/arXiv arXiv 2022

  9. [9]

    Chen, T., Guestrin, C.: XGBoost: A scalable tree boosting system. Proc. ACM SIGKDD Int. Conf. Knowledge Discovery and Data Mining, 785–794 (2016) https: //doi.org/10.1145/2939672.2939785 arXiv:1603.02754 [cs.LG]

    arXiv 2016

  10. [10]

    In: Proc

    Gilmer, J.,et al.: Neural message passing for quantum chemistry. In: Proc. Int. Conf. Machine Learning (ICML), vol. 70, pp. 1263–1272 (2017)

    2017

  11. [11]

    (eds.): High-Luminosity Large Hadron Collider (HL-LHC): Technical Design Report

    B´ ejar Alonso, I., Br¨ uning, O., Fessia, P., Lamont, M., Rossi, L., Tavian, L., Zer- lauth, M. (eds.): High-Luminosity Large Hadron Collider (HL-LHC): Technical Design Report. CERN Yellow Reports: Monographs, vol. 10/2020. CERN, Geneva 25 (2020). https://doi.org/10.23731/CYRM-2020-0010

    doi:10.23731/cyrm-2020-0010 2020

  12. [12]

    JHEP05, 183 (2026) https://doi.org/10.1007/JHEP05(2026) 183 arXiv:2510.23755 [hep-ex]

    ATLAS Collaboration: Measurement of the Higgs boson production in association with top quarks in multilepton final states inppcollisions at √s= 13 TeV with the ATLAS detector. JHEP05, 183 (2026) https://doi.org/10.1007/JHEP05(2026) 183 arXiv:2510.23755 [hep-ex]

    Pith/arXiv arXiv 2026

  13. [13]

    ATLAS Collaboration: Observation of four-top-quark production in the mul- tilepton final state with the ATLAS detector. Eur. Phys. J. C83, 496 (2023) https://doi.org/10.1140/epjc/s10052-023-11573-0 arXiv:2303.15061 [hep-ex]

    Pith/arXiv arXiv 2023

  14. [14]

    Pfeffer, E.,et al.: A case study of sending graph neural networks back to the test bench for applications in high-energy particle physics. Comput. Softw. Big Sci.8, 13 (2024) https://doi.org/10.1007/s41781-024-00122-3 arXiv:2402.17386 [hep-ex]

    Pith/arXiv arXiv 2024

  15. [15]

    Onyisi, P., Shen, D., Thaler, J.: Comparing point cloud strategies for collider event classification. Phys. Rev. D108, 012001 (2023) https://doi.org/10.1103/ PhysRevD.108.012001 arXiv:2212.10659 [hep-ph]

    Pith/arXiv arXiv 2023

  16. [16]

    SciPost Phys.19, 028 (2025) https://doi.org/10.21468/SciPostPhys.19.1.028 arXiv:2211.05143 [hep-ph]

    Builtjes, L.,et al.: Attention to the strengths of physical interactions: Trans- former and graph-based event classification for particle physics experiments. SciPost Phys.19, 028 (2025) https://doi.org/10.21468/SciPostPhys.19.1.028 arXiv:2211.05143 [hep-ph]

    Pith/arXiv arXiv 2025

  17. [17]

    SciPost Phys.12, 043 (2022) https://doi.org/10.21468/SciPostPhys.12.1

    Aarrestad, T.,et al.: The Dark Machines Anomaly Score Challenge: Benchmark Data and Model Independent Event Classification for the Large Hadron Col- lider. SciPost Phys.12, 043 (2022) https://doi.org/10.21468/SciPostPhys.12.1. 043 arXiv:2105.14027 [hep-ph]

    Pith/arXiv arXiv 2022

  18. [18]

    JHEP07, 079 (2014) https://doi.org/10.1007/JHEP07(2014)079 arXiv:1405.0301 [hep-ph]

    Alwall, J.,et al.: The automated computation of tree-level and next-to-leading order differential cross sections, and their matching to parton showers. JHEP07, 079 (2014) https://doi.org/10.1007/JHEP07(2014)079 arXiv:1405.0301 [hep-ph]

    Pith/arXiv arXiv 2014

  19. [19]

    NNPDF Collaboration: Parton distributions from high-precision collider data. Eur. Phys. J. C77, 663 (2017) https://doi.org/10.1140/epjc/s10052-017-5199-5 arXiv:1706.00428

    Pith/arXiv arXiv 2017

  20. [20]

    Sj¨ ostrand, T.,et al.: An introduction toPythia8.2. Comput. Phys. Commun. 191, 159–177 (2015) https://doi.org/10.1016/j.cpc.2015.01.024 arXiv:1410.3012 [hep-ph]

    Pith/arXiv arXiv 2015

  21. [21]

    JHEP02, 057 (2014) https://doi.org/10.1007/ JHEP02(2014)057 arXiv:1307.6346 [hep-ex] 26

    Favereau, J.,et al.:Delphes3: a modular framework for fast simulation of a generic collider experiment. JHEP02, 057 (2014) https://doi.org/10.1007/ JHEP02(2014)057 arXiv:1307.6346 [hep-ex] 26

    Pith/arXiv arXiv 2014

  22. [22]

    Deciphering the Nature of the Higgs Sector (2016) https://doi.org/10

    LHC Higgs Cross Section Working Group: Handbook of LHC Higgs Cross Sec- tions: 4. Deciphering the Nature of the Higgs Sector (2016) https://doi.org/10. 23731/CYRM-2017-002 arXiv:1610.07922

    Pith/arXiv arXiv 2016

  23. [23]

    JHEP 04, 063 (2008) https://doi.org/10.1088/1126-6708/2008/04/063 arXiv:0802.1189 [hep-ph]

    Cacciari, M., Salam, G.P., Soyez, G.: The anti-k T jet clustering algorithm. JHEP 04, 063 (2008) https://doi.org/10.1088/1126-6708/2008/04/063 arXiv:0802.1189 [hep-ph]

    Pith/arXiv arXiv 2008

  24. [24]

    ATLAS Public Note ATL-PHYS-PUB-2015-022, CERN (2015)

    ATLAS Collaboration: Expected performance of the ATLASb-tagging algorithms in Run-2. ATLAS Public Note ATL-PHYS-PUB-2015-022, CERN (2015). https: //cds.cern.ch/record/2037697

    arXiv 2015

  25. [25]

    ATLAS Public Note ATL-PHYS-PUB-2015-045, CERN (2015)

    ATLAS Collaboration: Tau reconstruction and identification efficiency meas- urements usingW→τ νandZ→τ τdecays with the ATLAS detector. ATLAS Public Note ATL-PHYS-PUB-2015-045, CERN (2015). https://cds.cern. ch/record/2064383

    arXiv 2015

  26. [26]

    Neurocomputing568, 127063 (2024) https://doi.org/10.1016/j.neucom

    Su, J.,et al.: RoFormer: Enhanced transformer with rotary position embed- ding. Neurocomputing568, 127063 (2024) https://doi.org/10.1016/j.neucom. 2023.127063 arXiv:2104.09864 [cs.CL]

    arXiv 2024

  27. [27]

    In: Proc

    Zhai, X.,et al.: Scaling vision transformers. In: Proc. IEEE/CVF Conf. Computer Vision and Pattern Recognition (CVPR), pp. 12104–12113 (2022)

    2022

  28. [28]

    In: Proc

    Akiba, T.,et al.: Optuna: A next-generation hyperparameter optimization frame- work. In: Proc. ACM SIGKDD Int. Conf. Knowledge Discovery and Data Mining, pp. 2623–2631 (2019). https://doi.org/10.1145/3292500.3330701

    arXiv 2019

  29. [29]

    In: Proc

    Kipf, T.N., Welling, M.: Semi-supervised classification with graph convolutional networks. In: Proc. Int. Conf. Learning Representations (ICLR) (2017)

    2017

  30. [30]

    In: Proc

    Veliˇ ckovi´ c, P.,et al.: Graph attention networks. In: Proc. Int. Conf. Learning Representations (ICLR) (2018)

    2018

  31. [31]

    In: Advances in Neural Information Processing Systems (NeurIPS), vol

    Hamilton, W.L., Ying, R., Leskovec, J.: Inductive representation learning on large graphs. In: Advances in Neural Information Processing Systems (NeurIPS), vol. 30 (2017)

    2017

  32. [32]

    In: Proc

    Shi, Y.,et al.: Masked label prediction: Unified message passing model for semi- supervised classification. In: Proc. Int. Joint Conf. Artificial Intelligence (IJCAI), pp. 1548–1554 (2021)

    2021

  33. [33]

    Cowan, G.,et al.: Asymptotic formulae for likelihood-based tests of new physics. Eur. Phys. J. C71, 1554 (2011) https://doi.org/10.1140/epjc/s10052-011-1554-0 arXiv:1007.1727 [physics.data-an]. [Erratum: Eur.Phys.J.C 73, 2501 (2013)]

    Pith/arXiv arXiv 2011

  34. [34]

    CERN GitLab, https://gitlab.cern.ch/TRExStats/TRExFitter (2024)

    ATLAS Collaboration: TRExFitter: A framework to perform profile likelihood 27 fits. CERN GitLab, https://gitlab.cern.ch/TRExStats/TRExFitter (2024). https: //doi.org/10.5281/zenodo.15524537

    work page doi:10.5281/zenodo.15524537 2024

  35. [35]

    Com- put

    Barlow, R.J., Beeston, C.: Fitting using finite Monte Carlo samples. Com- put. Phys. Commun.77, 219–228 (1993) https://doi.org/10.1016/0010-4655(93) 90005-W

    work page doi:10.1016/0010-4655(93 1993

  36. [36]

    In: PHYSTAT 2011, pp

    Conway, J.S.: Incorporating Nuisance Parameters in Likelihoods for Multi- source Spectra. In: PHYSTAT 2011, pp. 115–120 (2011). https://doi.org/10.5170/ CERN-2011-006.115

    2011

  37. [37]

    Technical Report CERN-OPEN-2012-016, CERN (2012)

    Cranmer, K., et al.: HistFactory: A tool for creating statistical models for use with RooFit and RooStats. Technical Report CERN-OPEN-2012-016, CERN (2012)

    2012

  38. [38]

    Pair bias

    Grinsztajn, L., Oyallon, E., Varoquaux, G.: Why do tree-based models still outperform deep learning on typical tabular data? In: Advances in Neural In- formation Processing Systems (NeurIPS), Datasets and Benchmarks Track, vol. 35 (2022) Appendix A Dataset and features A.1 Per-channel event selection Thet ¯tHmultilepton analysis is split into six orthogon...

    arXiv 2022