Pith. sign in

REVIEW 4 major objections 4 minor 83 references

Tagging fully hadronic exotic decays of the vectorlike $\mathbf{B}$ quark using a graph neural network

T0 review · 4 major / 4 minor · reviewed 2026-08-15 · deepseek-v4-flash

Pith's one-line read The paper projects that a graph-neural-network search can reach 1.8 TeV discovery and 2.4 TeV exclusion for fully hadronic vectorlike B decays at the HL-LHC.

desk verdict A credible, well-executed GNN-based collider study whose headline reach depends on an unquantified assertion that the QCD multijet background is negligible. read the letter →

arxiv 2505.07769 v2 pith:OJJURPIL submitted 2025-05-12 hep-ph cs.LGhep-ex

classification hep-phcs.LGhep-ex
keywords vectorlikeBquarkexoticdecayfullyhadronicfinalstategraphneuralnetworksingletscalarHL-LHCreachbottom-quarktaggingLHCsearchstrategy
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper argues that the fully hadronic decay chain $pp \to B\bar{B} \to (b\Phi)(\bar{b}\Phi)$, with the singlet scalar or pseudoscalar $\Phi$ decaying to $gg$ or $b\bar{b}$, can be brought within reach of the HL-LHC even though the final state has no leptons to trigger on. The authors build a hybrid classifier that represents each collision event as a graph of jets, fatjets, and event-level features, passes it through a graph neural network, and finishes with a deep neural network. With this pipeline and 3000 fb$^{-1}$ of data, they project that a search in the $2b+4j/6b$ final state could exclude vectorlike $B$ masses up to about 2.4 TeV and discover them near 1.8 TeV when $B\to b\Phi$ saturates the branching ratio. If correct, this would make a fully hadronic search competitive with semileptonic searches for the same process.

What carries the argument

The central machinery is an event graph with heterogeneous node types: shared-attribute nodes carry the four-momenta of every reconstructed jet and fatjet, auxiliary jet and fatjet nodes carry substructure and b-tag information, a global node carries event-level variables, and a CLS token aggregates the embedding for classification. The graph is built sequentially, first connecting kinematic nodes in a clique, then linking each object to its attributes, then connecting nearby objects with a $k$-nearest-neighbour rule in the $\eta$-$\phi$ plane, and finally connecting everything to the CLS token. Attention-based graph-convolution message passing updates the embeddings, and a five-layer deep neural network performs the final signal-versus-background classification after a bias-adjusted loss is used in pre-training on all mass points and fine-tuning on each point.

What would settle it

Generate a large QCD multijet sample through the same detector simulation and selection C1–C5 with realistic b-tagging, apply the trained GNN, and add the survivors to $N_B$ in Eq. (16); if the resulting $5\sigma$ discovery contour drops below roughly 1.5 TeV for $BR(B\to b\Phi)=100\%$, the paper's central reach claim is contradicted.

Watch

Extended reading notes

Core claim

The paper's central claim is that a hybrid deep-learning classifier, a graph neural network that builds an event-level embedding from every reconstructed jet, fatjet, and global event feature followed by a deep neural network, can separate the fully hadronic $pp\to B\bar B\to (b\Phi)(\bar b\Phi)$ signal, with $\Phi\to gg/b\bar b$, from Standard Model backgrounds. Applied to 14 TeV proton collisions at the HL-LHC luminosity of 3000 fb$^{-1}$, the analysis projects a $5\sigma$ discovery reach around $M_B \simeq 1.8$ TeV and a $2\sigma$ exclusion reach up to about 2.4 TeV when $BR(B\to b\Phi)=100\%$, with the exclusion still reaching about 1.8 TeV at 40% branching. The same reach applies to singlet $B+\Phi$ and doublet $(T,B)+\Phi$ models, and it makes the fully hadronic search competitive with the semileptonic search that previously set the benchmark for this process.

Load-bearing premise

The load-bearing premise is that ordinary QCD multi-jet events essentially vanish after the selection cuts and b-tagging requirements, so they are left out of both the background count and the training data; if even a small fraction survive, the projected discovery and exclusion masses are too high.

Editorial extensions

If this is right

  • At 3000 fb$^{-1}$, masses up to about 2.4 TeV could be excluded for $BR(B\to b\Phi)\simeq 100\%$, with discovery sensitivity around 1.8 TeV; even at 40% branching the exclusion still reaches about 1.8 TeV.
  • The fully hadronic $2b+4j/6b$ channel becomes competitive with the semileptonic $b\Phi$ search, which had previously set the benchmark for this decay mode.
  • The same exclusion contours apply to both singlet $B+\Phi$ and doublet $(T,B)+\Phi$ models, extending beyond current recast LHC limits and beyond the monoleptonic reach of the authors' earlier study.
  • Because the signal yield scales as $BR(B\to b\Phi)^2$, the plotted contours can be rescaled to estimate sensitivity for any intermediate branching ratio or for additional $B$ decay modes.
  • Section V further states that, for part of the parameter plane, a $5\sigma$ discovery significance can be reached at $M_B$ above 2 TeV.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The paper's zero-multijet assumption can be tested before the LHC run: a data-driven sideband that measures how often the selection cuts pass ordinary QCD events would directly bound how much of the 2.4 TeV exclusion is real.
  • The pretrain-then-finetune recipe suggests the event-graph encoder could be reused as a single pretrained backbone for several hadronic resonance searches, with only the final classifier retrained for a given mass point.
  • Because single-$B$ production can become competitive with pair production above about 2 TeV, including single-$B$ and $pp\to B\Phi$ contributions could extend the reach beyond the pair-production-only contours shown here.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 4 minor

Summary. This paper proposes a search strategy for pair-produced vectorlike B quarks decaying through the exotic mode B -> b Phi with fully hadronic final states (2b + 4j or 6b), where Phi is a new gauge-singlet scalar or pseudoscalar decaying mainly to gg or bb. The authors generate signal and Standard Model background events at sqrt(s) = 14 TeV with MadGraph/Pythia/Delphes, apply a five-stage cut flow (C1-C5), and then train a graph neural network followed by a deep neural network to separate signal from background. They quote projected HL-LHC discovery and exclusion reaches on the (M_B, M_Phi) plane for branching fractions beta_bPhi = 0.4, 0.7, 1.0, with the headline claim that at L = 3000 fb^-1, MB up to about 2.4 TeV can be excluded and a discovery reach around 1.8 TeV can be achieved for BR(B -> bPhi) = 100%. The paper also translates the exclusion contours into the BR(B -> bPhi) versus MB plane for singlet and doublet B + Phi models, comparing with current LHC limits.

Significance. If the projected reach is reliable, the paper would be a useful contribution: it demonstrates that a dedicated fully hadronic search with a GNN classifier can compete with semileptonic searches for exotic vectorlike B decays, and it provides a concrete analysis pipeline, including public model files. The event-generation setup, cut definitions, and GNN architecture are described in enough detail to be reproduced, and the classifier training strategy is clearly explained. However, the central reach claims rest on two incompletely justified assumptions: that the QCD multijet background is negligible without a quantitative estimate, and that systematic uncertainties can be ignored in the significance formulas. These gaps affect the headline numbers and need to be addressed before the projected reach can be taken at face value.

major comments (4)
  1. [III A / Table II] The QCD multijet background is asserted to be 'essentially eliminated' by the strong pT cuts and the b-tagging requirement, but no QCD multijet sample is generated, no cross-section is quoted, and Table II contains no multijet row. The inclusive multijet cross-section is orders of magnitude larger than the processes listed, and neither real b-jets from gluon splitting nor mistagged light jets are excluded at the rates required to make this statement quantitative. Because the GNN was trained without any QCD multijet class, its rejection power for this background is unknown. Please provide a quantitative estimate from a generated QCD multijet sample passing C1-C5 (and, ideally, the GNN classifier), or a data-driven sideband estimate, and include the result in the background yield used in Eq. (16).
  2. [V, Eqs. (16)-(17)] The discovery and exclusion significances are computed from the Poisson counting formulas of Eqs. (16) and (17), which contain no systematic uncertainties. With the total background after C5 being 1.58e6 events and the surviving signal being a few hundred events at the exclusion edge, even a few percent uncertainty on the background normalization or on the b-tagging efficiency can shift the 5-sigma and 2-sigma contours in Fig. 9 substantially. The projected reach should be recomputed with nuisance parameters (for example, log-normal background uncertainties) or the authors should demonstrate explicitly that such systematics are negligible for this analysis.
  3. [Abstract / V] The abstract states a discovery reach of about M_B = 1.8 TeV for BR(B -> bPhi) = 100%, while Section V states that 'it is possible to attain a discovery significance score of 5 sigma at values M_B > 2 TeV.' These two statements are inconsistent, and the exact mass value corresponding to the beta_bPhi = 1.0 contour in Fig. 9(a) should be identified. Please reconcile the abstract with the body of the paper.
  4. [IV C] The text in Section IV C says that the threshold on the classifier response is scanned and the value maximizing the discovery sensitivity in Eq. (16) is selected, but it does not specify whether this optimization is performed on a validation set that is independent from the events used to compute the quoted N_S and N_B. If the same events are used both to choose the threshold and to evaluate the significance, the projected Z_D is biased upward. Please clarify the train/validation/test split and, if a separate validation set was used, state the resulting threshold selection procedure explicitly.
minor comments (4)
  1. [Table II] The background labels such as 'thth', 'ththH', 'ththWh', and 'ththZh' are not defined in the text or captions; please clarify which SM processes these abbreviations denote and how they map to the processes listed in Table I.
  2. [III A / Fig. 5] The selection efficiency in Fig. 5 is quoted without statistical uncertainties; adding them would help assess whether the small differences across the parameter grid are meaningful.
  3. [II A, Eq. (2)] The notation for the Higgs doublet H and its vev is standard, but the sign conventions in Eqs. (2)-(4) should be stated more explicitly so that the mixing angles and the signs of the off-diagonal couplings are unambiguous.
  4. [V / Fig. 10] The comparison with the previous semileptonic result in Ref. [27] is stated only in the text; showing the corresponding exclusion curve in Fig. 10 would make the claimed improvement easier to verify.

Circularity Check

1 steps flagged · score 6.0 of 10

The projected HL-LHC reach is an in-sample optimum: the DNN is finetuned per mass point and the classifier threshold is chosen to maximize the significance on the same events used to compute that significance.

  1. fitted input called prediction [Section IV C (Finetuning), used in Section V Eqs. (16)-(17) and Figs. 8-10]
    "We scan the threshold values of the classifier response curve and select the best value that maximises the discovery sensitivity [66], given as ZD = sqrt(2(NS+NB) ln((NS+NB)/NB) - 2NS), (16) where NS and NB are the number of surviving signal and background events at the HL-LHC, respectively."

    The same MC events are used to fine-tune the DNN head for each mass point and to scan the classifier threshold; no held-out validation or test set is described before Eq. (16) is evaluated. The reported ZD is therefore not an independent prediction but the maximum of the fitting objective over the threshold and over the fine-tuned DNN weights, evaluated on the same events that entered the fit: ZD_reported = max_threshold ZD(threshold, model_fit_to_same_events). Since the 5-sigma discovery and 2-sigma exclusion contours in Figs.

full rationale

Apart from the threshold and fine-tuning issue, the paper's derivation is largely self-contained: the reach is computed from MadGraph/Pythia/Delphes Monte Carlo samples with external higher-order cross sections, the GNN is a standard architecture, and the model constraints from the authors' prior papers enter as inputs rather than as the output being derived. The omission of the QCD multijet background (Section III A and Table II) is a completeness and validation concern, not a circularity, because it is an asserted assumption rather than an equation that reduces to a fitted parameter. The comparison to the semileptonic reach uses the authors' own previous projection [27], which is self-citation, but the new reach is computed independently and that comparison is not used to derive the new numbers. The one genuine circular step is that the classifier and threshold are optimized on the same MC events that are then inserted into Eqs. (16)-(17) to quote ZD/ZE; without a described held-out set, the reported contours and M_B reach are the in-sample maximum of the fitting objective, so the 'prediction' is statistically forced by the optimization procedure.

Assumptions & free parameters 4 free parameters · 5 assumptions · 1 invented entities

The reported reach rests on the assumed BSM spectrum, a specific Phi decay pattern, an omitted QCD multijet contribution, fast detector simulation, and a threshold selection that fits the quoted significance. These are the main entries the reader is paying for rather than receiving as derived content.

free parameters (4)
  • ML hyperparameters (GATv2Conv layers, hidden sizes, dropout, AdamW weight decay, k for kNN, learning rate, batch size… = Final values are not reported; a Bayesian search was run.
    These choices determine the classifier efficiency and hence the projected reach, so they are fitted inputs rather than predictions.
  • Classification threshold per mass point = Values are not reported.
    The threshold is chosen by scanning the response curve and maximizing Z_D on the evaluated samples, which fits the quoted significance to the data.
  • Branching fraction beta_bPhi = 1.0 for the headline reaches; contours also at 0.4 and 0.7.
    The signal yield scales as beta_bPhi squared, and the 2.4 TeV exclusion claim assumes beta_bPhi equals 1.0.
  • Benchmark model couplings (lambda_a_Phi, lambda_b_Phi, mu_B1) = Examples include 0.11, 0.24, and 8.81 GeV; other benchmarks appear in Figure 2.
    These hand-selected benchmark points produce B to bPhi dominance and Phi to gg or bb dominance, so they are inputs rather than derived quantities.
assumptions (5)
  • domain assumption The vectorlike B quark and the singlet scalar or pseudoscalar Phi exist, with couplings generated by electroweak mixing as in Eqs. (2) to (6).
    This is the beyond-Standard-Model hypothesis under study, and no direct experimental evidence is presented in this paper.
  • domain assumption Phi decays predominantly to gg or bb in the parameter region considered.
    The 2b+4j/6b signature and the two-prong fatjet tag require this decay pattern, as established in Section II and Figure 2.
  • domain assumption The QCD multijet background is negligible after cuts C1 to C5 and can be excluded from Table I and from training.
    This is stated in Section III A with no quantitative simulation or data-driven estimate, yet it is load-bearing for the background model.
  • domain assumption Delphes 3 with the DeepCSV medium working point adequately models detector response for b-tagging and fatjet observables.
    All signal and background counts pass through this fast simulation, and no Geant4-level validation is provided.
  • domain assumption The asymptotic significance formulas (16) and (17) are valid without systematic uncertainty terms.
    No systematic uncertainties on background rates or taggers are included in the reported Z_D and Z_E values.
invented entities (1)
  • Singlet scalar or pseudoscalar Phi, inherited from the authors' earlier model series
    purpose: Mediates the exotic decay B to b Phi and then decays to gg or bb, producing the jets that define the search signature.
    The Phi is carried over from Refs. [25] to [27]; this paper provides no independent detection channel or direct evidence for it, so the reach is conditional on its existence.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Tagging fully hadronic exotic decays of the vectorlike $\mathbf{B}$ quark using a graph neural network." pith.science (2026). https://pith.science/paper/OJJURPIL

@misc{pith2026250507769,
  author       = {Pith},
  title        = {Pith review of: Tagging fully hadronic exotic decays of the vectorlike $\mathbfB$ quark using a graph neural network},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/OJJURPIL}},
  note         = {Machine review of arXiv:2505.07769}
}
abstract

Following up on our earlier study in [J. Bardhan et al., Machine learning-enhanced search for a vectorlike singlet B quark decaying to a singlet scalar or pseudoscalar, Phys. Rev. D 107 (2023) 115001; arXiv:2212.02442], we investigate the LHC prospects of pair-produced vectorlike $B$ quarks decaying exotically to a new gauge-singlet (pseudo)scalar field $\Phi$ and a $b$ quark. After the electroweak symmetry breaking, the $\Phi$ decays predominantly to $gg/bb$ final states, leading to a fully hadronic $2b+4j$ or $6b$ signature. Because of the large Standard Model background and the lack of leptonic handles, it is a difficult channel to probe. To overcome the challenge, we employ a hybrid deep learning model containing a graph neural network followed by a deep neural network. We estimate that such a state-of-the-art deep learning analysis pipeline can lead to a performance comparable to that in the semi-leptonic mode, taking the discovery (exclusion) reach up to about $M_B=1.8\:(2.4)$ TeV at HL-LHC when $B$ decays fully exotically, i.e., BR$(B \to b\Phi) = 100\%$.

Figures

Figures reproduced from arXiv: 2505.07769 by the authors.

Figure 1
Figure 1. FIG. 1. Signal topology [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. FIG. 2. Decays of [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. FIG. 3. Parameter scans for VLQ [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (6 more)
Figure 5
Figure 5. Figure 5: FIG. 5. Selection efficiency after [PITH_FULL_IMAGE:figures/full_fig_p005_5.png]
Figure 6
Figure 6. Figure 6: FIG. 6 [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]
Figure 7
Figure 7. Figure 7: FIG. 7. Schematic of the event classification pipeline. We perform an event-level signal vs. background classification on the learned [PITH_FULL_IMAGE:figures/full_fig_p008_7.png]
Figure 8
Figure 8. Figure 8: FIG. 8. Discovery sensitivity [defined in Eq. ( [PITH_FULL_IMAGE:figures/full_fig_p009_8.png]
Figure 9
Figure 9. Figure 9: FIG. 9. LHC reach plots for [PITH_FULL_IMAGE:figures/full_fig_p010_9.png]
Figure 10
Figure 10. Figure 10: FIG. 10. Projected LHC reach for (a) the Singlet [PITH_FULL_IMAGE:figures/full_fig_p010_10.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

83 extracted references · 13 canonical work pages

  1. [1]

    We then perform a round of message passing

    GNN layer 1-2: In the first step, the SA nodes are connectedwitheachother(i.e.,acliquegraph)toen- able global interactions among the kinematics of the reconstructed objects and to ensure that each subse- quent connection has some information about other nodes present at the global scale. We then perform a round of message passing

  2. [2]

    We identify b1 as the physical bottom quark andb2 is mostly theB quark

    The matrix is diagonalised by a bi-orthogonal rotation with two mixing anglesθL andθR: bL/R BL/R = cL/R sL/R −sL/R cL/R b1L/R b2L/R , (4) where sP = sinθP and cP = cosθP for the two chirality projections, andb1 and b2 are the mass eigenstates. We identify b1 as the physical bottom quark andb2 is mostly theB quark. Hence, we use the notationsB andb2 inter-...

  3. [3]

    The strongest limit on Φ comes from the clean two-photon search data

    recast the heavy resonance searches to put constraints on Φ production at hadron colliders. The strongest limit on Φ comes from the clean two-photon search data. (See Refs. [25, 27] for a detailed discussion on the constraints on the VLQ+Φ models from the LHC data.) Fig. 3 shows different projections of the allowed parameter region af- ter these constrain...

  4. [4]

    GNN layer 2-3: Each SA node is connected to its cor- responding auxiliary jet or fatjet node in the second step. Thisallowsforobject-specificembeddinglearn- ing, where the kinematic information gathered from the previous layer and the node-level input from the current one are used to construct meaningful em- beddings for each reconstructed object. A round...

  5. [5]

    For instance, a fatjet will be close to its constituent jets

    GNN layer 3-5: Jets and fatjets located nearby are likely related. For instance, a fatjet will be close to its constituent jets. Furthermore, independent jets with a common parent are also likely to be close. To incorporate this, we connect nodes to their nearest objects in theη−ϕ plane in the next step using thek Nearest Neighbours (kNN) algorithm. We al...

  6. [6]

    This idea is commonly used in Transformers in language-related tasks [54]

    GNN layer 5-6: Once all message passings are per- formed, and each node has an appropriate embed- ding, weconnectallthenodestothe CLStoken/node to extract the graph-level class information and per- form a round of message passing. This idea is commonly used in Transformers in language-related tasks [54]. Node-type Input Features Shared attributes (SA) nod...

  7. [7]

    Pre-training: Wepre-trainourmodelonallsignalpa- rameter points we consider

  8. [8]

    Ourstrategyismainlymotivatedbythemountingevidence that large models trained on vast amounts of data gener- alise better than specific models [62–65]

    Finetuning: After pre-training the classifier, we fine- tune our model on each mass point. Ourstrategyismainlymotivatedbythemountingevidence that large models trained on vast amounts of data gener- alise better than specific models [62–65]. Models trained onlargerdatasetsarebetterabletocapturetheunderlying properties. Our signal is characterised byMB andM...

Show all 83 references
  1. [9]

    LHC Signatures of a Vector-like b’,

    Shrihari Gopalakrishna, Tanumoy Mandal, Subhadip Mitra, and Rakesh Tibrewala, “LHC Signatures of a Vector-like b’,” Phys. Rev. D84, 055001 (2011), arXiv:1107.4306 [hep- ph]

  2. [10]

    LHC Signatures of Warped-space VectorlikeQuarks,

    Shrihari Gopalakrishna, Tanumoy Mandal, Subhadip Mitra, and Grégory Moreau, “LHC Signatures of Warped-space VectorlikeQuarks,”JHEP 08,079(2014),arXiv:1306.2656 [hep-ph]

  3. [11]

    Vector-like singlet quarks: A roadmap,

    João M. Alves, G. C. Branco, A. L. Cherchiglia, C. C. Nishi, J. T. Penedo, Pedro M. F. Pereira, M. N. Rebelo, and J. I. Silva-Marcos, “Vector-like singlet quarks: A roadmap,” Phys. Rept. 1057, 1–69 (2024), arXiv:2304.10561 [hep- ph]

  4. [12]

    Review of searches for vector-like quarks, vector-like leptons, and heavy neutral leptons in proton–proton collisions at√s = 13 TeV at the CMS experiment,

    Aram Hayrapetyan et al. (CMS), “Review of searches for vector-like quarks, vector-like leptons, and heavy neutral leptons in proton–proton collisions at√s = 13 TeV at the CMS experiment,” Phys. Rept.1115, 570–677 (2025), arXiv:2405.17605 [hep-ex]

  5. [13]

    Beyond the Minimal Top Partner Decay,

    Javi Serra, “Beyond the Minimal Top Partner Decay,” JHEP 09, 176 (2015), arXiv:1506.05110 [hep-ph]

  6. [14]

    Phenomenology of 2HDM with vectorlike quarks,

    A. Arhrib, R. Benbrik, S. J. D. King, B. Manaut, S. Moretti, and C. S. Un, “Phenomenology of 2HDM with vectorlike quarks,” Phys. Rev. D97, 095015 (2018), arXiv:1607.08517 [hep-ph]

  7. [15]

    Six Top Messages of New Physics at the LHC,

    Huayong Han, Li Huang, Teng Ma, Jing Shu, Tim M. P. Tait, and Yongcheng Wu, “Six Top Messages of New Physics at the LHC,” JHEP10, 008 (2019), arXiv:1812.11286 [hep- ph]

  8. [16]

    Common exotic decays of top partners,

    Nicolas Bizot, Giacomo Cacciapaglia, and Thomas Flacke, “Common exotic decays of top partners,” JHEP 06, 065 (2018), arXiv:1803.00021 [hep-ph]

  9. [17]

    Signaturesofvector-liketoppartners 11 decaying into new neutral scalar or pseudoscalar bosons,

    RachidBenbrik et al.,“Signaturesofvector-liketoppartners 11 decaying into new neutral scalar or pseudoscalar bosons,” JHEP 05, 028 (2020), arXiv:1907.05929 [hep-ph]

  10. [18]

    Light scalars in composite Higgs mod- els,

    Giacomo Cacciapaglia, Gabriele Ferretti, Thomas Flacke, and Hugo Serôdio, “Light scalars in composite Higgs mod- els,” Front. in Phys.7, 22 (2019), arXiv:1902.06890 [hep- ph]

  11. [19]

    Exotic vectorlike quark phenomenology in the minimal linearσ model,

    J. A. Aguilar-Saavedra, J. Alonso-González, L. Merlo, and J. M. No, “Exotic vectorlike quark phenomenology in the minimal linearσ model,” Phys. Rev. D101, 035015 (2020), arXiv:1911.10202 [hep-ph]

  12. [20]

    Exotic decays of top partners: mind the search gap,

    Giacomo Cacciapaglia, Thomas Flacke, Myeonghun Park, and Mengchao Zhang, “Exotic decays of top partners: mind the search gap,” Phys. Lett. B798, 135015 (2019), arXiv:1908.07524 [hep-ph]

  13. [21]

    Huntingfor top partner with a new signature at the LHC,

    DaohanWang, LeiWu, andMengchaoZhang,“Huntingfor top partner with a new signature at the LHC,” Phys. Rev. D 103, 115017 (2021), arXiv:2007.09722 [hep-ph]

  14. [22]

    Phenomenological aspects of compos- ite Higgs scenarios: exotic scalars and vector-like quarks,

    Avik Banerjeeet al., “Phenomenological aspects of compos- ite Higgs scenarios: exotic scalars and vector-like quarks,” (2022), arXiv:2203.07270 [hep-ph]

  15. [23]

    Modelling vector-like quarks in partial composite- ness framework,

    Avik Banerjee, Diogo Buarque Franzosi, and Gabriele Fer- retti, “Modelling vector-like quarks in partial composite- ness framework,” JHEP03, 200 (2022), arXiv:2202.00037 [hep-ph]

  16. [24]

    To- wardminimalcompositeHiggsmodelsfromregulargeome- tries in bottom-up holography,

    Daniel Elander, Ali Fatemiabhari, and Maurizio Piai, “To- wardminimalcompositeHiggsmodelsfromregulargeome- tries in bottom-up holography,” Phys. Rev. D107, 115021 (2023), arXiv:2303.00541 [hep-th]

  17. [25]

    Physics Beyond the Standard Model Associated with the Top Quark,

    Roberto Franceschini, “Physics Beyond the Standard Model Associated with the Top Quark,” Ann. Rev. Nucl. Part. Sci. 73, 397–420 (2023), arXiv:2301.04407 [hep-ph]

  18. [26]

    Heavy vector-like quarks decaying to exotic scalars: a case study with triplets,

    Avik Banerjee, Venugopal Ellajosyula, and Luca Panizzi, “Heavy vector-like quarks decaying to exotic scalars: a case study with triplets,” JHEP 01, 187 (2024), arXiv:2311.17877 [hep-ph]

  19. [27]

    Singlets in gaugetheorieswithfundamentalmatter,

    Ed Bennett, Ho Hsiao, Jong-Wan Lee, Biagio Lucini, Axel Maas, Maurizio Piai, and Fabian Zierler, “Singlets in gaugetheorieswithfundamentalmatter,”Phys.Rev.D 109, 034504 (2024), arXiv:2304.07191 [hep-lat]

  20. [28]

    Vector-like quarks: Status and new directions at the LHC,

    Avik Banerjee, Elin Bergeaas Kuutmann, Venugopal Ella- josyula, Rikard Enberg, Gabriele Ferretti, and Luca Panizzi, “Vector-like quarks: Status and new directions at the LHC,” SciPost Phys. Core7, 079 (2024), arXiv:2406.09193 [hep- ph]

  21. [29]

    Probing light scalars and vector-like quarks at the high-luminosity LHC,

    Umar Sohail Qureshi, Alfredo Gurrola, Andres Flórez, and Cristian Rodriguez, “Probing light scalars and vector-like quarks at the high-luminosity LHC,” Eur. Phys. J. C85, 379 (2025), arXiv:2410.17854 [hep-ph]

  22. [30]

    Searching for Dark Pho- tonswithMaverickTopPartners,

    Jeong Han Kim, Samuel D. Lane, Hye-Sung Lee, Ian M. Lewis, and Matthew Sullivan, “Searching for Dark Pho- tonswithMaverickTopPartners,”Phys.Rev.D 101,035041 (2020), arXiv:1904.05893 [hep-ph]

  23. [31]

    Exploring maverick top partner decays at the LHC,

    Shivam Verma, Sanjoy Biswas, Anirban Chatterjee, and Joy Ganguly, “Exploring maverick top partner decays at the LHC,” (2022), arXiv:2209.13888 [hep-ph]

  24. [32]

    Machine learning tagged boosted dark pho- ton: A signature of fermionic portal matter at the LHC,

    Shivam Verma, Sanjoy Biswas, Tanumoy Mandal, and Sub- hadip Mitra, “Machine learning tagged boosted dark pho- ton: A signature of fermionic portal matter at the LHC,” (2024), arXiv:2410.06925 [hep-ph]

  25. [33]

    Roadmap to explore vectorlike quarks decayingtoanewscalarorpseudoscalar,

    Akanksha Bhardwaj, Tanumoy Mandal, Subhadip Mitra, and Cyrin Neeraj, “Roadmap to explore vectorlike quarks decayingtoanewscalarorpseudoscalar,”Phys.Rev.D 106, 095014 (2022), arXiv:2203.13753 [hep-ph]

  26. [34]

    Discovery prospects of a vectorlike top partner decaying to a singlet boson,

    Akanksha Bhardwaj, Kartik Bhide, Tanumoy Mandal, Sub- hadip Mitra, and Cyrin Neeraj, “Discovery prospects of a vectorlike top partner decaying to a singlet boson,” Phys. Rev. D106, 075024 (2022), arXiv:2204.09005 [hep-ph]

  27. [35]

    Machine learning-enhanced search for a vectorlike singletBquarkdecayingtoasingletscalarorpseudoscalar,

    Jai Bardhan, Tanumoy Mandal, Subhadip Mitra, and Cyrin Neeraj, “Machine learning-enhanced search for a vectorlike singletBquarkdecayingtoasingletscalarorpseudoscalar,” Phys. Rev. D107, 115001 (2023), arXiv:2212.02442 [hep- ph]

  28. [36]

    FeynRules 2.0 - A complete toolboxfortree-levelphenomenology,

    Adam Alloul, Neil D. Christensen, Céline Degrande, Claude Duhr, and Benjamin Fuks, “FeynRules 2.0 - A complete toolboxfortree-levelphenomenology,”Comput.Phys.Com- mun. 185, 2250–2300 (2014), arXiv:1310.1921 [hep-ph]

  29. [37]

    UFO - The Universal FeynRules Output,

    Celine Degrande, Claude Duhr, Benjamin Fuks, David Grellscheid, Olivier Mattelaer, and Thomas Reiter, “UFO - The Universal FeynRules Output,” Comput. Phys. Commun. 183, 1201–1214 (2012), arXiv:1108.2040 [hep-ph]

  30. [38]

    The automated computation of tree-level and next-to- leading order differential cross sections, and their match- ing to parton shower simulations,

    J. Alwall, R. Frederix, S. Frixione, V. Hirschi, F. Maltoni, O.Mattelaer,H.S.Shao,T.Stelzer,P.Torrielli, andM.Zaro, “The automated computation of tree-level and next-to- leading order differential cross sections, and their match- ing to parton shower simulations,” JHEP07, 079 ...

  31. [39]

    An introduction to PYTHIA 8.2,

    Torbjörn Sjöstrand, Stefan Ask, Jesper R. Christiansen, RichardCorke, NishitaDesai, PhilipIlten, StephenMrenna, Stefan Prestel, Christine O. Rasmussen, and Peter Z. Skands, “An introduction to PYTHIA 8.2,” Comput. Phys. Commun. 191, 159–177 (2015), arXiv:1410.3012 [hep- ph]

  32. [40]

    DELPHES 3, A modular framework for fast simulation of a generic collider experiment,

    J. de Favereau, C. Delaere, P. Demin, A. Giammanco, V. Lemaître, A. Mertens, and M. Selvaggi (DELPHES 3), “DELPHES 3, A modular framework for fast simulation of a generic collider experiment,” JHEP02, 057 (2014), arXiv:1307.6346 [hep-ex]

  33. [41]

    Identificationofheavy-flavour jetswiththeCMSdetectorinppcollisionsat13TeV,

    A.M.Sirunyan et al.(CMS),“Identificationofheavy-flavour jetswiththeCMSdetectorinppcollisionsat13TeV,”JINST 13, P05011 (2018), arXiv:1712.07158 [physics.ins-det]

  34. [42]

    The anti-kt jet clustering algorithm,

    Matteo Cacciari, Gavin P. Salam, and Gregory Soyez, “The anti-kt jet clustering algorithm,” JHEP 04, 063 (2008), arXiv:0802.1189 [hep-ph]

  35. [43]

    Search for pair production of vectorlike quarks in the fully hadronic final state,

    Albert M Sirunyanet al. (CMS), “Search for pair production of vectorlike quarks in the fully hadronic final state,” Phys. Rev. D100, 072001 (2019), arXiv:1906.11903 [hep-ex]

  36. [44]

    Combi- nation of electroweak and QCD corrections to single W pro- ductionattheFermilabTevatronandtheCERNLHC,

    Giovanni Balossini, Guido Montagna, Carlo Michel Car- loni Calame, Mauro Moretti, Oreste Nicrosini, Fulvio Pic- cinini, Michele Treccani, and Alessandro Vicini, “Combi- nation of electroweak and QCD corrections to single W pro- ductionattheFermilabTevatronandtheCERNLHC,”JHEP 0...

  37. [45]

    Vector boson pro- duction at hadron colliders: A fully exclusive qcd calcula- tion at next-to-next-to-leading order,

    Stefano Catani, Leandro Cieri, Giancarlo Ferrera, Daniel de Florian, and Massimiliano Grazzini, “Vector boson pro- duction at hadron colliders: A fully exclusive qcd calcula- tion at next-to-next-to-leading order,” Phys. Rev. Lett.103, 082001 (2009)

  38. [46]

    Top Quark Pair Production beyond NNLO,

    Claudio Muselli, Marco Bonvini, Stefano Forte, Simone Marzani, and Giovanni Ridolfi, “Top Quark Pair Production beyond NNLO,” JHEP08, 076 (2015), arXiv:1505.02006 [hep-ph]

  39. [47]

    Theoretical results for electroweak- boson and single-top production,

    Nikolaos Kidonakis, “Theoretical results for electroweak- boson and single-top production,” PoS DIS2015, 170 (2015), arXiv:1506.04072 [hep-ph]. 12

  40. [48]

    Vector boson pair production at the LHC,

    John M. Campbell, R. Keith Ellis, and Ciaran Williams, “Vector boson pair production at the LHC,” JHEP07, 018 (2011), arXiv:1105.0020 [hep-ph]

  41. [49]

    Report from Working Group 2: Higgs Physics at the HL-LHC and HE-LHC,

    M. Cepeda et al., “Report from Working Group 2: Higgs Physics at the HL-LHC and HE-LHC,” CERN Yellow Rep. Monogr.7, 221–584 (2019), arXiv:1902.00134 [hep-ph]

  42. [50]

    Associatedproduc- tion of a top quark pair with a heavy electroweak gauge bo- sonatNLO +NNLLaccuracy,

    Anna Kulesza, Leszek Motyka, Daniel Schwartländer, TomaszStebel, andVincentTheeuwes,“Associatedproduc- tion of a top quark pair with a heavy electroweak gauge bo- sonatNLO +NNLLaccuracy,”Eur.Phys.J.C 79,249(2019), arXiv:1812.08622 [hep-ph]

  43. [51]

    Handbook of LHC Higgs Cross Sections: 4. De- ciphering the Nature of the Higgs Sector,

    D. de Florian et al. (LHC Higgs Cross Section Working Group), “Handbook of LHC Higgs Cross Sections: 4. De- ciphering the Nature of the Higgs Sector,”2/2017 (2016), 10.23731/CYRM-2017-002, arXiv:1610.07922 [hep-ph]

  44. [52]

    Matching matrix elements and shower evolution for top-quark production in hadronic col- lisions,

    Michelangelo L. Mangano, Mauro Moretti, Fulvio Piccinini, and Michele Treccani, “Matching matrix elements and shower evolution for top-quark production in hadronic col- lisions,” JHEP01, 013 (2007), arXiv:hep-ph/0611129

  45. [53]

    GraphNeuralNetworksinParticlePhysics,

    Jonathan Shlomi, Peter Battaglia, and Jean-Roch Vli- mant,“GraphNeuralNetworksinParticlePhysics,” (2020), 10.1088/2632-2153/abbf9a, arXiv:2007.13681 [hep-ex]

  46. [54]

    Object condensation: one-stage grid-free multi-object reconstruction in physics detectors, graph and image data,

    Jan Kieseler, “Object condensation: one-stage grid-free multi-object reconstruction in physics detectors, graph and image data,” Eur. Phys. J. C 80, 886 (2020), arXiv:2002.03605 [physics.data-an]

  47. [55]

    ParticleNet: Jet Tagging via Particle Clouds,

    Huilin Qu and Loukas Gouskos, “ParticleNet: Jet Tagging via Particle Clouds,” Phys. Rev. D 101, 056019 (2020), arXiv:1902.08570 [hep-ph]

  48. [56]

    ABCNet: An attention-based method for particle tagging,

    Vinicius Mikuni and Florencia Canelli, “ABCNet: An attention-based method for particle tagging,” Eur. Phys. J. Plus135,463(2020),arXiv:2001.05311[physics.data-an]

  49. [57]

    Probing the triple Higgs boson coupling with machine learning at the LHC,

    Murat Abdughani, Daohan Wang, Lei Wu, Jin Min Yang, and Jun Zhao, “Probing the triple Higgs boson coupling with machine learning at the LHC,” Phys. Rev. D 104, 056003 (2021), arXiv:2005.11086 [hep-ph]

  50. [58]

    Graph Neural Networks for IceCube Signal Classification,

    Nicholas Choma, Federico Monti, Lisa Gerhardt, Tomasz Palczewski, Zahra Ronaghi, Prabhat, Wahid Bhimji, Michael M. Bronstein, Spencer R. Klein, and Joan Bruna (IceCube), “Graph Neural Networks for IceCube Signal Classification,” (2018), arXiv:1809.06166 [cs.LG]

  51. [59]

    Graph Neural Networks for low-energy event classification & reconstruction in IceCube,

    R. Abbasi et al., “Graph Neural Networks for low-energy event classification & reconstruction in IceCube,” JINST17, P11003 (2022), arXiv:2209.03042 [hep-ex]

  52. [60]

    Edge-augmented graph trans- formers: Global self-attention is enough for graphs,

    Md. Shamim Hussain, Mohammed J. Zaki, and Dhar- mashankar Subramanian, “Edge-augmented graph trans- formers: Global self-attention is enough for graphs,” CoRR abs/2108.03348 (2021), 2108.03348

  53. [61]

    An analysis of virtual nodes in graph neural networks for link prediction,

    EunJeong Hwang, Veronika Thost, and Tengfei Ma Shib Sankar Dasgupta, “An analysis of virtual nodes in graph neural networks for link prediction,” Learning on Graphs (2022)

  54. [62]

    BERT: pre-training of deep bidirec- tional transformers for language understanding,

    Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova, “BERT: pre-training of deep bidirec- tional transformers for language understanding,” CoRR abs/1810.04805 (2018), 1810.04805

  55. [63]

    How at- tentive are graph attention networks?

    Shaked Brody, Uri Alon, and Eran Yahav, “How at- tentive are graph attention networks?” arXiv preprint arXiv:2105.14491 (2021)

  56. [64]

    Graph atten- tion networks,

    PetarVeličković, GuillemCucurull, ArantxaCasanova, Adri- ana Romero, Pietro Lio, and Yoshua Bengio, “Graph atten- tion networks,” (2017), arXiv:1710.10903 [stat.ML]

  57. [65]

    In- ductive representation learning on large graphs,

    Will Hamilton, Zhitao Ying, and Jure Leskovec, “In- ductive representation learning on large graphs,” Ad- vancesinneuralinformationprocessingsystems 30(2017), arXiv:1706.02216 [cs.SI]

  58. [66]

    Batch Normalization: Accelerating Deep Network Training by Reducing Internal Covariate Shift,

    Sergey Ioffe and Christian Szegedy, “Batch Normalization: Accelerating Deep Network Training by Reducing Internal Covariate Shift,” (2015), arXiv:1502.03167 [cs.LG]

  59. [67]

    Dropout: A simple way to prevent neural networks from overfitting,

    Nitish Srivastava, Geoffrey Hinton, Alex Krizhevsky, Ilya Sutskever, and Ruslan Salakhutdinov, “Dropout: A simple way to prevent neural networks from overfitting,” Journal of Machine Learning Research15, 1929–1958 (2014)

  60. [68]

    DecoupledWeightDecay Regularization,

    IlyaLoshchilovandFrankHutter,“DecoupledWeightDecay Regularization,” (2017), arXiv:1711.05101 [cs.LG]

  61. [69]

    Experiment tracking with weights and bi- ases,

    Lukas Biewald, “Experiment tracking with weights and bi- ases,” (2020), software available from wandb.com

  62. [70]

    Accelerating Resonance Searches via Signature- Oriented Pre-training,

    CongqiaoLi,AntoniosAgapitos,JovinDrews,JavierDuarte, Dawei Fu, Leyun Gao, Raghav Kansal, Gregor Kasieczka, Louis Moureaux, Huilin Qu, Cristina Mantilla Suarez, and Qiang Li, “Accelerating Resonance Searches via Signature- Oriented Pre-training,” (2024), arXiv:2405.12972 [hep- ph]

  63. [71]

    Llama: Open and efficient foundation language models,

    Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Mar- tinet, Marie-AnneLachaux, TimothéeLacroix, BaptisteRoz- ière, Naman Goyal, Eric Hambro, Faisal Azhar, Aurélien Ro- driguez, Armand Joulin, Edouard Grave, and Guillaume Lample, “Llama: Open and efficient foundation lang...

  64. [72]

    Gemini 1.5: Unlocking multimodal under- standing across millions of tokens of context,

    Machel Reid, Nikolay Savinov, Denis Teplyashin, Dmitry Lepikhin, Timothy Lillicrap, Jean-baptiste Alayrac, Radu Soricut, Angeliki Lazaridou, Orhan Firat, Julian Schrit- twieser, et al., “Gemini 1.5: Unlocking multimodal under- standing across millions of tokens of context,” (2...

  65. [73]

    Language models are few-shot learners,

    Tom B. Brown, Benjamin Mann, Nick Ryder, Melanie Sub- biah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakan- tan, Pranav Shyam, Girish Sastry, Amanda Askell, Sand- hini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel M. Ziegler,...

  66. [74]

    Asymptotic formulae for likelihood-based tests of new physics,

    Glen Cowan, Kyle Cranmer, Eilam Gross, and Ofer Vitells, “Asymptotic formulae for likelihood-based tests of new physics,” Eur. Phys. J. C 71, 1554 (2011), [Erra- tum: Eur.Phys.J.C 73, 2501 (2013)], arXiv:1007.1727 [physics.data-an]

  67. [75]

    GEANT4 - A Simulation Toolkit,

    S. Agostinelli et al. (GEANT4), “GEANT4 - A Simulation Toolkit,” Nucl. Instrum. Meth. A506, 250–303 (2003)

  68. [76]

    ProbingColorOctet Electrons at the LHC,

    TanumoyMandalandSubhadipMitra,“ProbingColorOctet Electrons at the LHC,” Phys. Rev. D87, 095008 (2013), arXiv:1211.6394 [hep-ph]

  69. [77]

    Sin- gle Productions of Colored Particles at the LHC: An Ex- ample with Scalar Leptoquarks,

    Tanumoy Mandal, Subhadip Mitra, and Satyajit Seth, “Sin- gle Productions of Colored Particles at the LHC: An Ex- ample with Scalar Leptoquarks,” JHEP 07, 028 (2015), arXiv:1503.04689 [hep-ph]

  70. [78]

    Particle Trans- former for Jet Tagging,

    Huilin Qu, Congqiao Li, and Sitian Qian, “Particle Trans- former for Jet Tagging,” (2022), arXiv:2202.03772 [hep- ph]

  71. [79]

    Explainableequivari- ant neural networks for particle physics: PELICAN,

    Alexander Bogatskiy, Timothy Hoffman, David W. Miller, 13 JanT.Offermann, andXiaoyangLiu,“Explainableequivari- ant neural networks for particle physics: PELICAN,” JHEP 03, 113 (2024), arXiv:2307.16506 [hep-ph]

  72. [80]

    Learning the language of QCD jets with trans- formers,

    Thorben Finke, Michael Krämer, Alexander Mück, and Jan Tönshoff, “Learning the language of QCD jets with trans- formers,” JHEP 06, 184 (2023), arXiv:2303.07364 [hep- ph]

  73. [81]

    Masked particle modeling on sets: to- wards self-supervised high energy physics foundation models,

    Tobias Golling, Lukas Heinrich, Michael Kagan, Samuel Klein, Matthew Leigh, Margarita Osadchy, and John An- drew Raine, “Masked particle modeling on sets: to- wards self-supervised high energy physics foundation models,” Mach. Learn. Sci. Tech. 5, 035074 (2024), arXiv:2401.135...

  74. [82]

    Jet tagging with more-interaction parti- cle transformer*,

    Yifan Wu, Kun Wang, Congqiao Li, Huilin Qu, and Jingya Zhu, “Jet tagging with more-interaction parti- cle transformer*,” Chin. Phys. C 49, 013110 (2025), arXiv:2407.08682 [hep-ph]

  75. [83]

    HEP-JEPA: A foundation model for collider physics using joint embedding predictive architecture,

    Jai Bardhan, Radhikesh Agrawal, Abhiram Tilak, Cyrin Neeraj, and Subhadip Mitra, “HEP-JEPA: A foundation model for collider physics using joint embedding predictive architecture,” (2025), arXiv:2502.03933 [cs.LG]. 14

Pith tools

Reviewed August 15, 2026 · model on record in the stance chip above.