Pith. sign in

REVIEW 4 major objections 5 minor 3 cited by

The paper claims that, at the precision frontier, explicitly encoding Lorentz symmetry in the architecture and implicitly learning physics from large-scale pretraining produce comparable performance, implying the gains from physics-aware ML

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

2026-08-03 02:31 UTC pith:XAQUOVDH

load-bearing objection Useful, mostly careful comparison of L-GATr vs OmniLearn on precision jet tasks, but the abstract's 'comparable across all benchmarks' is contradicted by the HERA results and needs revision. the 4 major comments →

arxiv 2603.08802 v3 pith:XAQUOVDH submitted 2026-03-09 hep-ph

Explicit or Implicit? Encoding Physics at the Precision Frontier

Victor Breso-Pla , Kevin Greif , Vinicius Mikuni , Benjamin Nachman , Tilman Plehn , Tanvi Wamorkar , Daniel Whiteson This is my paper
classification hep-ph
keywords physics priorsLorentz equivariancefoundation modelspretraininglikelihood ratio estimationunfoldinganomaly detection
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.

The paper asks whether it is better to build known particle-physics symmetries into a machine-learning model explicitly, or to let a large pretrained model absorb them from data. It compares two state-of-the-art representatives—L-GATr, a transformer whose operations are constrained to be Lorentz-equivariant, and OmniLearn, a foundation model pretrained on roughly one hundred million jets—on three tasks where the two classes to be separated are almost identical: reweighting-based unfolding in pp collisions, likelihood-ratio estimation in ep collisions at H1, and weakly supervised anomaly detection. On all three benchmarks the two approaches perform within statistical uncertainties of each other, and both clearly beat a randomly initialized network trained from scratch. The one clear exception, the H1 task, goes in favor of the implicitly pretrained model, and the authors trace that gap to the architecture's local feature processing rather than to the encoding strategy. The paper's central suggestion is that the sizeable efficiency gains from encoding known physics are largely method-independent: what matters is that physics is encoded, not whether it is hard-coded or learned.

Core claim

The paper's central claim is that, for precision classification where the classes are nearly identical, the explicit and implicit routes for injecting physics knowledge converge. Using L-GATr as the explicit representative and OmniLearn as the implicit one, the study finds comparable performance on unfolding-based reweighting for pp→Z+jets, likelihood-ratio estimation for ep deep inelastic scattering, and weakly supervised anomaly detection on the LHC Olympics benchmark. Both methods substantially outperform a PET network trained from scratch, showing that either form of physics prior adds real signal. The authors note one systematic exception: on the H1 ep task L-GATr consistently underperf

What carries the argument

The comparison is carried by two contrasting mechanisms for embedding physics. L-GATr maps each particle's four-momentum into the spacetime geometric algebra as a multivector, and restricts all linear and attention operations to act grade-by-grade so the whole network commutes with Lorentz transformations; symmetry can be broken dynamically by adding extra tokens. OmniLearn instead uses a Point-Edge Transformer pretrained on a large corpus of simulated jets (roughly 10^8), so the network acquires Lorentz-relevant structure and jet substructure correlations from data rather than from architectural constraints. A third ingredient, the same PET architecture trained from scratch, serves as the c

Load-bearing premise

The conclusion hinges on L-GATr and OmniLearn being fair representatives of the explicit and implicit strategies, so that observed differences reflect the encoding strategy rather than the specific architectures; the H1 result already challenges this, since the gap there is attributed to an architectural detail, not to the implicit prior.

What would settle it

Train an explicitly equivariant model that also processes local neighborhoods of constituents (e.g., an equivariant graph-transformer) on the H1 likelihood-ratio task and compare AUC with PET (about 0.569) and OmniLearn (about 0.570). If it does not close the gap, local feature processing is not the full explanation and the method-independence claim would need to be restricted further; if it does, the paper's attribution is confirmed and the equivalence claim stands only for tasks without strong locality requirements.

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

If this is right

  • On the three precision benchmarks studied, explicit Lorentz equivariance and large-scale pretraining are interchangeable in performance, so practitioners can choose between them based on compute, memory, and deployment constraints rather than expected accuracy.
  • Both strategies sharply outperform a from-scratch network, confirming that encoding known physics—by either route—is what drives the efficiency gain, not model size alone.
  • The H1 result shows the equivalence does not hold for every collision system: when classes are extremely similar and local constituent structure matters, an architecture with local feature processing can beat an equivariant transformer even without pretraining.
  • Because both approaches plateau at similar accuracy and are limited by finetuning dataset size, further gains on these tasks are more likely to come from more or better training data than from choosing one encoding strategy.
  • The two strategies can be combined; the paper notes there is no practical impediment to leveraging explicit equivariance and implicit pretraining at the same time.

Where Pith is reading between the lines

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

  • If method-independence generalizes beyond these benchmarks, then expensive equivariant architecture development and expensive foundation-model pretraining are partially substitute investments; future effort may concentrate on hybrid models that combine both, or on curating training data rather than inventing new architectures.
  • The HERA exception yields a testable prediction: adding local neighborhood aggregation to an explicitly equivariant model should recover PET-level performance, which would confirm the authors' architecture-detail explanation and narrow the method-independence claim to globally dominated tasks.
  • A natural extension is to test the same three tasks with an explicitly equivariant model that includes local message passing, or with an implicitly pretrained model that lacks it, to map exactly where the two strategies diverge.
  • For anomaly detection, the paper's uncertainties leave room for small but systematic differences (implicit better at low signal injection, explicit at high); a higher-statistics version of the LHC Olympics benchmark could resolve whether these are real or noise.

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 / 5 minor

Summary. This paper compares two strategies for encoding physics knowledge into machine-learning models for particle physics: explicit Lorentz equivariance, represented by L-GATr, and implicit learning through large-scale pretraining, represented by OmniLearn, a PET-based foundation model. On three precision benchmarks with nearly identical classes — reweighting-based unfolding for Z+jets at the LHC, likelihood-ratio estimation for HERA deep-inelastic scattering, and weakly supervised anomaly detection on the LHC Olympics dataset — the authors attempt to match inputs, parameter counts, epochs, and training protocol. They report that, across all benchmarks, the two methods achieve comparable performance given the statistical precision of the finetuning datasets, and conclude that the efficiency gains from encoding known physics structures are largely method-independent. The paper includes code release, a computational-resource analysis, and hyperparameter scans in appendices.

Significance. If the headline claim were correct, this would be an important result: it would suggest that physics structure need not be hard-coded into architectures and can instead be learned through pretraining, with substantial practical implications for model selection and computational cost. The study is carefully set up for the first two benchmarks, with matched inputs, parameter counts, epochs, public datasets, and multiple random seeds, and the code is released. However, the HERA benchmark (Table 4) shows L-GATr consistently underperforming even a from-scratch PET, and the authors themselves attribute this to an architectural detail (local feature processing) rather than to the implicit-pretraining strategy. This internal tension undermines the broad 'method-independent' conclusion. The paper is a useful transparent comparison of two specific state-of-the-art models, but the abstract-level generalization overstates the evidence.

major comments (4)
  1. [Abstract and Section 3.2, Table 4] The statement 'Across all benchmarks, both methods achieve comparable performance' is directly contradicted by the HERA results. L-GATr with 10^6 parameters reaches AUC 0.5603 and 1/epsilon_B = 2.396±0.003, while the from-scratch PET reaches 0.5691 and 2.467±0.002, and OmniLearn 0.5695 and 2.470±0.003. The authors themselves call L-GATr 'slightly, but consistently worse' in Section 3.2. These gaps are several times the reported 1/epsilon_B uncertainties, and no uncertainty is given for AUC. The 'given statistical precision' qualifier is therefore not substantiated for this benchmark. At minimum, the abstract must be qualified and the statistical criterion defined.
  2. [Section 3.2 and Outlook point 2] The paper attributes the HERA deficit to the local feature processing in PET, which is an architecture detail rather than a property of the implicit-pretraining strategy. Since each strategy is represented by a single model, the comparison cannot separate encoding strategy from architecture choice. The conclusion 'largely method-independent' is therefore not supported by the evidence presented. The results only support a comparison of these two specific models. The only task where the two strategies clearly differ is also the one where the explanation given is architectural, which further weakens the general claim.
  3. [Section 3.3, Figure 4] The anomaly detection comparison is not performed under the same matched protocol as the other two benchmarks. OmniLearn and the random PET results are quoted from Ref. [93] (by the same authors), while L-GATr is trained in this work. The text does not specify whether the OmniLearn/PET training used the same epochs, hyperparameter search, batch size, or ensemble size as the new L-GATr runs. This third benchmark therefore cannot carry the same weight as Sections 3.1–3.2 in supporting the 'comparable performance' claim, and the comparison is less controlled than implied by the paper's methodology description.
  4. [Section 3.1 and Appendix B] For the unfolding task, L-GATr results are reported as the best of 6 independent trainings (Appendix B: 'display the top performer'), while OmniLearn is described as stable and presumably reported as a single run. This asymmetry is favorable to L-GATr. It strengthens the conclusion that L-GATr genuinely underperforms on HERA (where best-of-N still loses), but it weakens the 'comparable' conclusion on the unfolding task, where the comparison is between L-GATr's best and OmniLearn's typical performance. The text should either use a consistent reporting rule or discuss the effect of this asymmetry on the stated conclusions.
minor comments (5)
  1. [References] References [40] and [93] cite the same paper (Mikuni & Nachman, arXiv:2502.14652); they should be consolidated to avoid duplicate citation.
  2. [Figures 1 and 2] The captions mention PET, but the legends show only L-GATr and OmniLearn. Either add PET to the legends or correct the captions.
  3. [Appendix A] Typographical issues: 'arnings' should be 'earnings' in the resources analysis, and 'set the set' should be 'set the'. These do not affect the content.
  4. [Table 1 caption] The notation 'f (PID)' is undefined; presumably it denotes a particle-identification feature. Please define it explicitly, as it is used in the input description.
  5. [Appendix B] The hyperparameter scan lists 'Learning rate:{10^5,3×10^5,10^4}' without negative exponents, which is inconsistent with the values (e.g., 3×10^-5) in Table 7. Please correct the formatting.

Circularity Check

0 steps flagged

Empirical comparison study with no derivation chain; no circular reduction to inputs.

full rationale

This paper is an empirical benchmark comparison, not a derivation. The central claim that explicit and implicit physics encoding are 'largely method-independent' is an inductive summary of the measured results in Tables 2–4 and Figure 4, not a quantity derived from a fitted input. No parameter is fitted to the reported performance metrics and then renamed a prediction: L-GATr is trained from scratch, OmniLearn is fine-tuned from public pretrained weights, and both are evaluated on fixed benchmark datasets (Z+jets, H1, LHCO). The HERA benchmark where L-GATr underperforms PET and OmniLearn is explicitly reported in Table 4 and acknowledged in Section 3.2 and Outlook point 2; this is a potential overgeneralization or external-validity concern, not circular reasoning. The self-citations to Refs. [20,24,38,39,93] define the representative models and supply prior benchmark numbers, but those numbers are empirical, externally checkable results from public datasets and model releases, not results whose validity depends on the present paper's conclusions. No uniqueness theorem, ansatz-via-citation, or renaming of a known result into new coordinates is present. The practice of reporting the best of three seeded runs weakens statistical claims but does not constitute a fitted-input-called-prediction loop. Overall, no circularity is found.

Axiom & Free-Parameter Ledger

4 free parameters · 4 axioms · 0 invented entities

No physical constants are fitted and no new entities are postulated. The free parameters are ML training choices (listed in Table 7 and Appendix B), chosen by hand and via small sweeps to match the OmniLearn protocol; they are not fitted to the benchmark observables. The central claim rests on domain assumptions about Lorentz symmetry, OmniFold's classifier-reweighting surrogate, and the fairness of the matched-training protocol; the last is the most fragile because best-of-N selection is applied to L-GATr but not to OmniLearn.

free parameters (4)
  • L-GATr hyperparameters (learning rate, weight decay, batch size) = LR 3e-5 / 2e-6 / 5e-4; WD 0.1 / 0.2 / 1e-3; BS 512 / 1024
    Chosen by hand and via small sweeps to match OmniLearn's setup; not fitted to physics observables.
  • L-GATr architecture size = 1e6 / 2e6 / 1.8e6 parameters
    Chosen to match OmniLearn's parameter count; performance changes little when increased.
  • Training epochs = 20 (unfolding/HERA), 60 (anomaly)
    Matched to OmniLearn; validation loss saturates by epoch 10.
  • Feature preprocessing and normalization = Feature sets in Tables 1 and 5
    Input selection and standardization follow Ref [39] and Refs [93,94]; these choices affect results but are not fitted to the benchmark.
axioms (4)
  • domain assumption Lorentz equivariance is the relevant symmetry for jet and event classification at LHC and HERA; L-GATr's geometric-algebra construction preserves it.
    Section 2.1; if wrong, the comparison of explicit encoding loses meaning for these tasks.
  • domain assumption A classifier trained for reweighting in OmniFold estimates the likelihood ratio needed for unfolding.
    Section 3.1 and Refs [54,55]; the entire unfolding benchmark depends on this surrogate.
  • ad hoc to paper Matching parameter counts, inputs, epochs, and using best-of-N L-GATr runs is a fair comparison protocol.
    Section 3.1 and Appendix B; this is not derived and biases toward L-GATr.
  • ad hoc to paper The finetuning datasets are large enough that remaining performance differences are within 'statistical precision'.
    Abstract; no quantitative threshold is given, and Section 3.2 shows a consistent if small gap.

pith-pipeline@v1.3.0-alltime-deepseek · 18155 in / 12226 out tokens · 119723 ms · 2026-08-03T02:31:20.351956+00:00 · methodology

0 comments
read the original abstract

High-performance machine learning tools in particle physics rest on two complementary directions: encoding symmetries explicitly in the architecture, and implicitly learning the structure of the data through large-scale (pre-) training. We compare the performance of the representative L-GATr and OmniLearn models on three especially challenging tasks: reweighting-based unfolding, likelihood-ratio estimation, and weakly supervised anomaly detection. Across all benchmarks, both methods achieve comparable performance given the statistical precision of the finetuning datasets, suggesting that the significant efficiency gains from encoding known particle physics structures are largely method-independent.

discussion (0)

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

Forward citations

Cited by 3 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. Generative models on phase space

    hep-ph 2026-04 unverdicted novelty 8.0

    Generative diffusion and flow models are constructed to remain exactly on the Lorentz-invariant massless N-particle phase space manifold during sampling for particle physics applications.

  2. One Generator, Any Process: LLM-Conditioning for the LHC

    hep-ph 2026-06 unverdicted novelty 7.0

    LLM embeddings condition generative networks for LHC events, yielding faster convergence, higher quality, and generalization to unseen processes.

  3. One Generator, Any Process: LLM-Conditioning for the LHC

    hep-ph 2026-06 unverdicted novelty 6.0

    LLM embeddings condition a generative transformer to enable faster convergence, better performance, and generalization to unseen LHC processes using a single model.

Reference graph

Works this paper leans on

93 extracted references · 1 canonical work pages · cited by 2 Pith papers

  1. [1]

    Radovic, M

    A. Radovic, M. Williams, D. Rousseau, M. Kagan, D. Bonacorsi, A. Himmel et al.,Machine learning at the energy and intensity frontiers of particle physics,Nature560(2018) 41–48

    2018

  2. [2]

    Karagiorgi, G

    G. Karagiorgi, G. Kasieczka, S. Kravitz, B. Nachman and D. Shih,Machine Learning in the Search for New Fundamental Physics,2112.03769

    Pith/arXiv arXiv

  3. [3]

    Feickert and B

    M. Feickert and B. Nachman,A Living Review of Machine Learning for Particle Physics, 2102.02770

    Pith/arXiv arXiv

  4. [4]

    Plehn, A

    T. Plehn, A. Butter, B. Dillon, T. Heimel, C. Krause and R. Winterhalder,Modern Machine Learning for LHC Physicists,2211.01421

    Pith/arXiv arXiv

  5. [6]

    Akhound-Sadegh, L

    T. Akhound-Sadegh, L. Perreault-Levasseur, J. Brandstetter, M. Welling and S. Ravanbakhsh, Lie Point Symmetry and Physics-Informed Networks, inAdvances in Neural Information Processing Systems 36, p. 42468, 2023

    2023

  6. [7]

    A. A. Elhag, T. K. Rusch, F. Di Giovanni and M. Bronstein,Relaxed Equivariance via Multitask Learning,2410.17878. – 20 –

    arXiv

  7. [8]

    Nabat, A

    S. Nabat, A. Ghosh, E. Witkowski, G. Kasieczka and D. Whiteson,Learning broken symmetries with approximate invariance,Phys. Rev. D111(2025) 072002, [2412.18773]

    Pith/arXiv arXiv 2025

  8. [9]

    Hebbar, T

    P. Hebbar, T. Madula, V. Mikuni, B. Nachman, N. Outmezguine and I. Savoray,SEAL - A Symmetry EncourAging Loss for High Energy Physics,2511.01982

    arXiv

  9. [10]

    Bogatskiy, B

    A. Bogatskiy, B. Anderson, J. T. Offermann, M. Roussi, D. W. Miller and R. Kondor,Lorentz Group Equivariant Neural Network for Particle Physics, inProceedings of the 37th International Conference on Machine Learning, vol. 119 ofProceedings of Machine Learning Research, pp. 992–1002, 2020,2006.04780

    Pith/arXiv arXiv 2020

  10. [11]

    Villar, D

    S. Villar, D. W. Hogg, K. Storey-Fisher, W. Yao and B. Blum-Smith,Scalars are universal: Equivariant machine learning, structured like classical physics, inAdvances in Neural Information Processing Systems 34, 2021,2106.06610

    Pith/arXiv arXiv 2021

  11. [12]

    S. Gong, Q. Meng, J. Zhang, H. Qu, C. Li, S. Qian et al.,An efficient Lorentz equivariant graph neural network for jet tagging,JHEP07(2022) 030, [2201.08187]

    Pith/arXiv arXiv 2022

  12. [13]

    S. Qiu, S. Han, X. Ju, B. Nachman and H. Wang,Holistic approach to predicting top quark kinematic properties with the covariant particle transformer,Phys. Rev. D107(2023) 114029, [2203.05687]

    Pith/arXiv arXiv 2023

  13. [14]

    Bogatskiy, S

    A. Bogatskiy, S. Ganguly, T. Kipf, R. Kondor, D. W. Miller, D. Murnane et al.,Symmetry Group Equivariant Architectures for Physics,2203.06153

    Pith/arXiv arXiv

  14. [15]

    Bogatskiy, T

    A. Bogatskiy, T. Hoffman, D. W. Miller and J. T. Offermann,PELICAN: Permutation Equivariant and Lorentz Invariant or Covariant Aggregator Network for Particle Physics, 2211.00454

    Pith/arXiv arXiv

  15. [16]

    C. Li, H. Qu, S. Qian, Q. Meng, S. Gong, J. Zhang et al.,Does Lorentz-symmetric design boost network performance in jet physics?,Phys. Rev. D109(2024) 056003, [2208.07814]

    Pith/arXiv arXiv 2024

  16. [17]

    Z. Hao, R. Kansal, J. Duarte and N. Chernyavskaya,Lorentz group equivariant autoencoders, Eur. Phys. J. C83(2023) 485, [2212.07347]

    Pith/arXiv arXiv 2023

  17. [18]

    Bogatskiy, T

    A. Bogatskiy, T. Hoffman, D. W. Miller, J. T. Offermann and X. Liu,Explainable equivariant neural networks for particle physics: PELICAN,JHEP03(2024) 113, [2307.16506]

    Pith/arXiv arXiv 2024

  18. [19]

    Batatia, M

    I. Batatia, M. Geiger, J. Munoz, T. Smidt, L. Silberman and C. Ortner,A General Framework for Equivariant Neural Networks on Reductive Groups, inAdvances in Neural Information Processing Systems 36, p. 55260, 2023

    2023

  19. [20]

    Brehmer, V

    J. Brehmer, V. Bres´ o, P. de Haan, T. Plehn, H. Qu, J. Spinner et al.,A Lorentz-equivariant transformer for all of the LHC,SciPost Phys.19(2025) 108, [2411.00446]

    arXiv 2025

  20. [21]

    Favaro, G

    L. Favaro, G. Gerhartz, F. A. Hamprecht, P. Lippmann, S. Pitz, T. Plehn et al., Lorentz-Equivariance without Limitations,2508.14898

    arXiv

  21. [22]

    H. Bahl, N. Elmer, L. Favaro, M. Haussmann, T. Plehn and R. Winterhalder,Accurate surrogate amplitudes with calibrated uncertainties,SciPost Phys. Core8(2025) 073, [2412.12069]

    arXiv 2025

  22. [23]

    Bogatskiy, T

    A. Bogatskiy, T. Hoffman and J. T. Offermann,19 Parameters Is All You Need: Tiny Neural Networks for Particle Physics, in37th Conference on Neural Information Processing Systems, 10, 2023,2310.16121

    Pith/arXiv arXiv 2023

  23. [24]

    Petitjean, T

    A. Petitjean, T. Plehn, J. Spinner and U. K¨ othe,Economical Jet Taggers – Equivariant, Slim, and Quantized,2512.17011. – 21 –

    arXiv

  24. [25]

    M. P. Kuchera, R. Ramanujan, J. Z. Taylor, R. R. Strauss, D. Bazin, J. Bradt et al.,Machine learning methods for track classification in the AT-TPC,Nucl. Instrum. Meth. A940(2019) 156, [1810.10350]

    Pith/arXiv arXiv 2019

  25. [26]

    B. M. Dillon, G. Kasieczka, H. Olischlager, T. Plehn, P. Sorrenson and L. Vogel,Symmetries, safety, and self-supervision,SciPost Phys.12(2022) 188, [2108.04253]

    Pith/arXiv arXiv 2022

  26. [27]

    B. M. Dillon, R. Mastandrea and B. Nachman,Self-supervised anomaly detection for new physics,Phys. Rev. D106(2022) 056005, [2205.10380]

    Pith/arXiv arXiv 2022

  27. [28]

    Chappell and L

    A. Chappell and L. H. Whitehead,Application of transfer learning to neutrino interaction classification,Eur. Phys. J. C82(2022) 1099, [2207.03139]

    Pith/arXiv arXiv 2022

  28. [29]

    F. A. Dreyer, R. Grabarczyk and P. F. Monni,Leveraging universality of jet taggers through transfer learning,Eur. Phys. J. C82(2022) 564, [2203.06210]

    Pith/arXiv arXiv 2022

  29. [30]

    Beauchesne, Z.-E

    H. Beauchesne, Z.-E. Chen and C.-W. Chiang,Pre-training strategies using contrastive learning and corpus transfer for jet physics,2312.06152

    Pith/arXiv arXiv

  30. [31]

    B. M. Dillon, L. Favaro, F. Feiden, T. Modak and T. Plehn,Anomalies, representations, and self-supervision,SciPost Phys. Core7(2024) 056, [2301.04660]

    Pith/arXiv arXiv 2024

  31. [32]

    J. Birk, A. Hallin and G. Kasieczka,OmniJet-α: The first cross-task foundation model for particle physics,2403.05618

    Pith/arXiv arXiv

  32. [33]

    Harris, M

    P. Harris, M. Kagan, J. Krupa, B. Maier and N. Woodward,Foundation models for particle physics event reconstruction,2403.07066

    Pith/arXiv arXiv

  33. [34]

    Golling, L

    T. Golling, L. Heinrich, M. Kagan, S. Klein, M. Leigh, M. Osadchy et al.,Masked particle modeling on sets: Towards self-supervised high energy physics foundation models,Mach. Learn. Sci. Tech.5(2024) 035074, [2401.13537]

    Pith/arXiv arXiv 2024

  34. [35]

    Leigh, S

    M. Leigh, S. Klein, F. Charton, T. Golling, L. Heinrich, M. Kagan et al.,Scaling masked particle modeling on sets,2409.12589

    Pith/arXiv arXiv

  35. [36]

    Hallin,OmniJet-α-2: A cross-task foundation model update,2509.21434

    A. Hallin,OmniJet-α-2: A cross-task foundation model update,2509.21434

    arXiv

  36. [37]

    Bardhan, R

    J. Bardhan, R. Agrawal, A. Tilak, C. Neeraj and S. Mitra,A self-supervised foundation model for jet physics,2502.03933

    Pith/arXiv arXiv

  37. [38]

    Brehmer, P

    J. Brehmer, P. de Haan, S. Behrends and T. Cohen,Geometric algebra transformer, 2023

    2023

  38. [39]

    Mikuni and B

    V. Mikuni and B. Nachman,Solving key challenges in collider physics with foundation models, Phys. Rev. D111(2025) L051504, [2404.16091]

    arXiv 2025

  39. [41]

    Bhimji, C

    W. Bhimji, C. Harris, V. Mikuni and B. Nachman,OmniLearned: A Foundation Model Framework for All Tasks Involving Jet Physics,2510.24066

    arXiv

  40. [42]

    Mikuni, I

    V. Mikuni, I. Elsharkawy and B. Nachman,OmniCosmos: Transferring Particle Physics Knowledge Across the Cosmos,2512.24422

    arXiv

  41. [43]

    Elsharkawy, V

    I. Elsharkawy, V. Mikuni, W. Bhimji and B. Nachman,OmniMol: Transferring Particle Physics Knowledge to Molecular Dynamics with Point-Edge Transformers,2601.10791

    Pith/arXiv arXiv

  42. [44]

    Butter et al.,The Machine Learning landscape of top taggers,SciPost Phys.7(2019) 014, [1902.09914]

    A. Butter et al.,The Machine Learning landscape of top taggers,SciPost Phys.7(2019) 014, [1902.09914]

    Pith/arXiv arXiv 2019

  43. [45]

    H. Qu, C. Li and S. Qian,Particle Transformer for Jet Tagging,2202.03772. – 22 –

    Pith/arXiv arXiv

  44. [46]

    Cranmer, J

    K. Cranmer, J. Brehmer and G. Louppe,The frontier of simulation-based inference, Proceedings of the National Academy of Sciences117(2020) 30055–30062, [https://www.pnas.org/doi/pdf/10.1073/pnas.1912789117]

    doi:10.1073/pnas.1912789117 2020

  45. [47]

    Canelli et al.,A practical guide to unbinned unfolding,Eur

    F. Canelli et al.,A practical guide to unbinned unfolding,Eur. Phys. J. C86(2026) 106, [2507.09582]

    arXiv 2026

  46. [48]

    Kasieczka et al.,The LHC Olympics 2020: A Community Challenge for Anomaly Detection in High Energy Physics,2101.08320

    G. Kasieczka et al.,The LHC Olympics 2020: A Community Challenge for Anomaly Detection in High Energy Physics,2101.08320

    Pith/arXiv arXiv 2020

  47. [49]

    Aarrestad et al.,The Dark Machines Anomaly Score Challenge: Benchmark Data and Model Independent Event Classification for the Large Hadron Collider,2105.14027

    T. Aarrestad et al.,The Dark Machines Anomaly Score Challenge: Benchmark Data and Model Independent Event Classification for the Large Hadron Collider,2105.14027

    Pith/arXiv arXiv

  48. [50]

    Belis, P

    V. Belis, P. Odagiu and T. K. Aarrestad,Machine learning for anomaly detection in particle physics,Rev. Phys.12(2024) 100091, [2312.14190]

    Pith/arXiv arXiv 2024

  49. [51]

    Petitjean, A

    A. Petitjean, A. Butter, K. Greif, S. Palacios Schweitzer, T. Plehn, J. Spinner et al., Generative Unfolding of Jets and Their Substructure,2510.19906

    arXiv

  50. [52]

    Batson and Y

    J. Batson and Y. F. Kahn,Scaling laws in jet classification,SciPost Phys. Core8(2025) 034, [2312.02264]

    Pith/arXiv arXiv 2025

  51. [53]

    M. Vigl, N. Hartman, M. Kagan and L. Heinrich,Neural Scaling Laws for Boosted Jet Tagging, 2602.15781

    arXiv

  52. [54]

    Andreassen, P

    A. Andreassen, P. T. Komiske, E. M. Metodiev, B. Nachman and J. Thaler,OmniFold: A Method to Simultaneously Unfold All Observables,Phys. Rev. Lett.124(2020) 182001, [1911.09107]

    Pith/arXiv arXiv 2020

  53. [55]

    Andreassen, P

    A. Andreassen, P. T. Komiske, E. M. Metodiev, B. Nachman, A. Suresh and J. Thaler, Scaffolding Simulations with Deep Learning for High-dimensional Deconvolution, 2105.04448

    Pith/arXiv arXiv

  54. [56]

    H¨ ocker and V

    A. H¨ ocker and V. Kartvelishvili,SVD approach to data unfolding,Nucl. Instrum. Meth. A372 (1996) 469–481, [hep-ph/9509307]

    Pith/arXiv arXiv 1996

  55. [57]

    G. D’Agostini,A multidimensional unfolding method based on bayes’ theorem,Nuclear Instruments and Methods in Physics Research Section A: Accelerators, Spectrometers, Detectors and Associated Equipment362(1995) 487–498

    1995

  56. [58]

    Cowan,A survey of unfolding methods for particle physics,Conf

    G. Cowan,A survey of unfolding methods for particle physics,Conf. Proc. C0203181(2002) 248–257

    2002

  57. [59]

    Blobel,Unfolding Methods in Particle Physics, inPHYSTAT 2011, (Geneva), pp

    V. Blobel,Unfolding Methods in Particle Physics, inPHYSTAT 2011, (Geneva), pp. 240–251, CERN, 2011, DOI

    2011

  58. [60]

    Schmitt,TUnfold: an algorithm for correcting migration effects in high energy physics, JINST7(2012) T10003, [1205.6201]

    S. Schmitt,TUnfold: an algorithm for correcting migration effects in high energy physics, JINST7(2012) T10003, [1205.6201]

    Pith/arXiv arXiv 2012

  59. [61]

    Behnke, K

    O. Behnke, K. Kr¨ oninger, T. Sch¨ orner-Sadenius and G. Schott, eds.,Data analysis in high energy physics: A practical guide to statistical methods. Wiley-VCH, Weinheim, Germany, 2013

    2013

  60. [62]

    Brenner, R

    L. Brenner, R. Balasubramanian, C. Burgard, W. Verkerke, G. Cowan, P. Verschuuren et al., Comparison of unfolding methods using RooFitUnfold,Int. J. Mod. Phys. A35(2020) 2050145, [1910.14654]

    Pith/arXiv arXiv 2020

  61. [63]

    Bellagente, A

    M. Bellagente, A. Butter, G. Kasieczka, T. Plehn and R. Winterhalder,How to GAN away Detector Effects,SciPost Phys.8(2020) 070, [1912.00477]. – 23 –

    Pith/arXiv arXiv 2020

  62. [64]

    Bellagente, A

    M. Bellagente, A. Butter, G. Kasieczka, T. Plehn, A. Rousselot, R. Winterhalder et al., Invertible Networks or Partons to Detector and Back Again,SciPost Phys.9(2020) 074, [2006.06685]

    Pith/arXiv arXiv 2020

  63. [65]

    Huetsch et al.,The landscape of unfolding with machine learning,SciPost Phys.18(2025) 070, [2404.18807]

    N. Huetsch et al.,The landscape of unfolding with machine learning,SciPost Phys.18(2025) 070, [2404.18807]

    Pith/arXiv arXiv 2025

  64. [66]

    Ore and T

    A. Ore and T. Plehn,Unfolding without Iterations, Adversaries, or Surrogates,2602.24282

    arXiv

  65. [67]

    Andreassen, P

    A. Andreassen, P. T. Komiske, E. M. Metodiev, B. Nachman and J. Thaler,Pythia/herwig + delphes jet datasets for omnifold unfolding, 2019. 10.5281/zenodo.3548091

    doi:10.5281/zenodo.3548091 2019

  66. [68]

    Bahr et al.,Herwig++ Physics and Manual,Eur

    M. Bahr et al.,Herwig++ Physics and Manual,Eur. Phys. J. C58(2008) 639–707, [0803.0883]

    Pith/arXiv arXiv 2008

  67. [69]

    Bellm et al.,Herwig 7.0/Herwig++ 3.0 release note,Eur

    J. Bellm et al.,Herwig 7.0/Herwig++ 3.0 release note,Eur. Phys. J. C76(2016) 196, [1512.01178]

    Pith/arXiv arXiv 2016

  68. [70]

    Bellm et al.,Herwig 7.1 Release Note,1705.06919

    J. Bellm et al.,Herwig 7.1 Release Note,1705.06919

    Pith/arXiv arXiv

  69. [71]

    Sj¨ ostrand, S

    T. Sj¨ ostrand, S. Ask, J. R. Christiansen, R. Corke, N. Desai, P. Ilten et al.,An introduction to PYTHIA 8.2,Comput. Phys. Commun.191(2015) 159–177, [1410.3012]. [72]ATLAS Pythia 8 Tunes to 7 TeV Data, tech. rep., CERN, Geneva, 2014. [73]DELPHES 3collaboration, J. de Favereau, C. Delaere, P. Demin, A. Giammanco, V. Lema ˆ ıtre, A. Mertens et al.,DELPHES ...

    Pith/arXiv arXiv 2015

  70. [74]

    Selvaggi,DELPHES 3: A modular framework for fast-simulation of generic collider experiments,J

    M. Selvaggi,DELPHES 3: A modular framework for fast-simulation of generic collider experiments,J. Phys. Conf. Ser.523(2014) 012033

    2014

  71. [75]

    Mertens,New features in Delphes 3,J

    A. Mertens,New features in Delphes 3,J. Phys. Conf. Ser.608(2015) 012045

    2015

  72. [76]

    Cacciari and G

    M. Cacciari and G. P. Salam,Dispelling theN 3 myth for thek t jet-finder,Phys. Lett. B641 (2006) 57–61, [hep-ph/0512210]

    Pith/arXiv arXiv 2006

  73. [77]

    Cacciari, G

    M. Cacciari, G. P. Salam and G. Soyez,The anti- kt jet clustering algorithm,JHEP04(2008) 063, [0802.1189]

    Pith/arXiv arXiv 2008

  74. [78]

    Cacciari, G

    M. Cacciari, G. P. Salam and G. Soyez,FastJet User Manual,Eur. Phys. J. C72(2012) 1896, [1111.6097]

    Pith/arXiv arXiv 2012

  75. [79]

    Arratia, D

    M. Arratia, D. Britzger, O. Long and B. Nachman,Reconstructing the Kinematics of Deep Inelastic Scattering with Deep Learning,2110.05505

    Pith/arXiv arXiv

  76. [80]

    Arratia, D

    M. Arratia, D. Britzger, O. Long and B. Nachman,Optimizing observables with machine learning for better unfolding,JINST17(2022) P07009, [2203.16722]

    Pith/arXiv arXiv 2022

  77. [81]

    Long and B

    O. Long and B. Nachman,Designing observables for measurements with deep learning,Eur. Phys. J. C84(2024) 776, [2310.08717]

    Pith/arXiv arXiv 2024

  78. [82]

    Charchula, G

    K. Charchula, G. A. Schuler and H. Spiesberger,Combined QED and QCD radiative effects in deep inelastic lepton - proton scattering: The Monte Carlo generator DJANGO6,Comput. Phys. Commun.81(1994) 381–402

    1994

  79. [83]

    Jung,Hard diffractive scattering in high-energy e p collisions and the Monte Carlo generator RAPGAP,Comput

    H. Jung,Hard diffractive scattering in high-energy e p collisions and the Monte Carlo generator RAPGAP,Comput. Phys. Commun.86(1995) 147–161

    1995

  80. [84]

    Peez,Search for deviations from the standard model in high transverse energy processes at the electron proton collider HERA, other thesis, 10, 2003

    M. Peez,Search for deviations from the standard model in high transverse energy processes at the electron proton collider HERA, other thesis, 10, 2003. 10.3204/DESY-THESIS-2003-023. – 24 –

    work page doi:10.3204/desy-thesis-2003-023 2003

Showing first 80 references.