Pith. sign in

REVIEW 4 major objections 5 minor 42 references

Pre-training on large simulated event samples, then fine-tuning per decay channel, matches or beats training from scratch across 12 charmed-baryon decay channels and improves angular resolution by up to 31 percent.

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 13:51 UTC pith:VAFMMPJF

load-bearing objection Solid, well-scoped ML benchmark for BESIII transfer learning; the core claim is plausible but only MC-validated, and the comparisons need error bars. the 4 major comments →

arxiv 2607.29088 v1 pith:VAFMMPJF submitted 2026-07-31 physics.data-an hep-ex

Learning transferable event representations for charmed baryon physics at BESIII

Kaixuan Huang , Yangu Li , Junpeng Zhao , Peilian Li , Peirong Li , Xiaorui Lyu , Yunxuan Song , Shengsen Sun
show 1 more author
Yangheng Zheng
This is my paper
classification physics.data-an hep-ex
keywords transfer learningParticle Transformercharmed baryon physicsLambda_c+BESIIIevent classificationmomentum-direction regressionMonte Carlo simulation
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.

This paper tries to establish that a single pre-trained deep network can serve many different physics analyses at the BESIII experiment instead of each analysis training its own model from scratch. Using production of the charmed baryon Lambda_c+ as a test case, the authors pre-train a Particle Transformer on roughly 71 million simulated events to classify event categories and to regress the momentum direction of the baryons. They then either apply the model directly or fine-tune it on each of 12 decay channels. The central claims are that fine-tuning matches or beats from-scratch training on every channel, with the largest gains where training data are scarce, and that the pre-trained regression model reduces angular resolution errors by about 31 percent in polar angle and 23 percent in azimuthal angle compared with conventional reconstruction. If these results hold on real data, the strategy would cut duplicated model development and improve sensitivity for rare or hard-to-reconstruct charm decays.

Core claim

On the paper's own terms, the central claim is that a Particle Transformer pre-trained on large inclusive Monte Carlo samples learns event representations that transfer to individual charmed-baryon decay channels with minimal fine-tuning. The classification model separates Lambda_c+ anti-Lambda_c- events from D-meson and light-quark backgrounds with an AUC of 0.988, rejecting 97.0% of background at 90% signal efficiency. Fine-tuning from the pre-trained model beats training from scratch on 11 of the 12 benchmark channels, with the largest gains at the smallest training-sample sizes. The regression model reduces angular resolution by 31.1% in theta and 22.8% in phi relative to conventional si

What carries the argument

The carrying mechanism is a Particle Transformer whose input is the event as an object-level point cloud: charged tracks and calorimeter clusters with their angles, momenta or energies, and low-level measurements used for particle identification, plus pairwise kinematic features (opening angles, energy ratios, and energy-momentum-conservation residuals) injected as a bias into the attention weights. The backbone uses particle-attention layers and a class-attention token to aggregate variable-length events into a fixed event representation. Pre-training performs two tasks - three-category event classification and regression of the Lambda_c+ and anti-Lambda_c+ momentum directions - and fine-tu

Load-bearing premise

Everything is trained and evaluated on Monte Carlo simulation; if the simulated events differ substantially from real BESIII data in detector response, background composition, or reconstruction efficiency, the reported gains could shrink or disappear on real data.

What would settle it

On real BESIII data, select double-tag events in which one Lambda_c+ decays hadronically and the other side is also fully reconstructed; use the pre-trained regression model to predict the momentum direction of the tag side and compare with its reconstructed direction. If the residual width on data is much larger than the 0.34 rad level found on simulation, the pre-trained representation has not transferred. Similarly, check the classifier's background rejection on real-data sidebands: if the 97 percent rejection does not hold at 90 percent signal efficiency on data, the simulation-only traini

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

If this is right

  • Fine-tuning from the pre-trained model yields higher AUC and background rejection than training from scratch on channels with limited Monte Carlo statistics, so analyses of rare modes need fewer simulated events to reach the same discrimination.
  • The pre-trained classifier can be applied directly, without fine-tuning, and still reject 31.6-88.1 percent of background at 90 percent signal efficiency across the 12 channels, making it a ready-made event filter.
  • The pre-trained regression model improves polar-angle resolution by about 31 percent over conventional single-tag reconstruction across the benchmark channels, so downstream measurements using these angles inherit finer resolution.
  • For final states with missing particles, such as semileptonic decays, the fine-tuned regression model recovers the Lambda_c+ direction substantially better than a model trained from scratch, reducing the need for double-tag reconstruction and its loss of signal statistics.
  • Because the same pre-trained backbone transfers across both tasks and across all tested channels, the strategy offers a scalable template for other BESIII analyses and for future e+e- experiments.

Where Pith is reading between the lines

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

  • The authors evaluate only on Monte Carlo simulation. If the simulated detector response matches real BESIII data, the same pre-trained weights would likely cut the required Monte Carlo sample sizes for many channels; a direct test on real double-tag data would be the natural next step.
  • Because the pre-training already includes a D-meson category, the same framework could be extended to charmed-meson decays without changing the input representation; a multi-class pre-training covering D0, D+, Ds and Lambda_c+ might yield one universal charm-physics foundation model.
  • The pairwise interaction features encode energy-momentum conservation in a spherical e+e- geometry, a design choice that could transfer directly to proposed Super Tau-Charm and CEPC detectors, which have similar event topologies.
  • A testable extension would be to replace the supervised pre-training categories with self-supervised objectives, such as masked-object prediction, to see whether the learned representations generalize to decay channels that are not among the 12 benchmarks.

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. The paper presents a Particle Transformer-based transfer-learning framework for BESIII charmed-baryon physics. The authors pre-train two models on large Monte Carlo samples at sqrt(s)=4.600–4.700 GeV: a classification model distinguishing Lambda_c+ Lambda_cbar, D Dbar(h), and q qbar events, and a regression model predicting the momentum directions of the Lambda_c baryons. These models are then evaluated by direct application, by fine-tuning, and by training from scratch on 12 hadronic Lambda_c decay channels, plus the semileptonic channel Lambda_c -> p K- e+ nu_e for the regression task. The central claims are: (i) fine-tuned classification models achieve comparable or better signal/background discrimination than from-scratch models, with the largest gains in low-statistics channels, and (ii) the pre-trained regression model improves angular resolution by ~31% (theta) and ~23% (phi) over conventional single-tag reconstruction, with further fine-tuning gains in the semileptonic mode. All quantitative results are obtained exclusively from Geant4/EvtGen simulation; no real BESIII data are used.

Significance. If the reported results hold, the framework would be a useful contribution to the growing literature on pre-trained, transferable architectures for collider physics, extending the CMS/ATLAS jet- and event-level studies to a threshold e+e- experiment. The paper’s strengths include a physically motivated benchmark, a consistent comparison of three training strategies across twelve channels, and large-scale pre-training. The architecture and fine-tuning recipe are clearly described. However, the central claim of transferability is currently supported only by simulation-based metrics; the absence of a real-data closure test and the absence of statistical uncertainties on the headline numbers make the quantitative claims provisional. The paper is publishable only after these points are addressed or the claims are appropriately scoped.

major comments (4)
  1. [§4.2 and §5] All reported performance numbers—the 97.0% background rejection, the 11/12 fine-tuning advantage in Table 1, the resolution improvements in Table 2 and Figure 7—are computed from MC simulation. The manuscript states that the simulation configuration is “consistent with that used in published BESIII physics analyses,” but this is not a validation of model transfer. Reconstruction-level mismatches in tracking, PID, calorimeter response, or background composition can make MC-only gains simulation-specific. Because the abstract and introduction frame the contribution as a transferable strategy “at BESIII,” a real-data closure test is needed, e.g., applying the fine-tuned classifier and regression model to a double-tagged Lambda_c+ Lambda_cbar control sample and comparing with MC expectations. Absent such a test, the transferability claim is unverified.
  2. [Tables 1–2 and Fig. 5] None of the performance comparisons carries a statistical uncertainty. Table 1 reports rejection rates at 90% signal efficiency to one decimal place; the differences used to conclude that fine-tuning wins in 11/12 channels are sometimes 0.1–0.3 percentage points (e.g., 88.3% vs 88.5% for p pi+ pi-, 85.8% vs 85.0% for p K- pi+ pi0) and may be within sampling noise. No bootstrap confidence intervals, repeated-seed training, or uncertainty propagation is provided for any AUC, rejection rate, or resolution. This is load-bearing for the central claim that fine-tuning improves performance, especially in the low-statistics regime emphasized in the abstract.
  3. [§4.2 and §5.1] The definition of the “direct application” of the pre-trained classification model to a specific decay channel is ambiguous. The pre-trained model’s output layer was trained to distinguish Lambda_c+ Lambda_cbar, D Dbar(h), and q qbar. Yet §4.2 states that the same event categories (signal, Lambda_c background, hadronic background) are used for fine-tuning, from-scratch, and direct application. It is not described how the pre-trained output is mapped onto channel-specific signal/background ROC curves. If the generic Lambda_c+ Lambda_cbar class is used as a proxy for “signal,” the comparison with fine-tuned and from-scratch models is not head-to-head. Please specify the exact score or class-probability transformation used for the pre-trained model in Table 1 and Figures 4–5.
  4. [§5.2] The baseline “conventional single-tag reconstruction” is not defined in sufficient detail for the reader to reproduce the quoted average resolutions of 0.325 rad and 0.626 rad. No algorithm, selection, or estimator description is given. Since the 31.1%/22.8% resolution improvement is a headline result, the baseline must be specified precisely or referenced to a concrete BESIII analysis procedure.
minor comments (5)
  1. [§5.2] Typographical/content error: “processes with partially detection of the final state particles” should read “processes with partial detection of the final-state particles.”
  2. [§3.1, Eq. (3.1)] The components of U_BES have mixed physical dimensions (dimensionless, GeV, and GeV^2). This may be intentional since the encoding is learned, but it should be stated explicitly to avoid confusion.
  3. [§4.2] “D+s” should be written as "D_s^+" for consistency with standard notation.
  4. [Fig. 2 caption] The panels are labeled only by “t-SNE 1” and “t-SNE 2”; it would help to state the epoch and dataset explicitly in the caption, and to indicate that the right panel corresponds to the best epoch by validation.
  5. [§5.2] The inclusive regression resolution (sigma(theta)=0.411 rad, sigma(phi)=0.858 rad) is much larger than the average over the 12 benchmark channels (0.224 and 0.484 rad). This is presumably because the inclusive sample contains a mixture of decay topologies, but the discrepancy is not explained; a sentence clarifying why would prevent a perceived inconsistency.

Circularity Check

0 steps flagged

No circularity found: an empirical pre-training/fine-tuning benchmark with independent baselines and held-out MC evaluation.

full rationale

The paper's central claims are empirical machine-learning results, not analytic derivations, and no load-bearing step reduces to its own inputs by construction. Pre-training uses inclusive MC samples with category or regression labels derived from simulation truth (Section 4.2); fine-tuning and from-scratch training use the same downstream datasets and splits, and performance is compared across three strategies (Section 5). The classification 'predictions' are category probabilities learned from labeled events, while the regression targets are MC-truth angles that are not part of the model input; evaluation measures residuals against those targets. This is a conventional supervised benchmark, so there is no equation-level identity or fitted-parameter-renamed-as-prediction. The improvements over from-scratch training and over conventional single-tag reconstruction are direct comparisons under a shared simulation framework, not self-referential definitions. References to the authors' previous BESIII analyses (e.g., [12,16,18,27,36]) are used for configuration, physics motivation, and context, but they are not invoked as uniqueness theorems or as substitutes for the present experiments; self-citations are therefore not load-bearing. The main weakness noted in the manuscript is that all reported numbers are obtained from Geant4/EvtGen Monte Carlo without a real-data closure test ('the MC simulation configuration is consistent with that used in published BESIII physics analyses', Section 4.2). That is an external-validity/generalization risk, not circularity: the simulation may misrepresent data, but the derivation does not assume what it claims to establish. Accordingly, the circularity score is 0.

Axiom & Free-Parameter Ledger

3 free parameters · 3 axioms · 0 invented entities

The central empirical claims rest on treating MC as data, on the sufficiency of the object-level point cloud, and on the hand-designed pairwise interaction matrix U_BES. Model hyperparameters and category balance are free design choices made without sensitivity analysis. No new physical entities are introduced.

free parameters (3)
  • Pre-training hyperparameters = batch size 512; initial LR 1e-3; epochs 26 (classification) and 78 (regression); dropout 0.1; fine-tuning LR multiplier
    All comparative results depend on these hand-set choices, and no ablations or sensitivity tests are reported (Section 4.3).
  • Architecture dimensions = 8 attention heads; 8 particle-attention layers; 3 class-attention layers; embedding widths 96/384/96; pairwise conv chan
    Chosen to increase capacity relative to previous BESIII models; the paper does not study how sensitive the transfer gains are to these sizes (Section 4.3).
  • Pre-training category balance and data split = balanced categories totaling ~7.1e7 events; regression pre-training ~2.4e7 events; train:val:test = 12:3:5
    Balancing categories and the split ratio are design choices that directly affect the reported 97.0% rejection and all fine-tuning comparisons (Section 4.2).
axioms (3)
  • domain assumption BESIII MC simulation samples are valid ground truth for physics and detector response
    All training, validation and test sets are MC events generated at sqrt(s)=4.600-4.700 GeV, with the paper assuming they match published BESIII analyses and hence real data (Section 4.2). No real-data validation is reported.
  • domain assumption Object-level point-cloud representation preserves enough information for transfer
    The model sees reconstructed tracks and clusters, not raw hits; the paper assumes this representation retains the information needed for classification and regression across channels (Section 3.1).
  • ad hoc to paper U_BES pairwise interaction features are an appropriate inductive bias
    The four hand-designed terms in Eq. (3.1) are introduced for BESIII's spherical geometry and energy-momentum conservation, but no derivation or ablation is given; the attention mechanism's success partly rests on this encoding.

pith-pipeline@v1.3.0-daily-deepseek · 12080 in / 13402 out tokens · 142637 ms · 2026-08-03T13:51:49.014099+00:00 · methodology

0 comments
read the original abstract

Deep learning has become an essential tool in high-energy physics, where the ability to learn transferable event representations can significantly improve model generalization across related physics processes. In this work, we present a Particle Transformer-based framework for learning such representations for charmed baryon physics in the BESIII experiment. The framework is implemented through large-scale pre-training on Monte Carlo simulation samples and subsequent fine-tuning for downstream analyses. Using the production and decays of the charmed baryon $\Lambda_c^+$ as a benchmark, we develop pre-trained models for both event classification and momentum-direction regression. The classification model learns discriminative event representations for the dominant physics categories, rejecting 97.0\% of background events at a signal efficiency of 90.0\%. Across 12 benchmark $\Lambda_c^+$ decay channels, fine-tuning from the pre-trained model achieves performance comparable or better than training from scratch, with particularly clear improvements in low-statistics regimes. For the regression task, the pre-trained model improves the momentum-direction prediction across the same benchmark channels. Further improvement is obtained after fine-tuning in the representative semileptonic decay $\Lambda_c^+ \to p K^- e^+ \nu_e$. This strategy provides a scalable solution for a wide range of physics cases at BESIII and can be extended to other high energy experiments.

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

42 extracted references · 21 linked inside Pith

  1. [1]

    Albertsson et al.,Machine Learning in High Energy Physics Community White Paper,J

    K. Albertsson et al.,Machine Learning in High Energy Physics Community White Paper,J. Phys. Conf. Ser.1085(2018) 022008 [1807.02876]

    Pith/arXiv arXiv 2018

  2. [2]

    Feickert and B

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

    Pith/arXiv arXiv

  3. [3]

    Goodfellow, Y

    I. Goodfellow, Y. Bengio and A. Courville,Deep Learning, MIT Press (2016)

    2016

  4. [4]

    Guest, K

    D. Guest, K. Cranmer and D. Whiteson,Deep Learning and its Application to LHC Physics, Ann. Rev. Nucl. Part. Sci.68(2018) 161 [1806.11484]

    Pith/arXiv arXiv 2018

  5. [5]

    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. [6]HEP Softw are Foundationcollaboration,A Roadmap for HEP Software and Computing R&D for the 2020s,Comput. Softw. Big Sci.3(2019) 7 [1712.06982]

    Pith/arXiv arXiv 2018

  6. [7]

    Bellis et al.,HEP Software Foundation Community White Paper Working Group – Visualization,1811.10309

    M. Bellis et al.,HEP Software Foundation Community White Paper Working Group – Visualization,1811.10309

    Pith/arXiv arXiv

  7. [8]

    Guest, J

    D. Guest, J. Collado, P. Baldi, S.-C. Hsu, G. Urban and D. Whiteson,Jet Flavor Classification in High-Energy Physics with Deep Neural Networks,Phys. Rev. D94(2016) 112002 [1607.08633]. [9]CMScollaboration,A Deep Neural Network for Simultaneous Estimation of b Jet Energy and Resolution,Comput. Softw. Big Sci.4(2020) 10 [1912.06046]

    Pith/arXiv arXiv 2016

  8. [10]

    Qian et al.,Vertex and energy reconstruction in JUNO with machine learning methods, Nucl

    Z. Qian et al.,Vertex and energy reconstruction in JUNO with machine learning methods, Nucl. Instrum. Meth. A1010(2021) 165527 [2101.04839]

    Pith/arXiv arXiv 2021

  9. [11]

    X. Jia, X. Qin, T. Li, X. Huang, X. Zhang, N. Yin et al.,BESIII track reconstruction algorithm based on machine learning,EPJ Web Conf.295(2024) 09006. [12]BESIIIcollaboration,Observation of a rare beta decay of the charmed baryon with a Graph Neural Network,Nature Commun.16(2025) 681 [2410.13515]

    arXiv 2024

  10. [13]

    H. Qu, C. Li and S. Qian,Particle Transformer for Jet Tagging, inProceedings of the 39th International Conference on Machine Learning, vol. 162 ofProceedings of Machine Learning Research, pp. 18281–18292, PMLR, 2022, https://proceedings.mlr.press/v162/qu22b.html [2202.03772]

    Pith/arXiv arXiv 2022

  11. [14]

    Y. Wu, K. Wang, C. Li, H. Qu and J. Zhu,Jet tagging with more-interaction particle transformer*,Chin. Phys. C49(2025) 013110 [2407.08682]. – 14 –

    Pith/arXiv arXiv 2025

  12. [15]

    Weaver: a streamlined machine-learning framework for high-energy physics

    H. Qu et al., “Weaver: a streamlined machine-learning framework for high-energy physics.” GitHub repository,https://github.com/hqucms/weaver-core(accessed 15 March 2026), 2026. [16]BESIIIcollaboration,Search for radiative leptonic decayD + →γe +νe using deep learning, Chin. Phys. C49(2025) 083001. [17]BESIIIcollaboration,Measurement of the singly Cabibbo-...

    arXiv 2026

  13. [19]

    Improving Language Understanding by Generative Pre-Training

    A. Radford, K. Narasimhan, T. Salimans and I. Sutskever, “Improving Language Understanding by Generative Pre-Training.” OpenAI technical report, https://cdn.openai.com/research-covers/language-unsupervised/language_ understanding_paper.pdf, 2018

    2018

  14. [20]

    OpenAI,OpenAI GPT-5 System Card,2601.03267

    Pith/arXiv arXiv

  15. [21]

    Rives, J

    A. Rives, J. Meier, T. Sercu, S. Goyal, Z. Lin, J. Liu et al.,Biological structure and function emerge from scaling unsupervised learning to 250 million protein sequences,Proc. Natl. Acad. Sci. U.S.A.118(2021) e2016239118

    2021

  16. [22]

    Z. Lin, H. Akin, R. Rao, B. Hie, Z. Zhu, W. Lu et al.,Evolutionary-scale prediction of atomic-level protein structure with a language model,Science379(2023) 1123. [23]CMScollaboration,The CMS Experiment at the CERN LHC,JINST3(2008) S08004. [24]ATLAScollaboration,The ATLAS Experiment at the CERN Large Hadron Collider,JINST 3(2008) S08003

    2023

  17. [25]

    Z. Zhao, F. Mokhtar, R. Kansal, H. Li and J. Duarte,Large-Scale Pretraining and Finetuning for Efficient Jet Classification in Particle Physics, in22nd International Workshop on Advanced Computing and Analysis Techniques in Physics Research: Foundation Models for Physics - Nexus of Computation and Physics through Embracing the Era of Foundation Models, Ma...

    Pith/arXiv arXiv 2024

  18. [26]

    J. Ho, B.R. Roberts, S. Han and H. Wang,Pretrained Event Classification Model for High Energy Physics Analysis,2412.10665

    Pith/arXiv arXiv

  19. [27]

    Li and X.-R

    H.-B. Li and X.-R. Lyu,Study of the standard model with weak decays of charmed hadrons at BESIII,Natl. Sci. Rev.8(2021) nwab181 [2103.00908]. [28]BESIIIcollaboration,Design and Construction of the BESIII Detector,Nucl. Instrum. Meth. A614(2010) 345 [0911.4960]

    Pith/arXiv arXiv 2021

  20. [29]

    Huang, Z.-J

    K.-X. Huang, Z.-J. Li, Z. Qian, J. Zhu, H.-Y. Li, Y.-M. Zhang et al.,Method for detector description transformation to Unity and application in BESIII,Nucl. Sci. Tech.33(2022) 142 [2206.10117]

    Pith/arXiv arXiv 2022

  21. [30]

    Yu et al.,BEPCII Performance and Beam Dynamics Studies on Luminosity, in Proceedings of the 7th International Particle Accelerator Conference (IPAC’16), no

    C. Yu et al.,BEPCII Performance and Beam Dynamics Studies on Luminosity, in Proceedings of the 7th International Particle Accelerator Conference (IPAC’16), no. 7 in International Particle Accelerator Conference, (Busan, Korea), pp. 1014–1018, JACoW Publishing, 2016, DOI. [31]BESIIIcollaboration,Future Physics Programme of BESIII,Chin. Phys. C44(2020) 0400...

    arXiv 2016

  22. [32]

    J.D. Lu, Y.J. Xiao and X.B. Ji,Online monitoring of the center-of-mass energy from real data at BESIII,Radiat. Detect. Technol. Methods4(2020) 337

    2020

  23. [33]

    Zhang et al.,Suppression of top-up injection backgrounds with offline event filter in the BESIII experiment,Radiat

    J.W. Zhang et al.,Suppression of top-up injection backgrounds with offline event filter in the BESIII experiment,Radiat. Detect. Technol. Methods6(2022) 289

    2022

  24. [34]

    Liao, J.-S

    M.-H. Liao, J.-S. Liu, X.-N. Wang, S.-S. Sun and Z.-Y. You,Experimental dataset from BESIII detector at Beijing electron–positron collider,Nucl. Sci. Tech.36(2025) 218 [2509.16066]

    arXiv 2025

  25. [35]

    Asner et al.,Physics at BES-III,Int

    D.M. Asner et al.,Physics at BES-III,Int. J. Mod. Phys. A24(2009) S1 [0809.1869]

    Pith/arXiv arXiv 2009

  26. [36]

    Li, X.-R

    P.-R. Li, X.-R. Lyu and Y. Zheng,Experimental overview on the charmed baryon decays, Chin. Phys. C50(2026) 022002 [2509.19141]

    arXiv 2026

  27. [37]

    Wang et al.,Transfer learning empowers material Z classification with muon tomography,Nucl

    H.-C. Wang et al.,Transfer learning empowers material Z classification with muon tomography,Nucl. Sci. Tech.37(2026) 77 [2504.12305]

    Pith/arXiv arXiv 2026

  28. [38]

    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

  29. [39]

    J. Birk, A. Hallin and G. Kasieczka,OmniJet-α: the first cross-task foundation model for particle physics,Mach. Learn. Sci. Tech.5(2024) 035031 [2403.05618]

    Pith/arXiv arXiv 2024

  30. [40]

    Pan and Q

    S.J. Pan and Q. Yang,A Survey on Transfer Learning,IEEE Trans. Knowledge Data Eng. 22(2010) 1345. [41]Particle Data Groupcollaboration,Review of Particle Physics,Phys. Rev. D110(2024) 030001. [42]MARK-IIIcollaboration,Direct Measurements of Charmed-D-Meson Hadronic Branching Fractions,Phys. Rev. Lett.56(1986) 2140

    2010

  31. [43]

    Agostinelli et al.,Geant4—a simulation toolkit,Nucl

    S. Agostinelli et al.,Geant4—a simulation toolkit,Nucl. Instrum. Meth. A506(2003) 250

    2003

  32. [44]

    Lange,The EvtGen particle decay simulation package,Nucl

    D.J. Lange,The EvtGen particle decay simulation package,Nucl. Instrum. Meth. A462 (2001) 152

    2001

  33. [45]

    Ping,Event generators at BESIII,Chin

    R.G. Ping,Event generators at BESIII,Chin. Phys. C32(2008) 599

    2008

  34. [46]

    Chen, G.S

    J.C. Chen, G.S. Huang, X.R. Qi, D.H. Zhang and Y.S. Zhu,Event generator forJ/ψand ψ(2S)decay,Phys. Rev. D62(2000) 034003

    2000

  35. [47]

    Yang, R.G

    R.L. Yang, R.G. Ping and H. Chen,Tuning and Validation of the Lundcharm Model with J/ψDecays,Chin. Phys. Lett.31(2014) 061301

    2014

  36. [48]

    Zou et al.,Offline data processing system of the BESIII experiment,Eur

    J. Zou et al.,Offline data processing system of the BESIII experiment,Eur. Phys. J. C84 (2024) 937. [49]BESIIIcollaboration,Measurements of the absolute branching fractions of theΛ + c hadronic decays,JHEP06(2026) 269 [2601.01503]. [50]BESIIIcollaboration,First Observation of the Semileptonic DecayΛ + c →pK −e+νe,Phys. Rev. D106(2022) 112010 [2207.11483]

    arXiv 2024

  37. [51]

    van der Maaten and G

    L. van der Maaten and G. Hinton,Visualizing data using t-SNE,J. Mach. Learn. Res.9 (2008) 2579

    2008

  38. [52]

    Bardhan, R

    J. Bardhan, R. Agrawal, A. Tilak, C. Neeraj and S. Mitra,HEP-JEPA: A foundation model for collider physics using joint embedding predictive architecture,2502.03933. – 16 –

    Pith/arXiv arXiv

  39. [53]

    M. Chen, R. Kansal, A. Gandrakota, Z. Hao, J. Ngadiuba and M. Spiropulu,An Evaluation of Representation Learning Methods in Particle Physics Foundation Models,2511.12829

    arXiv

  40. [54]

    Ai et al.,Conceptual design report of the Super Tau-Charm Facility: the accelerator, Nucl

    X.-C. Ai et al.,Conceptual design report of the Super Tau-Charm Facility: the accelerator, Nucl. Sci. Tech.36(2025) 242 [2509.11522]

    arXiv 2025

  41. [55]

    Cheng, X.-R

    H.-Y. Cheng, X.-R. Lyu and Z.-Z. Xing,Charm Physics at the Superτ-Charm Factory, Chin. Phys. Lett.42(2025) 010201 [2203.03211]

    Pith/arXiv arXiv 2025

  42. [56]

    The CEPC Study Group,CEPC Conceptual Design Report: Volume 2—Physics & Detector, 1811.10545. – 17 –

    Pith/arXiv arXiv