Pith. sign in

REVIEW 3 major objections 4 minor 35 references

An 82k-parameter student jet tagger can match and beat two multi-million-parameter teachers by explicitly modeling when to trust them.

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

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

T0 review · deepseek-v4-flash

2026-08-01 23:02 UTC pith:QI2R3DLE

load-bearing objection A genuinely new reliability-coupled distillation scheme for jet taggers, with honest ablations and code; the 'exceeds teacher' claim is plausible but needs a plain-supervised baseline and uncertainty on teacher rows. the 3 major comments →

arxiv 2607.15531 v2 pith:QI2R3DLE submitted 2026-07-17 hep-ph

JetCoRD: Reliability-Aware Cross-Experiment Distillation of Jet Taggers with Adaptive Corrective Representation

Liu-Long Gao , Zheng-Kun Huang , Xiao-Wei Jiang , Jia-Feng Li , Gong-Xing Sun This is my paper
classification hep-ph
keywords jet taggingknowledge distillationreliability estimationmodel compressioncross-experiment transferteacher-student learninghigh-energy physicsparticle transformer
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 prove that knowledge distillation of jet taggers works even when the teacher is imperfect, provided the student learns a per-sample reliability signal that controls both training and inference. The authors distill two large, detector-specific teachers (ATLAS GN2 and CMS ParT, together 7M parameters) into a single 82k-parameter student, which matches both teachers on overall accuracy and exceeds them at four physics-actionable working points. The core idea is that the same reliability estimate that down-weights the distillation loss on teacher mistakes also anchors a learned gate that decides, per jet, whether to trust the teacher's logits or the student's own prediction. If true, this suggests that state-of-the-art tagging capability can be compressed to about 1% of the parameter budget and shared across experiments, while treating teacher errors as information rather than ground truth.

Core claim

The central claim is that a single per-sample reliability signal r_i can simultaneously carry three jobs: reweight the KL distillation loss so teacher mistakes contribute almost nothing, drive a prototype-based repair loss that pulls the student's representation toward the true class when the teacher is wrong, and anchor an inference-time gate g_i that mixes teacher and student logits via s_final = g_i * t_i + (1 - g_i) * s_student. The gate is trained to reproduce r_i through a BCE anchor loss, so the same notion of "when to trust the teacher" applies at train and test time. The authors show that this coupling allows an 82k-parameter student to exceed the teachers at b-vs-c (ε=0.77, +4.3%),

What carries the argument

The central object is the per-sample reliability signal r_i = 1{y_hat_T = y_i} · σ(κ · (top1(t_i) − top2(t_i))), which is nonzero only when the teacher's prediction agrees with the ground-truth label and scales with the teacher's logit margin. This scalar is used in three coupled mechanisms: (1) it reweights the temperature-scaled KL distillation loss; (2) it gates an EMA-prototype repair loss that pulls the student embedding toward the correct class prototype when the teacher is wrong; and (3) it anchors the learned gate g_i = σ(MLP(z_i)) through a BCE loss, so that the final prediction is a convex mixture of teacher and student logits. The repair strength per class is controlled by learnab

Load-bearing premise

The entire correction mechanism assumes that the teacher's 128-dimensional embedding and logits retain enough information to identify and fix the teacher's own mistakes; if a teacher error is caused by information that was discarded before the embedding, neither the repair loss nor the gate can recover the true label.

What would settle it

A concrete test would be to construct a test set where teacher errors are dominated by a known information-loss mechanism—for example, by zeroing out the embedding components that carry secondary-vertex topology—and then check whether the student's improvement over the teacher at b-vs-c and c-vs-u disappears. If the gate's correlation with reliability (currently 0.49–0.59) drops to near zero on such a set, the central claim that reliability-based repair is the cause of the gains would be falsified.

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

If this is right

  • A tiny 82k-parameter model can deliver teacher-level or better performance, making it feasible to deploy state-of-the-art jet tagging in low-latency or resource-limited settings.
  • A single student can serve two different detector environments, potentially removing the need for each experiment to train and maintain its own tagger.
  • Reliability-weighted distillation stops the student from inheriting the teacher's systematic mistakes, which is a known failure mode of vanilla knowledge distillation.
  • The learned repair budget is interpretable and sparse: the student automatically concentrates its corrective effort on classes where the teacher is statistically weak, which could guide further model development.
  • The gate mechanism provides a principled way to combine any imperfect teacher with a student, extending beyond jet tagging to other high-energy physics classification tasks.

Where Pith is reading between the lines

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

  • The student's ability to repair teacher errors is bounded by the information content of the 128-dimensional teacher embedding; if a teacher mistake is caused by a feature that is discarded before the embedding (e.g., secondary-vertex topology), the gate and repair loss cannot recover the true label.
  • The measured correlation between the inference gate and the training-time reliability is only 0.49–0.59, suggesting that the gate has room to better identify unreliable samples; a stronger anchor or a richer gate input could yield further gains.
  • The small but reproducible regression at extreme working points (ε ≥ 0.99) might be addressed by the saturation mask the paper suggests; this is a concrete, testable extension.
  • The design points toward a theoretical question: under bounded teacher calibration error, is a reliability-anchored gate asymptotically Bayes-optimal? Proving this would elevate the approach from an empirical recipe to a principled framework.

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

3 major / 4 minor

Summary. The paper proposes JetCoRD, an 82k-parameter student that jointly distills two pre-trained jet taggers — ATLAS GN2 (3 classes) and CMS ParT (10 classes). The student consumes the teacher embeddings and logits as inputs and is trained with a reliability-weighted distillation loss (A-CoRD) plus a reliability-gated inference-time convex mixture of teacher and student logits (RAI). The authors report that the student matches both teachers in overall accuracy and exceeds them at selected working points (e.g., +4.3% on b-vs-c at ε=0.77, +1.5% on c-vs-u at ε=0.30, +1.6% on T→bqq at ε=0.5, +1.4% on H→bb at ε=0.5). The contribution is framed as the first cross-experiment distillation in HEP and a new reliability-coupled training/inference framework.

Significance. If the central claims hold, the work is significant: it would demonstrate that a very small unified student can exploit teacher reliability signals to improve upon heavyweight teachers at physics-relevant working points, and it introduces a training/inference coupling (A-CoRD/RAI) that appears novel in HEP distillation. The paper has concrete strengths: it releases code and trained checkpoints, adopts a 10-seed protocol, and includes architecture and algorithm ablations that isolate the contributions of RAI and CCB. However, the statistical evidence for 'exceeds teacher' is currently weakened by the absence of a plain supervised baseline on the same teacher features, by selection of the headline working points from 19 without multiple-comparison correction, and by an implausible zero-standard-deviation entry. These issues are fixable and do not invalidate the overall idea, but they currently prevent firm acceptance.

major comments (3)
  1. [§3.1, §6.2] The paper asserts 'a simple classifier on teacher embeddings can at best match the teacher, never surpass it' (Sec. 3.1), but this assertion is unproven and load-bearing. The teacher's classification head is a simple function on e_T, while the student is a nonlinear MLP (two ResMLP blocks) operating on the same 128-dim embedding plus logits and domain token. A nonlinear supervised classifier on these features could plausibly already beat the teacher's head, independent of any reliability mechanism. Table 4 ablations compare only KD, CoRD, A-CoRD, and PCG — all of which contain the RAI mixture — but no plain CE-trained student on (e_T, t) with the same backbone is reported. The attribution of the reported gains to A-CoRD/RAI therefore lacks a necessary control.
  2. [Table 1, §5.2] The four headline working points (+4.3%, +1.5%, +1.6%, +1.4%) are selected from the 19 rows in Table 1, and no multiple-comparison correction is applied. Since these working points are positively correlated, the expected maximum improvement under a null hypothesis is non-negligible. In addition, the Tbqq@0.5 row reports 32786.9±0.0 over n=10 seeds, which is inconsistent with any seed variation and likely reflects a reporting artifact; this row is one of the four headline results. The teacher rows are shown without uncertainty (the teacher is deterministic, but student seed variation should be propagated to the Δ% columns). The 'exceeds teacher' claim needs a multi-seed, multiple-comparison-aware analysis.
  3. [Eq. (3.8), §3.3.2] The final logits are a convex mixture s_final = g·t + (1−g)·s_student. Consequently, if the gate g is close to 1, the student output is nearly identical to the teacher's, so the reported 'matches teacher in overall accuracy' is at least partly by construction rather than evidence of learned compression. The non-trivial part is the exceedance at selected WPs, which requires g<1 on relevant samples. The paper should report the distribution of g over the full test set (e.g., the fraction of jets with g>0.9, per domain) and demonstrate that the WP gains are not driven by a minuscule subset of jets where the gate happens to defer to the student. Currently Table 5 gives only conditional means, which is insufficient to rule out this explanation.
minor comments (4)
  1. [§8.1 vs Abstract] The conclusion says the student exceeds the teachers at 'three physics-actionable working points', but the abstract and Table 1 list four (including T→bqq). Please reconcile the count.
  2. [Eq. (3.4), Table 6] Eq. (3.4) uses acc_c^T, which must be a fraction (0–1), while Table 6 reports teacher accuracy as a percentage (e.g., 76.8%). The notation should explicitly state the normalization to avoid confusion.
  3. [Table 1] The abbreviations 'Hqql' and 'Tbl' in Table 1 are not introduced; use the full names H→q¯qℓ and t→bℓν, as elsewhere in the text.
  4. [§4.6, Table 5] Table 5 reports quantities (mean reliability, mean gate, correlations) without error bars. Since these are computed on the test set from one checkpoint (or averaged over 10 seeds?), please clarify whether these are single-run or seed-averaged, and include standard deviations if averaged.

Circularity Check

2 steps flagged

Teacher-matching is definitionally available from Eq. 3.8; gate-reliability correlation is partly implanted by Eq. 3.9. The selected-WP exceedances remain empirical and not circular.

specific steps
  1. self definitional [Sec. 3.3.2 (Eq. 3.8); Sec. 5.1 (Table 1)]
    "The final logits are a convex mixture: s_final_i = g_i · t_i + (1−g_i) · s_student_i. ... To prevent the trivial g_i ≡1 solution that would collapse the student, we introduce two auxiliary losses."

    Since g_i ∈ [0,1] and teacher logits t_i enter s_final linearly, the choice g_i = 1 makes the student output exactly the teacher logits. The paper explicitly names this the 'trivial solution'. Therefore the reported 'matches both teachers in overall accuracy' (e.g. KD 73.932±0.001 vs teacher 73.932; PCG 86.064 vs 86.051) is available by construction, not evidence that the 82k-parameter student independently learned the teacher's function. The non-trivial exceedances at selected working points are empirical, but the headline 'matches' clause is definitional.

  2. fitted input called prediction [Sec. 3.2.1 (Eq. 3.1); Sec. 3.3.2 (Eq. 3.9); Sec. 7.1 (Table 5)]
    "r_i = 1{ŷT_i = y_i} · σ(κ·m_i) ... L_gate = BCE(logit(g_i), r_i), w_gate = 0.03, which ties the inference-time gate to the A-CoRD training-time reliability r_i. ... The gate has learned to not trust the teacher when the teacher is wrong."

    By Eq. 3.1, r_i is zero exactly on teacher-wrong samples. Eq. 3.9 trains g_i with BCE target r_i, so the training objective directly pushes g_i down on those same samples. The Sec. 7.1 observations that ⟨g|wrong⟩ < ⟨g|correct⟩ (0.549 vs 0.744 ATLAS) and corr(g,r)=0.49–0.59 are therefore the supervised target generalizing to the test set, not an emergent 'causal test' of the architecture-algorithm coupling. The behavior is implanted via the gate-anchor loss; presenting it as independent evidence of the reliability coupling is a self-consistency check, not a discovery.

full rationale

JetCoRD's central exceedance claims are not circular: the student is trained on external ATLAS Open Tagging and JetClass labels and evaluated against the actual GN2/ParT checkpoints on held-out test splits, so the +4.3% b-vs-c, +1.5% c-vs-u, +1.6% T→bqq, and +1.4% H→bb numbers are empirical, externally benchmarked results. The A-CoRD loss (Eqs. 3.1–3.5) is not derived from the target working points, and no load-bearing self-citation chain exists. However, two parts of the presentation reduce to construction. First, Eq. 3.8 defines final logits as a convex mixture of teacher and student logits and explicitly admits g=1 as the 'trivial solution,' so the abstract's 'student matches both teachers in overall accuracy' is a built-in property, not a learned compression result. Second, the gate-reliability correlation used as evidence in Sec. 7.1 is partly implanted: the gate is trained by BCE regression to r_i, whose definition already encodes teacher correctness. The paper also asserts without a plain-CE baseline that 'a simple classifier on teacher embeddings can at best match the teacher, never surpass it'; that is a load-bearing correctness/ablation gap rather than circularity. Overall, the central 'exceeds teacher' claim has independent empirical content, but the 'matches teacher' claim and the gate-coupling demonstration are partially circular by construction, giving score 4.0.

Axiom & Free-Parameter Ledger

7 free parameters · 5 axioms · 0 invented entities

The central claim rests on one strong information assumption (teacher embeddings contain enough signal to repair teacher errors), one architectural assumption (a shared backbone can bridge two experiments), and several hand-tuned hyperparameters selected on a 16-config random search. The paper introduces no new physical entities; the 'reliability signal' and 'prototypes' are training constructs, not physics objects.

free parameters (7)
  • alpha_c (13 per-class repair offsets: ATLAS 3 + JetClass 10) = u in ATLAS +9.06, H->4q +10.27, others ~ -10
    Learned scalars in beta_c = beta0*sigmoid(alpha_c - s*acc_T^c) (Eq. 3.4). They decide which classes receive prototype repair; only u and H->4q are non-zero. Fitted to training data.
  • beta0 = 0.5
    Base repair-loss scale, fixed by random search (Sec. 4.4).
  • s = 5.0
    Accuracy-contraction slope in beta_c; chosen by random search.
  • kappa = 5.0
    Sharpness of the soft margin gate in reliability r_i (Eq. 3.1); chosen by random search.
  • w_gate = 0.03
    Weight of BCE gate-anchor loss; the most sensitive hyperparameter, selected from a 16-config random search (Sec. 6.3).
  • T_KD = 4.0
    Distillation temperature; random-search value.
  • w_stu = 0.5
    Auxiliary student-CE loss weight; default.
axioms (5)
  • domain assumption Teacher embeddings e_T and logits t carry enough information for the student to correct the teacher's own mistakes (student never sees raw particle inputs).
    Sec. 3.1 states the student maps (e_T, domain, t) -> s_i; if mistakes come from information discarded before the 128-dim embedding, the claimed repair cannot work.
  • domain assumption A shared backbone with per-domain tokens/heads can bridge disjoint detector responses and disjoint class spaces (ATLAS 3 classes vs JetClass 10 classes).
    Sec. 3.3 adds zero-initialized domain tokens and per-domain CCB embeddings; the paper assumes no explicit detector translation is needed.
  • domain assumption The convex mixture of teacher and student logits (Eq. 3.8) yields a valid score for the one-vs-one discriminant D = P_s/(P_s+P_b).
    Used in all evaluations (Sec. 4.5); mixing logits before softmax is a design choice, not a theorem.
  • domain assumption Ground-truth labels are available to compute r_i and to train the gate anchor; the gate will predict r_i from features at inference.
    r_i uses y_i in Eq. 3.1; at inference y_i is absent, and transfer is only approximate (corr(g,r)=0.49-0.59, Sec. 7.1).
  • ad hoc to paper Per-class EMA prototypes in the student's 64-dim embedding are stable class centers suitable for representation repair.
    Eq. 3.3 pulls z_i toward mu_{y_i}; there is no guarantee this repair is geometrically meaningful in the learned space.

pith-pipeline@v1.3.0-alltime-deepseek · 13929 in / 15008 out tokens · 160679 ms · 2026-08-01T23:02:23.966674+00:00 · methodology

0 comments
read the original abstract

Modern jet taggers based on graph and transformer networks deliver state-of-the-art performance but are expensive to train and difficult to share across experiments. Knowledge distillation can in principle achieve model compression, but it rests on the assumption that the teacher model acts as a perfect tagger. This assumption fails to hold at high-purity working points, where the teacher itself exhibits a jet prediction error rate of approximately $10\%$ to $30\%$. We introduce JetCoRD, the first cross-experiment distillation in HEP: an 82 k-parameter unified student distilling jointly from ATLAS GN2 (5 M params, 3 classes) and CMS ParT (2 M params, 10 classes). The central innovation is a single per-sample reliability signal $r_i$ that simultaneously weights the distillation loss, controls prototype-based teacher repair, and anchors the inference-time gate $g_i$ that mixes teacher and student logits. With only 82 k trainable parameters (1.2% of the combined 7 M teacher parameters), the student matches both teachers in overall accuracy and exceeds them at physics-actionable working points: $+4.3\%$ on $b$-vs-$c$ at $\varepsilon=0.77$, $+1.5\%$ on $c$-vs-$u$ at $\varepsilon=0.30$, $+1.6\%$ on $T\!\to\!bqq$ at $\varepsilon=0.5$ and $+1.4\%$ on $H\!\to\!bb$ at $\varepsilon=0.5$. The reliability-coupled design is novel in HEP distillation and applicable wherever an imperfectly calibrated teacher must be compressed.

Figures

Figures reproduced from arXiv: 2607.15531 by Gong-Xing Sun, Jia-Feng Li, Liu-Long Gao, Xiao-Wei Jiang, Zheng-Kun Huang.

Figure 2
Figure 2. Figure 2: KDE of the inference gate gi conditioned on teacher correctness, for ATLAS (left) and JetClass (right). – 16 – [PITH_FULL_IMAGE:figures/full_fig_p018_2.png] view at source ↗

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

35 extracted references · 2 linked inside Pith

  1. [1]

    The ATLAS Collaboration,Transforming jet flavour tagging at ATLAS, Nat Commun 17, 541 (2026)

    2026

  2. [2]

    H. Qu, C. Li, S. Qian,Particle Transformer for Jet Tagging, ICML (2022)

    2022

  3. [3]

    Hinton, O

    G. Hinton, O. Vinyals, J. Dean,Distilling the Knowledge in a Neural Network, arXiv:1503.02531

    Pith/arXiv arXiv

  4. [4]

    H. Qu, L. Gouskos,Jet Tagging via Particle Clouds, Phys. Rev. D 101, 056019 (2020)

    2020

  5. [5]

    Mikuni, F

    V. Mikuni, F. Canelli,ABCNet: an attention-based method for particle tagging, Eur. Phys. J. Plus 135, 463 (2020)

    2020

  6. [6]

    Gong et al.,An efficient Lorentz equivariant graph neural network for jet tagging, JHEP 07 (2022) 030

    S. Gong et al.,An efficient Lorentz equivariant graph neural network for jet tagging, JHEP 07 (2022) 030

    2022

  7. [7]

    S. I. Mirzadeh et al.,Improved Knowledge Distillation via Teacher Assistant, AAAI (2020)

    2020

  8. [8]

    Guo et al.,Online Knowledge Distillation via Collaborative Learning, CVPR (2020)

    Q. Guo et al.,Online Knowledge Distillation via Collaborative Learning, CVPR (2020)

    2020

  9. [9]

    Park et al.,Relational Knowledge Distillation, CVPR (2019)

    W. Park et al.,Relational Knowledge Distillation, CVPR (2019)

    2019

  10. [10]

    Shazeer et al.,Outrageously Large Neural Networks: The Sparsely-Gated Mixture-of-Experts Layer, ICLR (2017)

    N. Shazeer et al.,Outrageously Large Neural Networks: The Sparsely-Gated Mixture-of-Experts Layer, ICLR (2017)

    2017

  11. [11]

    Fedus et al.,Switch Transformers: Scaling to Trillion Parameter Models with Simple and Efficient Sparsity, JMLR (2022)

    W. Fedus et al.,Switch Transformers: Scaling to Trillion Parameter Models with Simple and Efficient Sparsity, JMLR (2022)

    2022

  12. [12]

    Touvron et al.,Going Deeper with Image Transformers, ICCV (2021)

    H. Touvron et al.,Going Deeper with Image Transformers, ICCV (2021). – 18 –

    2021

  13. [13]

    A. K. Menon et al.,A Statistical Perspective on Distillation, ICML (2021)

    2021

  14. [14]

    Stanton et al.,Does Knowledge Distillation Really Work?, NeurIPS (2021)

    S. Stanton et al.,Does Knowledge Distillation Really Work?, NeurIPS (2021)

    2021

  15. [15]

    Duarte et al.,Fast inference of deep neural networks in FPGAs for particle physics, JINST 13 (2018) P07027

    J. Duarte et al.,Fast inference of deep neural networks in FPGAs for particle physics, JINST 13 (2018) P07027

    2018

  16. [16]

    Furlanello et al.,Born-Again Neural Networks, ICML (2018)

    T. Furlanello et al.,Born-Again Neural Networks, ICML (2018)

    2018

  17. [17]

    Cranmer, J

    K. Cranmer, J. Brehmer, G. Louppe,The frontier of simulation-based inference, PNAS 117 (2020) 30055

    2020

  18. [18]

    Andreassen et al.,OmniFold: A Method to Simultaneously Unfold All Observables, Phys

    A. Andreassen et al.,OmniFold: A Method to Simultaneously Unfold All Observables, Phys. Rev. Lett. 124 (2020) 182001

    2020

  19. [19]

    Mikuni, B

    V. Mikuni, B. Nachman,Method to simultaneously facilitate all jet physics tasks, Phys. Rev. D 111 (2025) 054015

    2025

  20. [20]

    Spinner et al.,Lorentz-Equivariant Geometric Algebra Transformers for High-Energy Physics, NeurIPS (2024)

    J. Spinner et al.,Lorentz-Equivariant Geometric Algebra Transformers for High-Energy Physics, NeurIPS (2024)

    2024

  21. [21]

    Wu et al.,Jet Tagging with More-Interaction Particle Transformer, Chinese Phys

    Y. Wu et al.,Jet Tagging with More-Interaction Particle Transformer, Chinese Phys. C 49 (2025) 013110

    2025

  22. [22]

    M. Y. Chen et al.,Accelerating Resonance Searches via Signature-Oriented Pre-training, arXiv:2405.12972

    Pith/arXiv arXiv

  23. [23]

    J. Birk, A. Hallin, G. Kasieczka,OmniJet-α: The first cross-task foundation model for particle physics, MLST 5 (2024) 035031

    2024

  24. [24]

    N. I. Pond,Top quark pair events for heavy flavour tagging and vertexing at the LHC, Zenodo (2023)

    2023

  25. [25]

    Zhao et al.,Decoupled Knowledge Distillation, CVPR (2022)

    B. Zhao et al.,Decoupled Knowledge Distillation, CVPR (2022)

    2022

  26. [26]

    Sun et al.,Logit Standardization in Knowledge Distillation, CVPR (2024)

    S. Sun et al.,Logit Standardization in Knowledge Distillation, CVPR (2024)

    2024

  27. [27]

    Fan et al.,ScaleKD: Strong Vision Transformers Could Be Excellent Teachers, NeurIPS (2024)

    J. Fan et al.,ScaleKD: Strong Vision Transformers Could Be Excellent Teachers, NeurIPS (2024)

    2024

  28. [28]

    Hao et al.,VanillaKD: Revisit the Power of Vanilla Knowledge Distillation from Small Scale to Large Scale, NeurIPS (2023)

    Z. Hao et al.,VanillaKD: Revisit the Power of Vanilla Knowledge Distillation from Small Scale to Large Scale, NeurIPS (2023)

    2023

  29. [29]

    Huang et al.,Knowledge Distillation from A Stronger Teacher, NeurIPS (2022)

    T. Huang et al.,Knowledge Distillation from A Stronger Teacher, NeurIPS (2022)

    2022

  30. [30]

    Wei et al.,Scale Decoupled Distillation, CVPR (2024)

    S. Wei et al.,Scale Decoupled Distillation, CVPR (2024)

    2024

  31. [31]

    Li et al.,Curriculum Temperature for Knowledge Distillation, AAAI (2023)

    Z. Li et al.,Curriculum Temperature for Knowledge Distillation, AAAI (2023)

    2023

  32. [32]

    Zheng, E.-H

    K. Zheng, E.-H. Yang,Knowledge Distillation Based on Transformed Teacher Matching, ICLR (2024)

    2024

  33. [33]

    B. Zhao, R. Song, J. Liang,DOT: A Distillation-Oriented Trainer, ICCV (2023)

    2023

  34. [34]

    Golling et al.,Masked Particle Modeling on Sets: Towards Self-Supervised High Energy Physics Foundation Models, MLST 5 (2024) 035074

    T. Golling et al.,Masked Particle Modeling on Sets: Towards Self-Supervised High Energy Physics Foundation Models, MLST 5 (2024) 035074

    2024

  35. [35]

    Bogatskiy et al.,Explainable equivariant neural networks for particle physics: PELICAN, JHEP 03 (2024) 113

    A. Bogatskiy et al.,Explainable equivariant neural networks for particle physics: PELICAN, JHEP 03 (2024) 113. – 19 –

    2024