Physics and Computing Performance of the EggNet Tracking Pipeline

Brandon Wang; Jay Chan; Paolo Calafiura

REVIEW 4 major objections 6 minor 19 references

Physics and Computing Performance of the EggNet Tracking Pipeline

T0 review · 4 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read This paper reports that EggNet, trained on segmented azimuthal windows instead of full events, matches full-graph track efficiency (about 0.96) on TrackML while cutting training time and GPU memory by roughly a factor of ten.

desk verdict A credible, incremental benchmark of the authors' own EggNet on full TrackML with a compute-saving training trick; the 10x savings and 'minimal impact' claim need an ablation on the window size and a better look at the duplication rate. read the letter →

arxiv 2506.03415 v1 pith:HDF3MSNG submitted 2025-06-03 physics.data-an hep-exhep-ph

Jay Chan , Brandon Wang , Paolo Calafiura This is my paper

classification physics.data-anhep-exhep-ph

keywords particletrackreconstructiongraphneuralnetworksEggNetMLHL-LHCsegmentedtrainingcontrastivelossDBSCAN

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

EggNet is a one-shot graph neural network for particle track reconstruction that builds its graph iteratively from a point cloud of detector spacepoints instead of requiring a separately constructed input graph. This paper establishes that the pipeline can be trained on the full TrackML dataset—about 108,000 spacepoints and 9,000 particles per event—and reaches a track efficiency of $0.9587\pm0.0003$ with full-graph training and $0.9638\pm0.0003$ with segmented training, with fake rates near 0.002. The new computational result is segmented graph training: random azimuthal windows, initially $\Delta\phi=\pi/5$, are adaptively grown or shrunk, cutting training time and GPU memory by roughly a factor of ten while leaving physics performance nearly unchanged. If the comparison holds, learned tracking becomes a practical option at HL-LHC event sizes rather than a small-dataset demonstration.

What carries the argument

The mechanism that carries the physics is the EggNet architecture: node features are encoded, then for five iterations the model performs graph-attention message passing and rebuilds the graph by K-nearest-neighbor search in the learned embedding, and the final embedding is clustered with DBSCAN to extract track candidates. The mechanism that carries the computational claim is segmented graph training: each training step draws a random window in azimuthal angle, initially $\Delta\phi=\pi/5$, and adaptively expands or shrinks it by binary search so the subgraph stays large enough for KNN while remaining small. This segmentation is what delivers the tenfold reduction in training time and memory, and the claim is that the contrastive loss still learns essentially the same embedding because the windows retain complete particle tracks.

What would settle it

Rerun the segmented training on TrackML with initial window widths of $\pi/10$ and $2\pi/5$; if track efficiency changes by more than the quoted $\pm0.0003$ or the tenfold training saving disappears, the reported comparison is tied to the $\Delta\phi=\pi/5$ choice.

Watch

Extended reading notes

Core claim

On the full TrackML dataset, the EggNet pipeline reconstructs tracks at efficiency $0.9587\pm0.0003$ with full-graph training and $0.9638\pm0.0003$ with segmented graph training, with duplication rates of $0.0276\pm0.0002$ and $0.0428\pm0.0003$, and fake rates of $0.0019\pm0.0001$ and $0.0021\pm0.0001$. Training on segmented subgraphs, selected by random azimuthal windows with initial width $\Delta\phi=\pi/5$ and adaptive binary-search growth, reduces training time and GPU memory by about a factor of ten. Inference stays on the full graph and takes about 1.47 seconds per event on an A100 GPU. The authors' conclusion is that EggNet with segmented training is a viable one-shot alternative to combinatorial Kalman-filter tracking at HL-LHC event sizes.

Load-bearing premise

The result rests on the assumption that slicing each event into wedges around the detector axis, initially one-fifth of the full circle, with each wedge allowed to grow or shrink, keeps enough complete particle tracks for the model to learn the same embedding as full-event training; if that assumption fails, both the tenfold savings and the matching physics scores could change.

Editorial extensions

If this is right

Trained on full HL-LHC-scale TrackML events, EggNet reconstructs tracks at efficiency above 0.958 with a fake rate near 0.002, so learned one-shot tracking operates at realistic event sizes.
Segmented graph training gives a tenfold reduction in training time and GPU memory while matching full-graph efficiency, making single-GPU training possible.
Track efficiency is higher at low transverse momentum and lower at high transverse momentum, pointing to imbalanced training statistics as the next physics target.
The quadratic scaling of KNN is the main scalability limit; replacing KNN or cutting its input size would further improve the pipeline.

Reading between the lines

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

If adaptive-window training preserves complete tracks, the same segmentation idea should transfer to inference and to other graph-based trackers, even though the paper only applies it during training.
The size of the saving at equal accuracy suggests training cost is controlled by graph size rather than model capacity, so approximate-nearest-neighbor search could buy further reductions.
A direct test of the locality assumption is to compare particles whose hits cross several training windows with particles fully inside one window; a deficit for boundary-crossing particles would call the segmentation design into question.
Testing the strategy on a different detector geometry would show whether the result is specific to the TrackML layout or generic to cylindrical trackers.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note

read the letter

First thing you should know: this is a straightforward engineering follow-up to the authors' own EggNet paper, now evaluated on the full TrackML dataset, plus a training trick (azimuthal phi-windowed subgraphs) that cuts training time and memory by about 10x. The physics numbers look credible and internally consistent: efficiency 0.9587 (full) and 0.9638 (segmented), fake rate ~0.002. The segmented training gives a modest efficiency gain and a real compute win. That's the honest core of the paper.

What's new: the full-dataset benchmark and the phi-window segmentation strategy. Prior EggNet was on a reduced dataset. Both are useful, incremental extensions. There's no new architecture or theoretical insight, and that's fine — the value is in the careful evaluation.

What's soft: first, the 'minimal impact on physics performance' claim understates the duplication rate jump from 0.0276 to 0.0428 — a 55% relative increase. Duplication isn't efficiency, but in a tracking pipeline it costs real CPU downstream, so 'minimal' is too kind. Second, the factor-of-10 claim rests on a single hyperparameter, initial delta-phi = pi/5, and there is no ablation showing the result is robust to window size or to the adaptive binary-search rule. A different window size could change both the compute savings and the physics metrics, so the headline comparison is conditional. The paper acknowledges the window has to be 'large enough to retain track completeness' but never checks that. Third, timing and memory numbers have no error bars, and the DBSCAN epsilon (0.1) is described as 'optimal' without a held-out selection procedure. Fourth, no external baseline is shown: no Kalman filter, no Exa.TrkX, no other GNN pipeline. So the reader can't judge where EggNet sits relative to alternatives.

Citation pattern is fine: self-citation is appropriate since this is a continuation, and the prior work is correctly credited.

Who is this for? People working on learned tracking for HL-LHC, particularly those concerned with training cost. It's a practical contribution, not a physics breakthrough. I'd give it a serious referee: the measurements are concrete, the method is described well enough to reproduce, and the segmentation idea is worth scrutiny even if the current evidence is a single point. Recommend they be asked to add an ablation on window size and report computing metrics with uncertainties.

Referee Report

4 major / 6 minor

Summary. This paper evaluates the EggNet graph-neural-network track-reconstruction pipeline on the full TrackML dataset (2,000 training and 400 test events). The authors train EggNet both with full event graphs and with a proposed "segmented graph training" scheme in which each training step uses a random azimuthal-\phi window with initial width \Delta\phi = \pi/5, adaptively grown or shrunk by binary search. On the test split they report full-graph (segmented) track efficiencies of 0.9587\pm0.0003 (0.9638\pm0.0003), duplication rates of 0.0276\pm0.0002 (0.0428\pm0.0003), and fake rates of 0.0019\pm0.0001 (0.0021\pm0.0001). They also provide per-event runtime breakdowns and claim that segmented training reduces training time and GPU memory by roughly a factor of ten with minimal impact on physics performance.

Significance. If the computational comparison is robust, the paper is a valuable benchmark for learned tracking at HL-LHC event sizes. Its strengths are a clear evaluation on the full public TrackML test split, statistical uncertainties on the physics metrics, a decomposition of runtime into graph attention, KNN, loss, backward propagation, and DBSCAN, and a realistic GPU environment (A100). The potential practical benefit of reducing training memory and time by an order of magnitude is significant. However, the evidence for the factor-of-ten claim is thinner than for the physics metrics: it rests on a single segmentation hyperparameter, has no reported uncertainty on time or memory, and lacks a definition of training convergence for the comparison.

major comments (4)

[Section 3, Figs. 5a and 5b] The central computational claim is conditional on a single untested hyperparameter. The initial window width \Delta\phi = \pi/5 is chosen, and the binary-search rule for growing and shrinking windows is based only on spacepoint count, not on track completeness. A random \phi-window truncates any track crossing the boundary, changing the true-edge set available to the contrastive loss and the KNN neighborhoods relative to full-graph training. The paper provides no ablation over window size, no measure of the fraction of truncated tracks, and no demonstration that the result is stable. Without such sensitivity evidence, the statement that segmented training has "minimal impacts" on physics and reduces cost by "about a factor of 10" is not established. Please add a scan over initial window size reporting efficiency, duplication rate, fake rate, training time, and memory for each setting.
[Section 4, DBSCAN paragraph] The sentence "we find \epsilon=0.1 to be an optimal value" does not specify the data used for this choice or the search procedure. If \epsilon and min_samples were selected using the same 400-event test split on which the headline metrics are reported, then the efficiency and fake-rate results are optimistic due to test-set adaptation. Please state whether a held-out validation set was used, and report the sensitivity of the final metrics to \epsilon (for example, \epsilon = 0.08 and 0.12) and to min_samples.
[Section 4, Figures 4 and 5] The time and memory results are presented without uncertainties or a precise measurement definition. It is unclear whether "training time" in Figure 5a is per epoch, per optimization step, or total wall-clock time to convergence; likewise "GPU memory" may be peak allocated memory, reserved memory, or something else. The full-graph and segmented runs also have no stated epoch budgets or convergence criterion, so the reported factor of ten may reflect a difference in the number of optimizer steps rather than per-step efficiency. Please define these quantities, report run-to-run variability, and compare both strategies at matched convergence points or matched total compute.
[Section 1 and Section 5] The paper does not compare EggNet with any external tracking algorithm on the same data. The introduction's claim that EggNet showed "promising track performance compared the existing track reconstruction algorithms" is not supported by any baseline in this manuscript. Because the paper is framed as an evaluation of the pipeline's physics and computing performance, adding a baseline (for example, a CKF or Exa.TrkX reference under the same test conditions) or removing the comparative claim would make the significance of the reported numbers concrete.

minor comments (6)

[Section 2, Eq. (1)] The notation "max2(0, m-d)" is nonstandard; use "max(0, m-d)^2" or spell out the squared hinge loss.
[Section 2, text] The text contains small language errors: "multi-layer perception" should be "multi-layer perceptron" and "weighting sum" should be "weighted sum".
[Section 2, text] The passage "graph-attention-based [?] message passing" has a missing citation; the attention mechanism should be identified precisely.
[Section 3, text] The adaptive window-size rule is underspecified: "too many spacepoints are selected given the criteria" does not define the criteria or the binary-search termination condition. Please state the thresholds and the search procedure for reproducibility.
[Section 4, text] The statement "About 467K particles are considered for the evaluation" should clarify that this is the total over the 400 test events and should define how track candidates matched to the same particle are treated in the efficiency and duplication-rate calculations.
[Figures 3-5] Figure captions should explicitly define the plotted quantities, axis labels, and any error bars; currently Figures 4 and 5 rely on the body text to explain what is measured.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: central claims are external benchmark measurements, not derivations from the model's own equations.

full rationale

The paper is a measurement paper. The physics metrics in Section 4 are obtained by running the EggNet pipeline on the full TrackML test set and comparing DBSCAN clusters to particles; those numbers are external benchmark results, not derived from the model's loss, from a fitted parameter, or from a self-citation. The segmentation comparison in Section 5 is likewise an empirical measurement at a stated hyperparameter (initial window Delta-phi = pi/5); the absence of an ablation or error bars on the factor-of-10 savings is a robustness limitation, not a circularity. Citations to the authors' prior EggNet paper [9] describe the architecture and loss, but the load-bearing claims of this paper—full-dataset efficiency, duplication rate, fake rate, training time, and GPU memory—are not justified by those citations; they are reported measurements on a public dataset. No equation in the paper is defined in terms of the quantity it supposedly predicts, and no fitted parameter is renamed as a prediction. Therefore no specific reduction can be exhibited, and the paper is self-contained against an external benchmark.

Assumptions & free parameters 8 free parameters · 5 assumptions · 0 invented entities

No new physical entities, forces, or conserved quantities are introduced; the p-space embedding is a mathematical representation. The central claim rests on hyperparameters (loss margin, KNN k, DBSCAN thresholds, phi window, architecture dimensions) and on domain assumptions about TrackML truth, embedding separability, and segmentation unbiasedness.

free parameters (8)

Contrastive loss margin m = 1
Set to 1 in Section 2; controls the repulsive separation between fake edges and affects embedding geometry and clustering.
KNN k = 10
Section 4; determines graph sparsity and message passing range, directly affecting edge efficiency and purity.
DBSCAN epsilon = 0.1
Section 4 states 'we find epsilon=0.1 to be an optimal value'; a hand-tuned clustering radius that determines track candidates.
DBSCAN min_samples = 3
Section 4; minimum cluster size, set without an explicit tuning procedure.
Initial phi window size = pi/5
Section 3; segmentation hyperparameter that controls training cost and the completeness of particle tracks within each window.
Embedding dimensions = h-space 128, p-space 24
Section 4; architecture capacity choices that define the metric space used for KNN and DBSCAN.
EggNet iterations and message passing steps = 5 iterations, 8 steps
Section 4; architecture depth hyperparameters that trade off model expressiveness against computational cost.
Learning rate = 2e-4
Section 4; Adam optimizer setting chosen for training stability without a reported search.

assumptions (5)

domain assumption TrackML simulation labels and detector geometry are accurate enough to define ground truth track candidates.
The entire evaluation in Section 4 treats TrackML truth as the reference for efficiency, duplication, and fake rate.
domain assumption Contrastive loss in Euclidean p-space produces embeddings in which DBSCAN clusters correspond to particle tracks.
Sections 2 and 4 assume that the learned embedding separates particles well enough for fixed-radius DBSCAN to recover tracks.
domain assumption The KNN graph with k=10 in embedding space preserves the edges needed for message passing.
Section 2 iteratively rebuilds graphs from KNN in embedding space; if true edges fall outside the k nearest neighbors, they are lost.
ad hoc to paper Random phi-window segmentation provides a representative training distribution.
Section 3 introduces this segmentation strategy for training only; the paper assumes it does not bias the learned embedding.
ad hoc to paper DBSCAN with epsilon=0.1 and min_samples=3 identifies track candidates without further per-event tuning.
Section 4 fixes these clustering parameters globally, assuming they generalize across all test events.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Physics and Computing Performance of the EggNet Tracking Pipeline." pith.science (2026). https://pith.science/paper/HDF3MSNG

@misc{pith2026250603415,
  author       = {Pith},
  title        = {Pith review of: Physics and Computing Performance of the EggNet Tracking Pipeline},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/HDF3MSNG}},
  note         = {Machine review of arXiv:2506.03415}
}

read the original abstract

Particle track reconstruction is traditionally computationally challenging due to the combinatorial nature of the tracking algorithms employed. Recent developments have focused on novel algorithms with graph neural networks (GNNs), which can improve scalability. While most of these GNN-based methods require an input graph to be constructed before performing message passing, a one-shot approach called EggNet that directly takes detector spacepoints as inputs and iteratively apply graph attention networks with an evolving graph structure has been proposed. The graphs are gradually updated to improve the edge efficiency and purity, thus providing a better model performance. In this work, we evaluate the physics and computing performance of the EggNet tracking pipeline on the full TrackML dataset. We also explore different techniques to reduce constraints on computation memory and computing time.

Figures

Figures reproduced from arXiv: 2506.03415 by the authors.

**Figure 2.** The EggNet architecture. i corresponds to each EggNet iteration, and j corresponds to each message passing step. An EggNet iteration generally consists of a graph attention block and a KNN. The first iteration does not perform graph attention and the last iteration does not perform KNN. The EggNet model is trained with the contrastive loss. For each edge connection, the individual loss term is given by: l = yd 2 + (… view at source ↗

**Figure 3.** Track efficiency as a function of particle transverse momentum (pT), shown for both full graph and segmented graph training. About 467K particles are considered for the evaluation. In [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗

Figures from the paper (2 more)

**Figure 4.** Figure 4: Training (a) and inference (b) time for each event versus the number of spacepoints. [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]

**Figure 5.** Figure 5: (a) Comparison of the training time between full graph and segmented graph train [PITH_FULL_IMAGE:figures/full_fig_p006_5.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

19 extracted references · 7 canonical work pages

[1]

Strandlie, R

A. Strandlie, R. Frühwirth, Track and vertex reconstruction: From classical to adaptive methods, Rev. Mod. Phys.82, 1419 (2010). 10.1103/RevModPhys.82.1419

doi:10.1103/revmodphys.82.1419 2010
[2]

Aaboud et al

M. Aaboud et al. (ATLAS), Performance of the ATLAS Track Reconstruction Al- gorithms in Dense Environments in LHC Run 2, Eur. Phys. J. C77, 673 (2017), 1704.07983. 10.1140/epjc/s10052-017-5225-7

arXiv 2017
[3]

Chatrchyan et al

S. Chatrchyan et al. (CMS), Description and performance of track and primary- vertex reconstruction with the CMS tracker, JINST9, P10009 (2014),1405.6569. 10.1088/1748-0221/9/10/P10009

arXiv 2014
[4]

Ju et al

X. Ju et al. (Exa.TrkX), Performance of a geometric deep learning pipeline for HL-LHC particle tracking, Eur. Phys. J. C81, 876 (2021),2103.06995. 10.1140/epjc/s10052- 021-09675-8

arXiv 2021
[5]

Biscarat, S

C. Biscarat, S. Caillou, C. Rougier, J. Stark, J. Zahreddine, Towards a realistic track reconstruction algorithm based on graph neural networks for the HL-LHC, EPJ Web Conf.251, 03047 (2021),2103.00916. 10.1051/epjconf/202125103047

arXiv 2021
[6]

Caillou, P

S. Caillou, P. Calafiura, S.A. Farrell, X. Ju, D.T. Murnane, C. Rougier, J. Stark, A. Val- lier (ATLAS), Tech. rep., CERN, Geneva (2022),https://cds.cern.ch/record/ 2815578

work page 2022
[7]

Lieret, G

K. Lieret, G. DeZoort, D. Chatterjee, J. Park, S. Miao, P. Li, High Pileup Particle Track- ing with Object Condensation (2023),2312.03823

arXiv 2023
[8]

Accelerating the Inference of the Exa.TrkX Pipeline

A. Lazar et al., Accelerating the Inference of the Exa.TrkX Pipeline, J. Phys. Conf. Ser. 2438, 012008 (2023),2202.06929. 10.1088/1742-6596/2438/1/012008

work page Pith review arXiv 2023

Show all 19 references

[9]

Calafiura, J

P. Calafiura, J. Chan, L. Delabrouille, B. Wang, EggNet: An Evolving Graph-based Graph Attention Network for Particle Track Reconstruction (2024),2407.13925

2024 arXiv
[10]

Amrouche et al., The Tracking Machine Learning challenge : Accuracy phase (2019), 1904.06778

S. Amrouche et al., The Tracking Machine Learning challenge : Accuracy phase (2019), 1904.06778

2019 arXiv
[11]

Amrouche et al., The Tracking Machine Learning Challenge: Throughput Phase, Comput

S. Amrouche et al., The Tracking Machine Learning Challenge: Throughput Phase, Comput. Softw. Big Sci.7, 1 (2023),2105.01160. 10.1007/s41781-023-00094-w

2023 arXiv
[12]

Ester, H.P

M. Ester, H.P. Kriegel, J. Sander, X. Xu, A density-based algorithm for discovering clusters in large spatial databases with noise, inProceedings of the Second International Conference on Knowledge Discovery and Data Mining(AAAI Press, 1996), KDD’96, p. 226–231

1996
[13]

Elfwing, E

S. Elfwing, E. Uchibe, K. Doya, Sigmoid-weighted linear units for neural network func- tion approximation in reinforcement learning, Neural Networks107, 3 (2018), special issue on deep reinforcement learning. https://doi.org/10.1016/j.neunet.2017.12.012

2018 doi
[14]

J.L. Ba, J.R. Kiros, G.E. Hinton, Layer normalization (2016),1607.06450

2016 arXiv
[15]

Paszke, S

A. Paszke, S. Gross, F. Massa, A. Lerer, J. Bradbury, G. Chanan, T. Killeen, Z. Lin, N. Gimelshein, L. Antiga et al., inAdvances in Neural Infor- mation Processing Systems 32, edited by H. Wallach, H. Larochelle, A. Beygelzimer, F. d'Alché-Buc, E. Fox, R. Garnett (Curran Asso-...

2019
[16]

Falcon, The PyTorch Lightning team, PyTorch Lightning (2019),https:// github.com/Lightning-AI/lightning

W. Falcon, The PyTorch Lightning team, PyTorch Lightning (2019),https:// github.com/Lightning-AI/lightning

2019
[17]

Atkinson, S

M.J. Atkinson, S. Caillou, P. Clafiura, C. Collard, S.A. Farrell, B. Huth, X. Ju, R. Liu, T. Minh Pham, D.c.a. Murnane et al., gnn4itk,https://github.com/GNN4ITkTeam/ CommonFramework
[18]

Kingma, J

D. Kingma, J. Ba, Adam: A method for stochastic optimization (2014),1412.6980

2014 arXiv
[19]

Team, RAPIDS: Libraries for End to End GPU Data Science (2023),https: //rapids.ai

R.D. Team, RAPIDS: Libraries for End to End GPU Data Science (2023),https: //rapids.ai

2023

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.

[1] [1]

Strandlie, R

A. Strandlie, R. Frühwirth, Track and vertex reconstruction: From classical to adaptive methods, Rev. Mod. Phys.82, 1419 (2010). 10.1103/RevModPhys.82.1419

doi:10.1103/revmodphys.82.1419 2010

[2] [2]

Aaboud et al

M. Aaboud et al. (ATLAS), Performance of the ATLAS Track Reconstruction Al- gorithms in Dense Environments in LHC Run 2, Eur. Phys. J. C77, 673 (2017), 1704.07983. 10.1140/epjc/s10052-017-5225-7

arXiv 2017

[3] [3]

Chatrchyan et al

S. Chatrchyan et al. (CMS), Description and performance of track and primary- vertex reconstruction with the CMS tracker, JINST9, P10009 (2014),1405.6569. 10.1088/1748-0221/9/10/P10009

arXiv 2014

[4] [4]

Ju et al

X. Ju et al. (Exa.TrkX), Performance of a geometric deep learning pipeline for HL-LHC particle tracking, Eur. Phys. J. C81, 876 (2021),2103.06995. 10.1140/epjc/s10052- 021-09675-8

arXiv 2021

[5] [5]

Biscarat, S

C. Biscarat, S. Caillou, C. Rougier, J. Stark, J. Zahreddine, Towards a realistic track reconstruction algorithm based on graph neural networks for the HL-LHC, EPJ Web Conf.251, 03047 (2021),2103.00916. 10.1051/epjconf/202125103047

arXiv 2021

[6] [6]

Caillou, P

S. Caillou, P. Calafiura, S.A. Farrell, X. Ju, D.T. Murnane, C. Rougier, J. Stark, A. Val- lier (ATLAS), Tech. rep., CERN, Geneva (2022),https://cds.cern.ch/record/ 2815578

work page 2022

[7] [7]

Lieret, G

K. Lieret, G. DeZoort, D. Chatterjee, J. Park, S. Miao, P. Li, High Pileup Particle Track- ing with Object Condensation (2023),2312.03823

arXiv 2023

[8] [8]

Accelerating the Inference of the Exa.TrkX Pipeline

A. Lazar et al., Accelerating the Inference of the Exa.TrkX Pipeline, J. Phys. Conf. Ser. 2438, 012008 (2023),2202.06929. 10.1088/1742-6596/2438/1/012008

work page Pith review arXiv 2023