B-Jet Tagging with Retentive Networks: A Novel Approach and Comparative Study

Ayse Asu Guvenli; Bora Isildak

REVIEW 5 major objections 5 minor 1 cited by

B-Jet Tagging with Retentive Networks: A Novel Approach and Comparative Study

T0 review · 5 major / 5 minor · reviewed 2026-08-11 · deepseek-v4-flash

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

Pith's one-line read A 330k-parameter retentive network tags b-jets with 0.956 AUC, edging out an MLP twin.

desk verdict A clean, honest application of RetNet to b-jet tagging on CMS Open Data, but the headline performance numbers are validation-selected and the MLP comparison isn't yet controlled enough to support the efficiency claim. read the letter →

arxiv 2412.08134 v1 pith:D4FTAHMA submitted 2024-12-11 hep-ex

Ayse Asu Guvenli , Bora Isildak This is my paper · ORCID

classification hep-ex

keywords b-jettaggingretentivenetworksjetflavouridentificationlow-levelfeaturesparticlejetsdeeplearninginhigh-energyphysicsCMSOpenDataefficiency

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper proposes JetRetNet, a b-jet tagging model built on retentive networks, and claims it offers an efficient alternative for flavor tagging when computational resources or training data are limited. On a simulated sample of semileptonic top-quark pair events, JetRetNet reaches a validation AUC of 0.956 and F1 of 0.914 with about 330,000 trainable parameters, outperforming a same-shaped MLP baseline at every misidentification rate. Against much larger state-of-the-art taggers, it does not win at any working point, but remains competitive at tight mistag rates despite training on two orders of magnitude fewer jets. The authors read this as evidence that retentive networks can capture the track and secondary-vertex structure needed for b-tagging at low computational cost.

What carries the argument

The load-bearing mechanism is the retention module from retentive networks, which computes Key, Query, and Value matrices and multiplies their product by an exponential-decay matrix with a causal mask, so earlier sequence positions fade while later ones remain visible. It is implemented as a multi-scale retention module with per-head decay scales, a group norm, and a swish gate. JetRetNet applies two such modules in parallel, one over sorted track features and one over sorted secondary-vertex features, then concatenates their outputs with global jet features before a final feed-forward classifier. Because the retention mask imposes an order, the paper must supply an ordering, and it does so by sorting tracks and secondary vertices by decreasing transverse momentum as a proxy for the experimentally inaccessible time order.

What would settle it

Retrain JetRetNet on the same data with the order of tracks and secondary vertices randomly shuffled for every jet. If the validation AUC stays at about 0.956, the retention mask contributes nothing through ordering, and the paper's architectural claim collapses to a permutation-invariant set model.

Watch

Extended reading notes

Core claim

The central discovery claimed is that a compact retentive-network architecture, JetRetNet, can learn b-jet discrimination from low-level jet constituents about as well as an MLP with the same hyperparameters and nearly as well as much larger models. The model processes tracks and secondary vertices through separate multi-scale retention blocks, appends global jet features, and classifies with a small feed-forward head. On the held-out validation sample it reaches AUC 0.955-0.956 and F1 0.913-0.914, with training and validation curves nearly overlapping, and it maintains an advantage over the MLP across loose, medium, and tight misidentification working points in both b-versus-light and b-versus-c comparisons. The authors explicitly call the comparison with DeepJet and Particle Transformer unfair because of the dataset-size gap, and report that JetRetNet does not surpass them at any working point, yet it comes closest at the tight working point, which is where a low-resource tagger would be most useful.

Load-bearing premise

The assumption that sorting jet constituents by decreasing momentum gives the retention mask a meaningful order; if any ordering performs identically, the model's sequence machinery is not actually being used.

Editorial extensions

If this is right

At the claimed 0.956 validation AUC, JetRetNet operates within reach of taggers used in physics analyses while using roughly 330k parameters and modest training data, suggesting low-resource experiments can deploy a performant tagger.
Because JetRetNet beats the same-hyperparameter MLP at all misidentification rates, the paper implies that the retention-style sequence processing, not raw model size, contributes the edge in b-jet discrimination.
At the tight working point of 0.001 misidentification rate, the gap to much larger models narrows, so the architecture may be most useful for high-purity b-jet selection with small statistics.
The training and validation curves nearly overlap, indicating that the model is not visibly overfitting the training sample even with early stopping, which supports the claim that its capacity is being used productively.
A model of this size can be trained on a data slice of four million jets, so the same architecture is a plausible starting point for experiments that lack the multi-hundred-million-jet datasets used by the larger taggers.

Reading between the lines

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

The causal mask is the fragile part: because jet constituents have no true time order, if randomly permuting the pT-sorted input leaves performance unchanged, JetRetNet's gain over the MLP would not come from the retention sequence at all, and the model would reduce to a permutation-invariant set architecture.
The same 330k-parameter design could be applied to other object-tagging tasks where constituent ordering is arbitrary, such as quark-gluon discrimination or boosted-W tagging, and a permutation-invariant variant would be a useful control benchmark.
Since the paper identifies dataset size rather than architecture as the main handicap, training the same model on the full public simulated dataset might close more of the gap to DeepJet and Particle Transformer without changing the network.
The near-overlap of training and validation curves suggests the model is capacity-limited rather than data-limited, so increasing the hidden dimension or number of RetNet blocks could raise the performance ceiling without immediately hurting generalization.

Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note

read the letter

Two things to know. First, this is a genuine new application: Retentive Networks for b-jet tagging, with a custom architecture (JetRetNet) that processes tracks, secondary vertices, and global jet features separately, trained on a public CMS simulation with 330k parameters. Second, the central claim that this is an efficient alternative for limited-resource use rests on validation-set metrics, not on a held-out test set. That is the soft spot to worry about.

The paper does real things well. The data preprocessing is documented thoroughly, the train/validation split is clear, and the MLP baseline is trained with the same hyperparameters as a controlled comparison. The authors also explicitly label the DeepJet/ParT comparison as unfair because of the dataset size difference, which is the right call. The architecture is a reasonable assembly of existing components, and the pT-ordered track/SV sequences are a defensible design choice.

The weaknesses are proportionate. The stress-test note is correct: Section 3.3 describes early stopping on validation loss, Figure 7 reports the max validation AUC (0.955) and F1, and there is no separate test set. That makes the reported performance optimistically biased. The MLP comparison is the key evidence for efficiency, but the MLP parameter count is not reported, so the capacity difference between the two models is unknown. If the MLP has substantially fewer parameters, its validation-selected performance could be worse for reasons unrelated to the RetNet mechanism. The paper needs a held-out test set and ideally repeated runs to give error bars. A second concern is the causal masking assumption: pT ordering is a proxy for time, but the retention mask imposes a one-directional decay. If that mask doesn't actually capture useful structure, the model effectively becomes a set-based model, and the RetNet-specific advantage is unsubstantiated. This isn't a fatal problem, but it should be tested or at least discussed.

Overall, this is a solid engineering study with honest limitations. The novelty is modest, but the execution is careful and the dataset is public. The claimed advantage over an MLP and the efficiency story are not yet proven. Still, a good referee could get this into publishable shape with a proper test set and more careful model comparison.

I would accept this for peer review. The topic is relevant, the work is reproducible in principle, and the flaws are addressable. My own verdict would be conditional: require the held-out test set, metric uncertainty, and a reported MLP parameter count before I'd trust the efficiency claim. I would not cite it in my own work in its current form, but I'd bring it to a reading group to discuss the pT-ordering assumption in sequence models for jets.

Referee Report

5 major / 5 minor

Summary. The paper applies a Retentive Network (RetNet) to b-jet tagging on a simulated CMS Open Data t-tbar sample, using global jet, track, and secondary-vertex features. The proposed JetRetNet model is trained on 4 million jets and validated on 1 million jets, with reported validation AUC around 0.956 and F1 around 0.914. It is compared with a same-hyperparameter MLP baseline and with externally published DeepJet and Particle Transformer results, and the authors conclude that RetNet-based models are a promising low-parameter alternative. The central evidence is the validation-based comparison; no separate held-out test set is used.

Significance. If the reported performance estimates were unbiased, the paper would be a useful demonstration that a compact sequence model can approach state-of-the-art b-tagging performance with limited data. The use of public CMS Open Data, the clear description of the input features, and the controlled MLP baseline are commendable. The authors also correctly label the DeepJet/ParT comparison as unfair, which is an appropriate caveat. However, the headline numbers are selected from the validation set, and the manuscript lacks a held-out test evaluation, uncertainty estimates, and a test of the causal-ordering assumption that is central to the RetNet architecture. These gaps currently prevent the paper from supporting its main efficiency claim.

major comments (5)

[§2.3, §3.3, Fig. 7] The paper uses only a train/validation split (4M/1M jets; §2.3) and selects the model by early stopping on validation loss (§3.3). Figure 7 explicitly quotes the maximum validation AUC (0.955) and F1 (0.913), so these are validation-selected estimates, not unbiased generalization estimates. Since the abstract's central claim rests on exactly these numbers, the manuscript needs an untouched test set and a final evaluation on it, with the model-selection procedure fixed, before the efficiency claim can be supported.
[§3.3, Fig. 8] The MLP baseline is described as using the same hyperparameters (four layers, hidden dimension 128), but its trainable parameter count is not reported, and no repeated training runs or error bars are given. Because JetRetNet has 330k parameters, a capacity mismatch could explain part of the ROC advantage. Please report the MLP parameter count, train both models over multiple seeds, and show mean and standard deviation for AUC, F1, and the ROC curves.
[§2.3, Eqs. (1)–(2)] The retention mechanism applies causal masking through the matrix D in Eq. (2), while §2.3 orders tracks and SVs by decreasing pT as a proxy for time order. For jets, pT ordering is not a physically causal temporal sequence; if the causal mask simply imposes an arbitrary order, the RetNet-specific benefit over set-based aggregation is not demonstrated. A concrete test would be to compare against the same architecture with a bidirectional or no mask, or against a permutation-invariant baseline, and to report sensitivity to the track ordering.
[§2.3] The fixed maximum of 16 tracks and 5 SVs is applied by clipping or padding, yet the paper reports that only 2.63% of jets have fewer than 17 tracks and only 0.012% have fewer than 6 SVs. This means that the large majority of jets have at least 17 tracks and at least 6 SVs, so most jets are truncated and a substantial part of the low-level information is discarded in nearly every jet. The authors should report the fraction of jets with more than 16 tracks and more than 5 SVs, and investigate how performance changes with the maximum length.
[Abstract, §4] The conclusion that RetNet models are an efficient alternative for limited computational resources is supported only by the 330k parameter count; no training time, inference time, memory footprint, FLOPs, or throughput measurements are reported. An efficiency claim requires at least one compute metric beyond parameter count, especially because the model is compared with an MLP baseline whose computational cost is not characterized.

minor comments (5)

[Abstract] The phrase "alternative for b-jet with limited computational resources" appears to be missing the word "tagging" after "b-jet."
[Eq. (1)] The text above Eq. (1) contains the typo "Querry"; it should be "Query."
[Fig. 3] The last two panels of Figure 3 are both labeled "SV dxy"; one is likely intended to be SV dxy significance and the other SV dlen significance, matching the features listed in §2.2.
[Fig. 5] The caption of Figure 5 describes the panels as "Number tracks (right) and number of SV's (left)", but the surrounding text and figure layout suggest the left/right labels may be reversed; please confirm the panel order.
[§4, Figs. 9–10] The sentence "Also, both performances are given to make sure there are no discrepancies between the training and validation datasets" is ambiguous; it should state explicitly which curves correspond to training and which to validation, and whether the MLP is included in those figures.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the paper trains a standard classifier on public data using an externally published architecture, and the only explicit benchmark caveat is acknowledged rather than used to force a claim.

full rationale

The paper's derivation chain is an empirical machine-learning comparison, not a derivation of predictions from fitted inputs. The retention mechanism is imported from the external Retentive Network paper [12], which has no author overlap with the present work, so no self-citation is load-bearing. No uniqueness theorem is invoked, and no parameter is fitted to a subset of data and then reported as a prediction of a closely related quantity. The reported AUC (0.956) and F1 (0.914) are validation-set metrics selected over epochs, which is a statistical bias or experimental-design concern rather than a circularity: the metrics are not defined in terms of the paper's conclusion, and the MLP baseline is trained under the same protocol. The DeepJet and Particle Transformer comparison is explicitly labeled 'not fair' because of the smaller training set, so the paper does not use those external results to force its claim. The assumption that pT-decile ordering proxies for time order is a modeling assumption that could be wrong, but it is not a circular reduction of the output to the input. Under the required standard of exhibiting a specific equation or fit that is equivalent by construction to the claimed result, no circular step can be identified.

Assumptions & free parameters 4 free parameters · 4 assumptions · 0 invented entities

The central claim rests on standard ML assumptions and a few hand-chosen preprocessing settings. No new physical entities are introduced. The pT-ordering proxy and the c-jet-as-background definition are domain assumptions that could change results if altered.

free parameters (4)

Maximum number of tracks per jet = 16
Chosen to cover 97.37% of jets (Section 2.3); sets input width and model capacity.
Maximum number of secondary vertices per jet = 5
Chosen to cover 99.988% of jets (Section 2.3).
Track selection cuts = pT > 1.0 GeV, chi2/ndof < 5.0
Hand-applied to reduce noise (Section 2.2); affects all downstream features.
Mistag working points = 0.1, 0.01, 0.001
Selected to define loose, medium, and tight operating points for reporting efficiencies (Section 4).

assumptions (4)

domain assumption Sorting tracks and SVs by decreasing pT approximates the temporal order of jet formation, which the retention mechanism's causal mask is designed to exploit.
Section 2.3: 'This approach serves as a proxy for the time order of jet formation.' This is a key inductive-bias assumption for the RetNet backbone.
domain assumption The simulated semileptonic ttbar sample is representative for evaluating b-jet tagging performance.
Section 2.1: model is trained only on semileptonic ttbar events from CMS Open Data; no cross-check on other samples or production modes.
domain assumption Labeling c-jets as background alongside light jets is appropriate for the reported b-tagging metric.
Section 2.3: b-jets are assigned 1, and u,d,s,c,g jets are assigned 0. This is standard for b-tagging, but c-jet confusion is known to be harder; the paper reports b vs c separately.
standard math The RetNet retention mechanism as defined in Equations 1-3 is taken as a valid building block from the cited paper (Sun et al., 2023).
The paper reuses the architecture without deriving or questioning it; this is acceptable background.

how reviews work

0 comments

Cite this review

Pith. "Pith review of B-Jet Tagging with Retentive Networks: A Novel Approach and Comparative Study." pith.science (2026). https://pith.science/paper/D4FTAHMA

@misc{pith2026241208134,
  author       = {Pith},
  title        = {Pith review of: B-Jet Tagging with Retentive Networks: A Novel Approach and Comparative Study},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/D4FTAHMA}},
  note         = {Machine review of arXiv:2412.08134}
}

read the original abstract

Identifying jets originating from bottom quarks is vital in collider experiments for new physics searches. This paper proposes a novel approach based on Retentive Networks (RetNet) for b-jet tagging using low-level features of jet constituents along with high-level jet features. A simulated \ttbar dataset provided by CERN CMS Open Data Portal was used, where only semileptonic decays of \ttbar pairs produced by 13 TeV proton-proton collisions are included. The performance of the newly proposed Retentive Network model is compared with state-of-the-art models such as DeepJet and Particle Transformer, as well as with a baseline MLP (Multi-Layer-Perceptron) classifier. Despite using a relatively smaller dataset, the Retentive Networks demonstrate a promising performance with only 330k trainable parameters. Results suggest that RetNet-based models can be used as an efficient alternative for b-jet with limited computational resources.

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

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

A Survey of Retentive Network
cs.CL 2025-06 conditional novelty 2.0 of 10

A review that describes the RetNet architecture and enumerates its applications across many domains, without presenting new experimental results.

Reference graph

Works this paper leans on

18 extracted references · 15 canonical work pages · cited by 1 Pith paper

[1]

Technical report, CERN, Geneva, 2009

Algorithms for b Jet identification in CMS. Technical report, CERN, Geneva, 2009

work page 2009
[2]

Identification of b-quark jets with the cms experiment

The CMS collaboration. Identification of b-quark jets with the cms experiment. Journal of Instrumentation, 8(04):P04013–P04013, April 2013

work page 2013
[3]

Technical report, CERN, Geneva, 2016

Identification of b quark jets at the CMS Experiment in the LHC Run 2. Technical report, CERN, Geneva, 2016

work page 2016
[4]

Journal of Instrumentation, 11(04):P04008–P04008, April 2016

Performance of b-jet identification in the atlas experiment. Journal of Instrumentation, 11(04):P04008–P04008, April 2016

work page 2016
[5]

Journal of Instrumentation, 13(05):P05011–P05011, May 2018

Identification of heavy-flavour jets with the cms detector in pp collisions at 13 tev. Journal of Instrumentation, 13(05):P05011–P05011, May 2018

work page 2018
[6]

Tech- nical report, CERN, Geneva, 2020

Deep Sets based Neural Networks for Impact Parameter Flavour Tagging in ATLAS. Tech- nical report, CERN, Geneva, 2020. All figures including auxiliary figures are available at https://atlas.web.cern.ch/Atlas/GROUPS/PHYSICS/PUBNOTES/ATL-PHYS-PUB-2020-014

work page 2020
[7]

Technical report, CERN, Geneva, 2017

Identification of Jets Containing b-Hadrons with Recurrent Neural Networks at the ATLAS Experi- ment. Technical report, CERN, Geneva, 2017. All figures including auxiliary figures are available at https://atlas.web.cern.ch/Atlas/GROUPS/PHYSICS/PUBNOTES/ATL-PHYS-PUB-2017-003

work page 2017
[8]

Journal of Instrumentation, 13(05):P05011, may 2018

Identification of heavy-flavour jets with the cms detector in pp collisions at 13 tev. Journal of Instrumentation, 13(05):P05011, may 2018

work page 2018

Show all 18 references

[9]

E. Bols, J. Kieseler, M. Verzetti, M. Stoye, and A. Stakia. Jet flavour classification using deepjet. Journal of Instrumentation, 15(12):P12012–P12012, December 2020

2020
[10]

Jet tagging via particle clouds

Huilin Qu and Loukas Gouskos. Jet tagging via particle clouds. Physical Review D, 101(5), March 2020

2020
[11]

Particle transformer for jet tagging, 2024

Huilin Qu, Congqiao Li, and Sitian Qian. Particle transformer for jet tagging, 2024

2024
[12]

Retentive network: A successor to transformer for large language models, 2023

Yutao Sun, Li Dong, Shaohan Huang, Shuming Ma, Yuqing Xia, Jilong Xue, Jianyong Wang, and Furu Wei. Retentive network: A successor to transformer for large language models, 2023

2023
[13]

Cms open data primary datasets, 2012

CERN Open Data Portal. Cms open data primary datasets, 2012. Accessed: 2024-12-09

2012
[14]

The anti-ktjet clustering algorithm

Matteo Cacciari, Gavin P Salam, and Gregory Soyez. The anti-ktjet clustering algorithm. Journal of High Energy Physics, 2008(04):063–063, April 2008

2008
[15]

Machine learning algorithms for b-jet tagging at the atlas experiment, 2017

Michela Paganini. Machine learning algorithms for b-jet tagging at the atlas experiment, 2017

2017
[16]

Jet b-tag visualization with tikz

Izaak Neutelings. Jet b-tag visualization with tikz. https://tikz.net/jet_btag/, 2024. Accessed: December 9, 2024

2024
[17]

Gomez, Lukasz Kaiser, and Illia Polosukhin

Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need, 2023

2023
[18]

Pytorch: An imperative style, high-performance deep learning library, 2019

Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, Alban Desmaison, Andreas K¨ opf, Edward Yang, Zach DeVito, Martin Raison, Alykhan Tejani, Sasank Chilamkurthy, Benoit Steiner, Lu F...

2019

Pith tools

Reviewed August 11, 2026 · model on record in the stance chip above.

[1] [1]

Technical report, CERN, Geneva, 2009

Algorithms for b Jet identification in CMS. Technical report, CERN, Geneva, 2009

work page 2009

[2] [2]

Identification of b-quark jets with the cms experiment

The CMS collaboration. Identification of b-quark jets with the cms experiment. Journal of Instrumentation, 8(04):P04013–P04013, April 2013

work page 2013

[3] [3]

Technical report, CERN, Geneva, 2016

Identification of b quark jets at the CMS Experiment in the LHC Run 2. Technical report, CERN, Geneva, 2016

work page 2016

[4] [4]

Journal of Instrumentation, 11(04):P04008–P04008, April 2016

Performance of b-jet identification in the atlas experiment. Journal of Instrumentation, 11(04):P04008–P04008, April 2016

work page 2016

[5] [5]

Journal of Instrumentation, 13(05):P05011–P05011, May 2018

Identification of heavy-flavour jets with the cms detector in pp collisions at 13 tev. Journal of Instrumentation, 13(05):P05011–P05011, May 2018

work page 2018

[6] [6]

Tech- nical report, CERN, Geneva, 2020

Deep Sets based Neural Networks for Impact Parameter Flavour Tagging in ATLAS. Tech- nical report, CERN, Geneva, 2020. All figures including auxiliary figures are available at https://atlas.web.cern.ch/Atlas/GROUPS/PHYSICS/PUBNOTES/ATL-PHYS-PUB-2020-014

work page 2020

[7] [7]

Technical report, CERN, Geneva, 2017

Identification of Jets Containing b-Hadrons with Recurrent Neural Networks at the ATLAS Experi- ment. Technical report, CERN, Geneva, 2017. All figures including auxiliary figures are available at https://atlas.web.cern.ch/Atlas/GROUPS/PHYSICS/PUBNOTES/ATL-PHYS-PUB-2017-003

work page 2017

[8] [8]

Journal of Instrumentation, 13(05):P05011, may 2018

Identification of heavy-flavour jets with the cms detector in pp collisions at 13 tev. Journal of Instrumentation, 13(05):P05011, may 2018

work page 2018