Bidirectional Long Short-Term Memory (BLSTM) neural networks for reconstruction of top-quark pair decay kinematics

Fardin Syed; Pekka Sinervo; Riccardo Di Sipio

REVIEW 3 major objections 5 minor 14 references

Bidirectional Long Short-Term Memory (BLSTM) neural networks for reconstruction of top-quark pair decay kinematics

T0 review · 3 major / 5 minor · reviewed 2026-08-14 · deepseek-v4-flash

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

Pith's one-line read A BLSTM neural network reconstructs top-quark-pair decay kinematics about as well as a chi-squared jet-assignment fit.

desk verdict The BLSTM is a plausible new tool, but the paper's headline claim of comparable performance is contradicted by its own Table 1. read the letter →

arxiv 1909.01144 v1 pith:O4MFDY4T submitted 2019-09-03 hep-ex physics.comp-phphysics.data-an

Fardin Syed , Riccardo Di Sipio , Pekka Sinervo This is my paper · ORCID

classification hep-exphysics.comp-phphysics.data-an

keywords topquarkreconstructionBidirectionalLSTMrecurrentneuralnetworkssemileptonicttbardecaykinematicMonteCarlosimulationjetassignmentmachinelearninginhigh-energyphysics

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

Top quarks live too briefly to be observed directly, so their momenta must be inferred from the muon, missing energy, and jets left behind. This paper shows that a deep neural network built around a Bidirectional Long Short-Term Memory (BLSTM) layer can perform that inference for semileptonic top-quark-pair events: the network takes a fixed 6-by-6 event summary and outputs the momentum vectors of the two b quarks, two W bosons, and two top quarks. Trained on about five million simulated 13 TeV proton-proton collisions with pileup, the network's parton-level kinematic distributions agree with the generator-level truth about as well as those from a benchmark chi-squared fit that searches over jet-to-parton permutations. The payoff is that a machine-learning reconstruction can match a standard algorithmic fitter without an explicit permutation search, and its flexibility opens a route to event topologies where such fits struggle.

What carries the argument

The load-bearing object is the event representation: a 6-by-6 matrix whose rows are the muon's momentum components, muon arrival time, missing transverse energy, missing-energy azimuthal angle, and, for up to five jets, the jet momentum components, energy, mass, and a binary b-tag flag. Missing jets are zero-padded. This matrix is processed by a Bidirectional Long Short-Term Memory layer, meaning a recurrent network cell with a memory state that scans the six columns both forward and backward, so the prediction for each object can depend on the whole event context rather than on a fixed jet order. The network outputs a 6-by-3 matrix of Cartesian momentum components for the hadronic and semileptonic b quarks, W bosons, and top quarks, with the top and W masses fixed at 172.5 and 80.4 GeV. Because the mapping is continuous and probabilistic, the network never has to choose a discrete jet-to-parton permutation, which is the step that a chi-squared fit must solve by enumeration.

What would settle it

Run the same training and evaluation on a detailed full detector simulation of the same $t\bar{t}$ sample, or on a large real-data sample whose parton-level kinematics are known from an independent high-precision reconstruction, and recompute the per-observable $\chi^2/\mathrm{DOF}$ comparisons of Table 1. If the BLSTM's agreement with truth degrades far more than the chi-squared-fit benchmark's does, for instance if the semileptonic top-$p_T$ $\chi^2/\mathrm{DOF}$ rises from 5.81 toward or above the benchmark's 67.08, then the claimed parity at parton level does not survive realistic detector conditions.

Watch

Extended reading notes

Core claim

On its own terms, the paper establishes a proof of feasibility: a BLSTM network with 329,913 trainable parameters, trained with a mean-squared-error loss on 90% of about five million simulated events and evaluated on the remaining 10%, maps detector-level inputs to parton-level four-momenta for all six decay products of the $t\bar{t}$ system in the muon+jets channel. The comparison with the chi-squared-fit benchmark is mixed. For the semileptonic top quark's transverse momentum the network is much closer to Monte Carlo truth ($\chi^2/\mathrm{DOF}=5.81$ versus 67.08), and it also improves the hadronic top's pseudorapidity (16.28 versus 471.56); for the b-quark transverse momenta it is much worse (1092--1497 versus 63--107), and for azimuthal angles the chi-squared fit is nearly perfect while the network develops 10--20% structures. The authors state the overarching result as comparability, and interpret the mixed pattern as evidence that machine-learning reconstruction is a viable route and that the architecture can be refined.

Load-bearing premise

The whole comparison rests on the simulated events, with their generator-level truth, parton shower, and fast detector simulation with pileup overlay, being faithful enough to real high-energy proton-proton collisions that a network trained on those events reconstructs parton-level kinematics in actual data at the same level of agreement.

Editorial extensions

If this is right

Semileptonic $t\bar{t}$ reconstruction becomes a regression problem rather than a combinatorial assignment problem, so events with a lost jet or an ambiguous b-tag are not automatically discarded.
The same BLSTM can be adapted to the boosted regime by replacing resolved jets with large-radius jets, since the input is a fixed-size array and no permutation logic has to change.
Users of such a network know which observables to trust: the paper's numbers say to trust top-quark and semileptonic W-boson kinematics and to distrust b-quark transverse momenta and azimuthal angles.
The architecture is flexible enough to accept new inputs, for example improved b-tagging scores or pileup-related variables, by adding or modifying layers without redesigning the reconstruction algorithm.

Reading between the lines

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

The shared $p_T$ underestimation between the BLSTM and the chi-squared fit suggests that the dominant resolution limit for inclusive top kinematics is the detector response itself, not the jet-assignment ambiguity; if so, further gains will come from better jet energy calibration rather than from a smarter fitter.
The network's failure on azimuthal angles is plausibly tied to its Cartesian input features: because rotations of the event around the beam axis are not encoded as an explicit symmetry, the network has to relearn them, and small angular correlations are hard to capture. A testable variant would add $\phi$, $\eta$, or relative angular coordinates and check whether the $\phi$ asymmetries vanish.
Because the output is continuous and the network has no built-in uncertainties, a natural next step left implicit by the paper is to add a variance head or an ensemble of BLSTMs to obtain per-event uncertainties, which would make the method usable in precision measurements.

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

Fardin and colleagues have put a BLSTM on a real problem — regressing all six parton four-momenta in t-tbar events from detector-level inputs — and that's new in the context of the cited baselines. The architecture is well-established, and the paper doesn't pretend otherwise. What it does do is give us a first look at how a recurrent model handles an inherently combinatorial task without explicit jet-permutation scoring. The training setup is described clearly, the code is promised (though the link is garbled), and the authors are candid about the network's weak spots.

The honest parts are also where the central claim falls apart. The abstract says agreement is 'comparable' to the chi2-fit benchmark. Table 1 shows that's only true for a subset of the top-quark and W-boson pT and eta variables. On azimuthal angles, BLSTM chi2/DOF sits between 35 and 273, while chi2-fit is below 1.2. On b-quark pT and eta, the BLSTM is 10-20 times worse. Section 6 admits the phi problem and the poor b-quark performance. So the 'comparable' claim is not supportable as stated; at best it's 'comparable on some variables, much worse on others.'

There are also technical issues. Eq. (5.2) defines chi2/NDF but never defines sigma_i^pred, and no uncertainties are given for the Table 1 entries. That weakens even the variable-subset claims. The benchmark is chi2-fit, not KLFitter, so the comparison is not to state-of-the-art. The code link is broken in the arxiv version. None of this is fatal — the paper is explicit about being a proof-of-concept — but each issue needs attention.

The MC-only evaluation (same generator for training and test) is a real limitation but not a fatal one; the network is learning a mapping, not fitting the generator, and the fixed masses are a reasonable simplification. I would have liked a separate event generator or a small closure test with different settings, but for a first demonstration it's acceptable.

Who is this for? Someone working on ML-based top reconstruction, or on combining deep recurrent models with physics constraints, will get a useful starting point and an honest set of failure modes. The paper deserves a serious referee — the idea is worth pursuing, and the overstatement is fixable. I'd send it out, but require a revised abstract, a defined comparison metric with uncertainties, and a working code link.

My recommendation: engage with it, treat the results as preliminary, and don't quote the 'comparable' claim without the qualifiers.

Referee Report

3 major / 5 minor

Summary. The paper introduces AngryTops, a BLSTM neural network trained on roughly five million MadGraph5+Pythia8 events with Delphes3 detector simulation to reconstruct the parton-level four-momenta of the top quarks, W bosons, and b quarks in the muon+jets ttbar decay channel. The network output is compared with a benchmark chi2-fit permutation method on 18 kinematic distributions using a chi2/DOF histogram metric. The abstract claims that the BLSTM's agreement with Monte Carlo predictions at parton level is comparable to that of the chi2-fit method. The paper also reports qualitative observations in Section 6 and concludes that machine-learning approaches can be competitive with standard reconstruction algorithms.

Significance. If the central claim were fully supported, the paper would demonstrate a viable machine-learning alternative to combinatorial kinematic fitting for top-quark reconstruction, with the practical advantage of handling variable jet multiplicities and not requiring explicit transfer functions. The strengths of the work include the public release of the AngryTops code, a clearly described simulation and training setup, and a direct comparison against a standard benchmark method. However, the evidence presented in Table 1 and the qualifying language in Section 6 show that the BLSTM is not competitive in several important distributions, so the significance as stated in the abstract is not established.

major comments (3)

[Abstract and Table 1] The central claim that the BLSTM's agreement with Monte Carlo predictions is "comparable" to the chi2-fit method is not supported by the paper's own quantitative results. In Table 1, for azimuthal-angle distributions the BLSTM chi2/DOF values range from 35.30 to 273.47, while the chi2-fit values are between 0.55 and 1.19; for b-quark transverse momentum distributions the BLSTM values are 1092-1497 versus 62-107 for the chi2-fit. Section 6 explicitly concedes that the chi2-fit "significantly outperforms AngryTops in the phi variable" and that b-quark kinematics are "perhaps the most poorly reconstructed observables." The abstract and the concluding claim of competitiveness therefore overstate the evidence presented in the manuscript.
[Section 5, Eq. (5.2)-(5.3)] Equation (5.2) defines chi2/NDF using a per-bin uncertainty sigma_i = sqrt((sigma_i^MC)^2 + (sigma_i^predicted)^2), but the quantity sigma_i^predicted is never defined anywhere in the paper. It is unclear whether this term represents the statistical uncertainty of the predicted histogram, a systematic uncertainty, or some other quantity. In addition, Table 1 reports chi2/DOF values without any uncertainties or information about the number of bins and the number of degrees of freedom, so the reader cannot assess whether the differences between the BLSTM and chi2-fit are statistically significant. A fair comparison requires a complete definition of the metric and appropriate error propagation.
[Sections 2 and 4] The network is trained and evaluated on events produced by the same Monte Carlo generator (MadGraph5+Pythia8) with the same Delphes3 detector simulation. Since the evaluation compares the reconstruction output with the parton-level quantities from this same simulation chain, the reported agreement is a self-consistency check on that simulation rather than a test against an independent benchmark or real data. The paper should explicitly state this limitation and temper the abstract's phrasing, which currently reads as a general statement about reconstruction performance without noting that the entire study is performed within a single simulation framework.

minor comments (5)

[Abstract and repository URL] The repository URL in the abstract, "yttps:@@gztyus.tom@IMFrruz@AngryTops", appears garbled and should be corrected to a proper HTTPS link.
[Section 1 and throughout] The paper describes the method as a "probabilistic reconstruction" (abstract and Section 1), but the network outputs deterministic point estimates of the four-momenta and no uncertainty or probability distribution is produced. The terminology should be clarified or the network should be augmented to output uncertainties.
[Section 3, Eq. (3.1)] The input matrix includes "muon arrival time of flight T0" as one of the six rows, but the text never explains how this quantity is defined or why it is used. If it is an important input, it deserves a definition; otherwise it may be removed.
[Section 6, first paragraph] The sentence "We first note that the chi2 comparisons are not particularly insightful" appears to conflict with the use of those same chi2/DOF values as the primary quantitative evidence in Table 1. The authors should either justify the metric or explain why it is not insightful while still using it as the basis for the comparison.
[References [13] and [14]] The references to Hyperopt and Tune give arXiv identifiers without full journal or version information; if the paper is intended for a journal, these should be completed.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the BLSTM reconstruction is a learned mapping trained on Monte Carlo truth labels, and the abstract's 'comparable' claim is a performance-claim issue, not a definitional or self-citation circularity.

full rationale

The paper's derivation chain is not circular. The BLSTM is trained to regress parton-level four-momenta from detector-level inputs (muon, jets, missing transverse energy, b-tagging) using Monte Carlo truth labels, which is a genuine supervised learning task rather than an algebraic rearrangement of its inputs. The benchmark χ2-fit is independently defined in Eq. (5.1) with fixed masses and hand-chosen widths, and the comparison metric in Eq. (5.2) is a standard histogram difference statistic; no parameter is fitted to the comparison data and then renamed as a prediction. There are no load-bearing self-citations: the paper cites standard tools (MadGraph5, Pythia8, Delphes3, KLFitter, etc.) as external software, not as the source of its own conclusions. The paper does contain evidence that the abstract's 'comparable' claim is overstated: Table 1 shows BLSTM χ2/DOF values of 35–273 for φ distributions versus 0.55–1.19 for χ2-fit, and 739–1497 for b-quark pT/η versus 62–707 for χ2-fit, and Section 6 concedes that χ2-fit 'significantly outperforms AngryTops in the φ variable' and that b-quark kinematics are 'perhaps the most poorly reconstructed observables by AngryTops.' These are correctness and presentation concerns, not circularity: the BLSTM result is not equivalent to its inputs by construction, nor is it forced by a self-citation chain. The fact that training and evaluation share the same Monte Carlo generator affects external validity but does not make the derivation circular, since the network output is a learned function, not a restatement of the training labels. The undefined σ_i^pred in Eq. (5.3) is an incompleteness in the comparison statistic, not a circular reduction. Overall, no circular step is present.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

No new physics entities are introduced. The central assumptions are the fidelity of the Monte Carlo and detector simulation used for training and evaluation, plus hand-chosen parameters in the selection and the benchmark objective.

free parameters (5)

sigma_t (chi2-fit objective) = 30 GeV
Hand-chosen width in the benchmark chi2 objective (Eq. 5.1), not learned from data.
sigma_W (chi2-fit objective) = 20 GeV
Hand-chosen width in the benchmark chi2 objective (Eq. 5.1).
pT selection threshold = 20 GeV
Cut applied to muons, W bosons, and b quarks in Section 2.
pseudorapidity selection threshold = 2.5
Cut applied to muons, W bosons, and b quarks in Section 2.
BLSTM network weights = 329,913 trainable parameters
Fitted to 4.5 million training events by minimizing MSE loss (Section 4). This is the learned mapping, not a physics constant.

assumptions (4)

domain assumption MadGraph5+Pythia8 with MLM matching correctly models ttbar production and decay at leading order.
Used to generate the training and test samples (Section 2). Any inaccuracy biases both training and evaluation.
domain assumption Delphes3 fast detector simulation with 25 pile-up collisions adequately represents the response of the ATLAS/CMS detectors.
Used to simulate detector resolution and pile-up (Section 2). Real-data performance depends on this fidelity.
ad hoc to paper The chi2/DOF histogram metric (Eq. 5.2) is a valid measure of reconstruction agreement.
Defined in Section 5 to compare methods; its validity as a ranking metric is not independently established.
domain assumption The top quark mass (172.5 GeV) and W boson mass (80.4 GeV) are fixed correctly in the output representation.
The network only predicts px, py, pz for each particle (Section 3), so the masses are assumed inputs.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Bidirectional Long Short-Term Memory (BLSTM) neural networks for reconstruction of top-quark pair decay kinematics." pith.science (2026). https://pith.science/paper/O4MFDY4T

@misc{pith2026190901144,
  author       = {Pith},
  title        = {Pith review of: Bidirectional Long Short-Term Memory (BLSTM) neural networks for reconstruction of top-quark pair decay kinematics},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/O4MFDY4T}},
  note         = {Machine review of arXiv:1909.01144}
}

abstract

A probabilistic reconstruction using machine-learning of the decay kinematics of top-quark pairs produced in high-energy proton-proton collisions is presented. A deep neural network whose core consists of a Bidirectional Long Short-Term Memory (BLSTM) is trained to infer the four-momenta of the two top quarks produced in the hard scattering process. The MadGraph5+Pythia8 Monte Carlo event generator is used to create a sample of top-quark pairs decaying in the $\mu$+jets channel, whose final-state objects are used to create the input to the deep neural network. Distortions due to limited resolution of the experimental apparatus are simulated with the Delphes3 fast detector simulator. The level of agreement between the Monte Carlo predictions and the BLSTM for kinematic distributions at parton level is comparable to that obtained using a benchmark method that finds the jet permutation that minimizes an objective function.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

14 extracted references · 10 canonical work pages

[1]

Top quark physics at hadron colliders,

A. Quadt, “Top quark physics at hadron colliders,”Eur. Phys. J., vol. C48, pp. 835–1000, 2006

work page 2006
[2]

A likelihood-based reconstruction algorithm for top-quark pairs and the KLFitter framework,

J. Erdmann, S. Guindon, K. Kroeninger, B. Lemmer, O. Nackenhorst, A. Quadt, and P. Stolte, “A likelihood-based reconstruction algorithm for top-quark pairs and the KLFitter framework,”Nucl. Instrum. Meth., vol. A748, pp. 18–25, 2014

work page 2014
[3]

Study of methods of resolved top quark reconstruction in semileptonict¯t decay,

J. Kvita, “Study of methods of resolved top quark reconstruction in semileptonict¯t decay,”Nucl. Instrum. Meth., vol. A900, pp. 84–100, 2018

work page 2018
[4]

Top quark mass measurement using the template method in the lepton + jets channel at CDF II,

A. Abulenciaet al., “Top quark mass measurement using the template method in the lepton + jets channel at CDF II,”Phys. Rev., vol. D73, p. 032003, 2006

work page 2006
[5]

The automated computation of tree-level and next-to-leading order diﬀerential cross sections, and their matching to parton shower simulations,

J. Alwall, R. Frederix, S. Frixione, V. Hirschi, F. Maltoniet al., “The automated computation of tree-level and next-to-leading order diﬀerential cross sections, and their matching to parton shower simulations,”JHEP, vol. 07, p. 079, 2014

work page 2014
[6]

An Introduction to PYTHIA 8.2,

T. Sjöstrand, S. Ask, J. R. Christiansen, R. Corke, N. Desai, P. Ilten, S. Mrenna, S. Prestel, C. O. Rasmussen, and P. Z. Skands, “An Introduction to PYTHIA 8.2,”Comput. Phys. Commun., vol. 191, p. 159, 2015

work page 2015
[7]

A New approach to multijet calculations in hadron collisions,

F. Caravaglios, M. L. Mangano, M. Moretti, and R. Pittau, “A New approach to multijet calculations in hadron collisions,”Nucl. Phys., vol. B539, pp. 215–232, 1999

work page 1999
[8]

The anti-kt jet clustering algorithm,

M. Cacciari, G. P. Salam, and G. Soyez, “The anti-kt jet clustering algorithm,”JHEP, vol. 04, p. 063, 2008

work page 2008

Show all 14 references

[9]

FastJet User Manual,

——, “FastJet User Manual,”Eur. Phys. J. C, vol. 72, p. 1896, 2012

2012
[10]

Cholletet al., “Keras,” https://keras.io, 2015

F. Cholletet al., “Keras,” https://keras.io, 2015

2015
[11]

TensorFlow: Large-scale machine learning on heterogeneous systems,

M. Abadi, A. Agarwal, P. Barham, E. Brevdo, Z. Chen, C. Citro, G. S. Corrado, A. Davis, J. Dean, M. Devin, S. Ghemawat, I. Goodfellow, A. Harp, G. Irving, M. Isard, Y. Jia, R. Jozefowicz, L. Kaiser, M. Kudlur, J. Levenberg, D. Mané, R. Monga, S. Moore, D. Murray, C. Olah, M. S...

2015
[12]

Adam: A method for stochastic optimization,

D. P. Kingma and J. Ba, “Adam: A method for stochastic optimization,” https://arXiv.org/1412.6980, 2014

2014 arXiv
[13]

Making a science of model search,

J. Bergstra, D. Yamins, and D. D. Cox, “Making a science of model search,” arXiv:1209.5111, 2012

2012 arXiv
[14]

Tune: A research platform for distributed model selection and training,

R. Liaw, E. Liang, R. Nishihara, P. Moritz, J. E. Gonzalez, and I. Stoica, “Tune: A research platform for distributed model selection and training,” arXiv:1807.05118, 2018. – 13 –

2018 arXiv

Pith tools

Reviewed August 14, 2026 · model on record in the stance chip above.

[1] [1]

Top quark physics at hadron colliders,

A. Quadt, “Top quark physics at hadron colliders,”Eur. Phys. J., vol. C48, pp. 835–1000, 2006

work page 2006

[2] [2]

A likelihood-based reconstruction algorithm for top-quark pairs and the KLFitter framework,

J. Erdmann, S. Guindon, K. Kroeninger, B. Lemmer, O. Nackenhorst, A. Quadt, and P. Stolte, “A likelihood-based reconstruction algorithm for top-quark pairs and the KLFitter framework,”Nucl. Instrum. Meth., vol. A748, pp. 18–25, 2014

work page 2014

[3] [3]

Study of methods of resolved top quark reconstruction in semileptonict¯t decay,

J. Kvita, “Study of methods of resolved top quark reconstruction in semileptonict¯t decay,”Nucl. Instrum. Meth., vol. A900, pp. 84–100, 2018

work page 2018

[4] [4]

Top quark mass measurement using the template method in the lepton + jets channel at CDF II,

A. Abulenciaet al., “Top quark mass measurement using the template method in the lepton + jets channel at CDF II,”Phys. Rev., vol. D73, p. 032003, 2006

work page 2006

[5] [5]

The automated computation of tree-level and next-to-leading order diﬀerential cross sections, and their matching to parton shower simulations,

J. Alwall, R. Frederix, S. Frixione, V. Hirschi, F. Maltoniet al., “The automated computation of tree-level and next-to-leading order diﬀerential cross sections, and their matching to parton shower simulations,”JHEP, vol. 07, p. 079, 2014

work page 2014

[6] [6]

An Introduction to PYTHIA 8.2,

T. Sjöstrand, S. Ask, J. R. Christiansen, R. Corke, N. Desai, P. Ilten, S. Mrenna, S. Prestel, C. O. Rasmussen, and P. Z. Skands, “An Introduction to PYTHIA 8.2,”Comput. Phys. Commun., vol. 191, p. 159, 2015

work page 2015

[7] [7]

A New approach to multijet calculations in hadron collisions,

F. Caravaglios, M. L. Mangano, M. Moretti, and R. Pittau, “A New approach to multijet calculations in hadron collisions,”Nucl. Phys., vol. B539, pp. 215–232, 1999

work page 1999

[8] [8]

The anti-kt jet clustering algorithm,

M. Cacciari, G. P. Salam, and G. Soyez, “The anti-kt jet clustering algorithm,”JHEP, vol. 04, p. 063, 2008

work page 2008