Pith. sign in

REVIEW 3 major objections 5 minor 39 references

Theory-informed neural networks for particle physics

T0 review · 3 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read A reinforcement-learning agent that maps final-state particles to partons by maximising the tree-level matrix element can reconstruct LHC events, tag W polarisation, and flag anomalies without labelled data.

desk verdict Clever DQN/ME parton-assignment pipeline that's novel and plausible, but the self-referential validation and missing assignment-accuracy metrics undercut the strong claims. read the letter →

arxiv 2507.13447 v1 pith:532PIHYJ submitted 2025-07-17 hep-ph hep-exhep-th

Barry M. Dillon , Michael Spannowsky This is my paper
classification hep-phhep-exhep-th
keywords reinforcementlearningdeepQ-networkmatrixelementmethodpartonassignmenteventreconstructionLHCphysicsanomalydetectionWbosonpolarization
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 argues that the hard combinatorial problem of assigning measured particles to the partons of a hypothesized scattering process can be solved without any labelled training data. The agent is a transformer-based Deep Q-Network that repeatedly swaps particle-to-parton assignments, receiving a reward equal to the logarithmic change in the exact tree-level matrix element after each swap. Because the reward comes from first-principles theory, the learned policy is interpretable and every reconstructed particle can be traced to a partonic origin. On simulated $t\bar{t}$, $t\bar{t}W$, and $t\bar{t}t\bar{t}$ events the reconstructed intermediate masses recover the true distributions, and the same machinery builds a longitudinal $WW$ polarisation tagger and a theory-informed anomaly score. If right, this unifies event reconstruction, tagging, and anomaly detection in one transparent pipeline for the High-Luminosity LHC.

What carries the argument

The load-bearing mechanism is the theory-informed reward: at each step of the Markov decision process the agent chooses a swap (or 'do nothing') action and receives $r_t = \log(m_{t+1}/m_t)$, the logarithmic change in the tree-level matrix element. A permutation-invariant transformer with a positional encoding for the parton identities approximates the Q-function via the Bellman equation, and the policy greedily selects the action with the highest Q-value over $T=N-3$ steps. This reward is what makes the approach label-free and interpretable: the network learns only the combinatorial mapping, while the matrix element carries all the physics, including spin correlations and interference effects.

What would settle it

Take the same trained agents and compare their final particle-to-parton assignments, event by event, against Monte Carlo truth labels rather than only comparing reconstructed mass peaks; if the agreement is no better than the best random assignment, or if the policy systematically prefers non-truth assignments in events where the matrix-element-maximising mapping differs from truth, the label-free reconstruction claim is refuted.

Watch

Extended reading notes

Core claim

The paper's central claim is that the mapping from final-state particles to matrix-element partons can be treated as an optimal-control problem whose reward is the exact tree-level matrix element itself. The agent's policy is obtained purely from physics: each action changes the assignment, and the reward $r_t = \log(m_{t+1}/m_t)$ tells the agent whether the new assignment is more consistent with the theory hypothesis. Once trained, the policy has no labels, no handcrafted observables, and no black-box decisions: the network only provides the permutation, while all classification and anomaly scores are computed from the analytic matrix element, so the scores respect the symmetries of the theory. The authors validate this on $t\bar{t}$, $t\bar{t}W$, and $t\bar{t}t\bar{t}$ reconstruction, on longitudinal $W^+W^-$ tagging, and on $t\bar{t}$-background anomaly detection, finding robust performance that scales with combinatorial complexity.

Load-bearing premise

The method's reward assumes that the assignment with the largest tree-level matrix element for the assumed process is the physical assignment, and the authors note this is not always the true assignment; if this mismatch is frequent, the reconstructed topology and all scores built on it inherit the error.

Editorial extensions

If this is right

  • Event reconstruction, polarisation tagging, and anomaly detection become one pipeline: the same trained policy supplies the particle-to-parton mapping, and the matrix element supplies the score.
  • No labels or simulated 'truth' are needed for training, so the method avoids simulation-dependent biases of supervised classifiers.
  • Because scores are computed from the matrix element, they automatically respect Lorentz invariance and the physical symmetries of the process.
  • The approach scales to large combinatorial complexity: reconstruction quality remains high from 720 mappings in $t\bar{t}$ to about $4\times10^8$ in $t\bar{t}t\bar{t}$.
  • The longitudinal $W^+W^-$ tagger and the $t\bar{t}$-background anomaly score outperform a plain AutoEncoder on the restricted-mass test, while retaining interpretability of which partons form which resonance.

Reading between the lines

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

  • Beyond the paper, the reward could be upgraded to include higher-order corrections or detector response, which would make the same assignment policy applicable to reconstructed objects rather than parton-level four-momenta.
  • A testable extension is to compare the agent's final assignment against parton-shower truth labels event-by-event; the paper only shows mass distributions, so an explicit assignment-accuracy number would sharpen the label-free claim.
  • Because the matrix-element-maximising assignment is not always the true one, the method's practical ceiling depends on how often the two agree in a given process; quantifying that overlap would predict where the approach will fail.
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.

Referee Report

3 major / 5 minor

Summary. The paper proposes a theory-informed reinforcement-learning framework for assigning final-state particles to partons in hadron collider events. A transformer-based Deep Q-Network is trained with a reward equal to the logarithmic change in the tree-level matrix element, and the resulting policy produces a parton-level event graph. The method is demonstrated on parton-level $t\bar{t}$, $t\bar{t}W$, and $t\bar{t}t\bar{t}$ events, with reconstructed mass peaks shown for intermediate $W$ bosons and top quarks. The same pipeline is then used to construct a $W^+W^-$ polarization tagger via the likelihood ratio in Eq. (7) and an anomaly detector based on the inverse background matrix element in Eq. (8). The central claims are that the method is label-free, fully interpretable, robust across processes, and naturally respects physical symmetries.

Significance. If the central claims are substantiated, the framework would be a useful contribution: it unifies event reconstruction, polarization tagging, and anomaly detection in a single theory-driven pipeline, and it makes the assignment step interpretable by construction. The use of matrix-element rewards rather than labels is a genuine conceptual step, and the demonstration on three processes of increasing combinatorial complexity is a strength. The paper also ships a public code repository, which is valuable for reproducibility. However, the evidence currently falls short of the claims. The validation never directly measures whether the ME-maximizing assignment coincides with the true parton-level assignment, and the claimed classifier performance is not quantified with AUCs or rejection factors. The parton-level approximation is acknowledged, but the 'robust performance' claim is supported only by qualitative histograms. The central idea is plausible and the manuscript is within the scope of the journal, but the missing quantitative validation is load-bearing for the paper's main conclusions.

major comments (3)
  1. [Section 3 and Fig. 3] The core validation compares the matrix element of the predicted assignment with the matrix element of the true assignment (right panels of Fig. 3). This is not an independent test of reconstruction: because the policy is trained to maximize the matrix element, the predicted ME will be close to the true ME whenever the true assignment is near the ME maximum, even if many individual partons are mismatched. The paper itself states in Section 3 that the ME-maximizing assignment is "not always the true assignment," and Section 4 notes that $t\bar{t}W$ and $t\bar{t}t\bar{t}$ events suffer from incorrect W-b-quark matching. The manuscript therefore needs a direct quantitative measure of assignment correctness, for example the fraction of events in which all parton assignments are correct, or per-parton matching accuracies, for each process. Until this is reported, the claims that the policy is "fully interpretable" and that Eq. (7) is optimal in the ideal scenario are not supported.
  2. [Section 5, Figs. 8 and 9] The proposed classifiers are shown only as ROC curves, with no reported AUC for the RL-based W-polarization tagger or for the theory-informed anomaly detector. The text quotes AUC 0.904 for the likelihood ratio computed with true assignments and AUC 0.645 for the autoencoder, but does not quote the corresponding numbers for the proposed method. To support "robust performance" and "powerful anomaly-detection performance," please report AUC and background-rejection factors, with statistical uncertainties, for the RL-based scores in both applications. This is particularly important because the theoretical optimality of Eq. (7) holds only for the correct particle-parton mapping.
  3. [Section 4 and Abstract] The claim that the method "maintains robust performance across all processes" is supported only by qualitative mass histograms. There is no per-process quantitative metric, such as reconstruction efficiency with a defined matching criterion, the fraction of correctly assigned events, or the widths of reconstructed mass peaks relative to the truth distribution, and no uncertainty estimate. Given the acknowledged parton-level approximation and the absence of detector effects, "robust performance" overstates the evidence presented. Please either add quantitative benchmarks for reconstruction accuracy or soften the claim in the abstract and Section 4.
minor comments (5)
  1. [Section 2, Eq. (1)] The text writes "P rt" where a summation is intended; please correct the notation and define the sum over time steps explicitly.
  2. [Section 2, training procedure] The sentence "The ensure that the agent is able to learn from a diverse range of state-action combinations" contains a typo and should read "To ensure."
  3. [Fig. 3] The right panels are described as histograms, but the axis labels and the meaning of the dashed line ("initial mapping") are not defined in the caption; please clarify.
  4. [Eq. (6) and Eq. (8)] The log-ratio reward in Eq. (6) and the inverse matrix-element score in Eq. (8) are undefined if the matrix element vanishes; since tree-level matrix elements can vanish at phase-space boundaries, please state how such cases are handled in the implementation.
  5. [Section 5, anomaly detection] The notation "$\epsilon_b^{-1}(\epsilon_s=0.3) \simeq 8$" is not standard and lacks a space; please define the background-rejection notation and explain how the autoencoder's ROC curve is evaluated.

Circularity Check

1 steps flagged · score 2.0 of 10

Minor self-referential validation: the matrix-element reward and the matrix-element/mass-peak evaluation metrics are the same quantity, but the central RL optimization still has independent content.

  1. self definitional [Section 3 'The reward' (Eq. 6) and Section 4 'Theory-informed event reconstruction' (Figs. 3–6)]
    "The reward for this action is defined as the log-ratio of the two matrix elements rt = log mt+1 mt (6) ... In Fig. 3 we show ... the relative difference between the matrix element of the predicted mappings and the true matrix element. ... we see very good agreement between the matrix elements of the predicted mappings and the true mappings."

    The RL agent's cumulative reward telescopes to log(m_final/m_initial), so the policy is explicitly trained to maximize the final matrix element. The validation in Fig. 3 then compares the matrix element of the predicted assignment with that of the true assignment. Because the predicted assignment is chosen to maximize the matrix element, ME(predicted) >= ME(true) by construction; close agreement of these two values therefore mostly reflects successful optimization of the training objective, not necessarily agreement of the assignments. Similarly, the mass-peak validations in Figs. 4–6 use the W and top propagator masses that are already encoded in the same tree-level matrix element used as the reward.

full rationale

The paper is not fundamentally circular: the core contribution is an RL policy that learns to search the combinatorial assignment space using a physics-based reward, and the policy optimization itself is a legitimate, nontrivial computation. The classifier and anomaly-detection demonstrations are evaluated against external objectives (ROC curves against true polarization labels and against a separately trained autoencoder), which provides independent content. The self-referential aspect is confined to the event-reconstruction validation: the reward is the matrix element, and the primary reconstruction-quality metrics are also matrix-element values and resonant masses that are already built into that same matrix element. This is a genuine but moderate issue, not a collapse of the derivation into its inputs. The many self-citations in the paper are to prior autoencoder and interpretability work used for comparison or context, not as load-bearing justification for the new method, so they do not raise the score further. Overall, the central claims are not forced by definition or by a self-citation chain, and the method retains independent computational and phenomenological content; hence a score of 2 is appropriate.

Assumptions & free parameters 1 free parameters · 4 assumptions · 0 invented entities

The method introduces no new physical entities. Its central assumptions are that the tree-level matrix element is the correct objective and that parton-level simulation is representative. The only manually tuned parameter is the discount factor, which is a standard RL hyperparameter and does not affect the physics conclusion.

free parameters (1)
  • discount factor gamma = 0.5 (ttbar, WW), 0.01 (ttbarW, ttbarttbar)
    Hand-tuned per process to ensure convergence; the paper states results were not sensitive to exact values.
assumptions (4)
  • domain assumption The tree-level matrix element is the appropriate reward for optimal parton assignment.
    Section 3 defines rt = log(m_{t+1}/m_t) and states the optimal assignment maximizes the matrix element; this is a modelling choice about what is 'correct'.
  • domain assumption Parton-level approximation (no showering, hadronization, detector effects) is sufficient for validation.
    Section 3, paragraph on states and actions: 'We will ignore any showering, hadronisation, or detector effects.' The paper's LHC-applicability claims rest on this approximation.
  • standard math MadGraph provides exact tree-level matrix elements for the reward.
    Section 4: 'We also use MadGraph to compute the matrix elements used in the rewards during training.'
  • standard math The state transition is deterministic and Markov, so Q-learning applies.
    Section 2 defines the MDP; swapping actions are deterministic, so the Bellman equation is valid.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Theory-informed neural networks for particle physics." pith.science (2026). https://pith.science/paper/532PIHYJ

@misc{pith2026250713447,
  author       = {Pith},
  title        = {Pith review of: Theory-informed neural networks for particle physics},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/532PIHYJ}},
  note         = {Machine review of arXiv:2507.13447}
}
abstract

We present a theory-informed reinforcement-learning framework that recasts the combinatorial assignment of final-state particles in hadron collider events as a Markov decision process. A transformer-based Deep Q-Network, rewarded at each step by the logarithmic change in the tree-level matrix element, learns to map final-state particles to partons. Because the reward derives solely from first-principles theory, the resulting policy is label-free and fully interpretable, allowing every reconstructed particle to be traced to a definite partonic origin. The method is validated on event reconstruction for $t\bar{t}$, $t\bar{t}W$, and $t\bar{t}t\bar{t}$ processes at the Large Hadron Collider. The method maintains robust performance across all processes, demonstrating its scaling with increasing combinatorial complexity. We demonstrate how this method can be used to build a theory-informed classifier for effective discrimination of longitudinal $W^{+}W^{-}$ pairs, and show that we can construct theory-informed anomaly-detection tools using background process matrix elements. Building on theoretical calculations, this method offers a transparent alternative to black-box classifiers. Being built on the matrix element, the classification and anomaly scores naturally respect all physical symmetries and are much less susceptible to the implicit biases common to other methods. Thus, it provides a framework for precision measurements, hypothesis testing, and anomaly searches at the High-Luminosity LHC.

Figures

Figures reproduced from arXiv: 2507.13447 by the authors.

Figure 1
Figure 1. Outline of the general training procedure in Deep Q-Learning. [PITH_FULL_IMAGE:figures/full_fig_p006_1.png] view at source ↗
Figure 2
Figure 2. Outline of the Theory-Informed Neural-Network approach to Deep Q [PITH_FULL_IMAGE:figures/full_fig_p009_2.png] view at source ↗
Figure 3
Figure 3. The left plots show the increase in the total reward in each episode over the [PITH_FULL_IMAGE:figures/full_fig_p010_3.png] view at source ↗
Figures from the paper (6 more)
Figure 4
Figure 4. Figure 4: Histogram showing the invariant masses of the reconstructed [PITH_FULL_IMAGE:figures/full_fig_p011_4.png]
Figure 5
Figure 5. Figure 5: Histogram showing the invariant masses of the reconstructed [PITH_FULL_IMAGE:figures/full_fig_p012_5.png]
Figure 6
Figure 6. Figure 6: Histogram showing the invariant masses of the reconstructed [PITH_FULL_IMAGE:figures/full_fig_p013_6.png]
Figure 7
Figure 7. Figure 7: Histogram showing the invariant masses of the reconstructed [PITH_FULL_IMAGE:figures/full_fig_p014_7.png]
Figure 8
Figure 8. Figure 8: ROC curve showing WW polarisation tagging performance with the signal [PITH_FULL_IMAGE:figures/full_fig_p014_8.png]
Figure 9
Figure 9. Figure 9: ROC curve showing anomaly-detection performance with a [PITH_FULL_IMAGE:figures/full_fig_p015_9.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

39 extracted references · 11 canonical work pages

  1. [1]

    Dynamical Likelihood Method for Reconstruction of Events with Missing Momentum. I. Method and Toy Models

    K. Kondo. “Dynamical Likelihood Method for Reconstruction of Events with Missing Momentum. I. Method and Toy Models”. In:J. Phys. Soc. Jap.57 (1988), pp. 4126–4140. doi: 10.1143/JPSJ.57.4126

    doi:10.1143/jpsj.57.4126 1988

  2. [2]

    A precision measurement of the mass of the top quark

    V.M. Abazov et al. “A precision measurement of the mass of the top quark”. In: Nature 429 (2004), pp. 638–642. doi: 10.1038/nature02589

    work page doi:10.1038/nature02589 2004

  3. [3]

    Finding physics signals with shower deconstruction

    Davison E. Soper and Michael Spannowsky. “Finding physics signals with shower deconstruction”. In: Phys. Rev. D84 (2011), p. 074002. doi: 10.1103/PhysRevD. 84.074002. arXiv: 1102.3480 [hep-ph]

    arXiv 2011

  4. [4]

    Finding top quarks with shower deconstruction

    Davison E. Soper and Michael Spannowsky. “Finding top quarks with shower deconstruction”. In: Phys. Rev. D87 (2013), p. 054012. doi: 10.1103/PhysRevD. 87.054012. arXiv: 1211.3140 [hep-ph]

    arXiv 2013

  5. [5]

    Finding physics signals with event deconstruction

    Davison E. Soper and Michael Spannowsky. “Finding physics signals with event deconstruction”. In: Phys. Rev. D 89.9 (2014), p. 094005. doi: 10 . 1103 / PhysRevD.89.094005. arXiv: 1402.1189 [hep-ph]

    work page Pith review arXiv 2014

  6. [6]

    Determining the Structure of Higgs Couplings at the LHC

    Tilman Plehn, David L. Rainwater, and Dieter Zeppenfeld. “Determining the Structure of Higgs Couplings at the LHC”. In: Phys. Rev. Lett. 88 (2002), p. 051801. doi: 10.1103/PhysRevLett.88.051801. arXiv: hep-ph/0105325

    arXiv 2002

  7. [7]

    Weighing Wimps with Kinks at Colliders: Invisible Particle Mass Measurements from Endpoints

    Alan J. Barr, Ben Gripaios, and Christopher G. Lester. “Weighing Wimps with Kinks at Colliders: Invisible Particle Mass Measurements from Endpoints”. In: JHEP 02 (2008), p. 014. doi: 10 . 1088 / 1126 - 6708 / 2008 / 02 / 014. arXiv: 0711.4008 [hep-ph]

    work page Pith review arXiv 2008

  8. [8]

    Jet substructure as a new Higgs search channel at the LHC

    Jonathan M. Butterworth et al. “Jet substructure as a new Higgs search channel at the LHC”. In: Phys. Rev. Lett. 100 (2008), p. 242001. doi: 10 . 1103 / PhysRevLett.100.242001. arXiv: 0802.2470 [hep-ph]

    arXiv 2008

Show all 39 references
  1. [9]

    Fat Jets for a Light Higgs

    Tilman Plehn, Gavin P. Salam, and Michael Spannowsky. “Fat Jets for a Light Higgs”. In: Phys. Rev. Lett.104 (2010), p. 111801. doi: 10.1103/PhysRevLett. 104.111801. arXiv: 0910.5472 [hep-ph]

    2010 arXiv

  2. [10]

    Playing Tag with ANN: Boosted Top Identification with Pattern Recognition

    Leandro G. Almeida et al. “Playing Tag with ANN: Boosted Top Identification with Pattern Recognition”. In: JHEP 07 (2015), p. 086. doi: 10 . 1007 / JHEP07(2015)086. arXiv: 1501.05968 [hep-ph]

    2015 arXiv

  3. [11]

    Jet-Images – Deep Learning Edition

    Luke de Oliveira et al. “Jet-Images – Deep Learning Edition”. In: JHEP 07 (2016), p. 069. doi: 10.1007/JHEP07(2016)069. arXiv: 1511.05190 [hep-ph]

    2016 arXiv

  4. [12]

    Parameterized neural networks for high-energy physics

    Pierre Baldi et al. “Parameterized neural networks for high-energy physics”. In: Eur. Phys. J. C76.5 (2016), p. 235. doi: 10.1140/epjc/s10052- 016- 4099- 4. arXiv: 1601.07913 [hep-ex]

    2016 arXiv

  5. [13]

    Mapping Machine-Learned Physics into a Human-Readable Space

    Taylor Faucett, Jesse Thaler, and Daniel Whiteson. “Mapping Machine-Learned Physics into a Human-Readable Space”. In: Phys. Rev. D103.3 (2021), p. 036020. doi: 10.1103/PhysRevD.103.036020. arXiv: 2010.11998 [hep-ph]. REFERENCES 19

    2021 arXiv

  6. [14]

    On the Problem of the Most Efficient Tests of Statistical Hypotheses

    J. Neyman and E. S. Pearson. “On the Problem of the Most Efficient Tests of Statistical Hypotheses”. In: Philosophical Transactions of the Royal Society of London Series A231 (Jan. 1933), pp. 289–337. doi: 10.1098/rsta.1933.0009

    1933

  7. [15]

    Interpretable deep learning models for the inference and classification of LHC data

    Vishal S. Ngairangbam and Michael Spannowsky. “Interpretable deep learning models for the inference and classification of LHC data”. In: JHEP 05 (2024), p. 004. doi: 10.1007/JHEP05(2024)004. arXiv: 2312.12330 [hep-ph]

    2024 arXiv

  8. [16]

    Playing Atari with Deep Reinforcement Learning

    Volodymyr Mnih et al. “Playing Atari with Deep Reinforcement Learning”. In: arXiv preprint arXiv:1312.5602 (2013). url: https://arxiv.org/abs/1312. 5602

    2013 arXiv

  9. [17]

    Human-level control through deep reinforcement learning

    Volodymyr Mnih et al. “Human-level control through deep reinforcement learning”. In: Nature 518.7540 (2015), pp. 529–533. doi: 10.1038/nature14236

    2015 doi

  10. [18]

    Q-learning

    Christopher Watkins and Peter Dayan. “Q-learning”. In: Machine Learning 8.3 (1992), pp. 279–292. issn: 1573-0565. doi: 10.1007/BF00992698

    1992 doi

  11. [19]

    Sutton and Andrew G

    Richard S. Sutton and Andrew G. Barto. Reinforcement Learning: An Introduction. Second. The MIT Press, 2018. isbn: 978-0262039246. url: http: //incompleteideas.net/book/the-book-2nd.html

    2018

  12. [20]

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

    J. Alwall et al. “The automated computation of tree-level and next-to-leading order differential cross sections, and their matching to parton shower simulations”. In: JHEP 07 (2014), p. 079. doi: 10.1007/JHEP07(2014)079. arXiv: 1405.0301 [hep-ph]

    2014 arXiv

  13. [21]

    Auto-Encoding Variational Bayes

    Diederik P Kingma and Max Welling. “Auto-Encoding Variational Bayes”. In: (2014). arXiv: 1312.6114 [stat.ML]

    2014 arXiv

  14. [22]

    Searching for New Physics with Deep Autoencoders

    Marco Farina, Yuichiro Nakai, and David Shih. “Searching for New Physics with Deep Autoencoders”. In: Phys. Rev. D 101 (2020), p. 075021. doi: 10 . 1103 / PhysRevD.101.075021. arXiv: 1808.08992 [hep-ph]

    2020 arXiv

  15. [23]

    QCD or What?

    Theo Heimel et al. “QCD or What?” In: SciPost Phys. 6 (2019), p. 030. doi: 10.21468/SciPostPhys.6.3.030. arXiv: 1808.08979 [hep-ph]

    2019 arXiv

  16. [24]

    Adversarially-trained autoencoders for robust unsupervised new physics searches

    Andrew Blance, Michael Spannowsky, and Philip Waite. “Adversarially-trained autoencoders for robust unsupervised new physics searches”. In: JHEP 10 (2019), p. 047. doi: 10.1007/JHEP10(2019)047. arXiv: 1905.10384 [hep-ph]

    2019 arXiv

  17. [25]

    Anomaly Detection under Coordinate Transformations

    Gregor Kasieczka et al. “Anomaly Detection under Coordinate Transformations”. In: (Sept. 2022). arXiv: 2209.06225 [hep-ph]

    2022 arXiv

  18. [26]

    Solving differential equations with neural networks: Applications to the calculation of cosmological phase transitions

    Maria Laura Piscopo, Michael Spannowsky, and Philip Waite. “Solving differential equations with neural networks: Applications to the calculation of cosmological phase transitions”. In: Phys. Rev. D 100.1 (2019), p. 016002. doi: 10.1103/PhysRevD.100.016002. arXiv: 1902.05563 [hep-ph]

    2019 arXiv

  19. [27]

    Symmetries, safety, and self-supervision

    Barry M. Dillon et al. “Symmetries, safety, and self-supervision”. In: SciPost Phys. 12.6 (2022), p. 188. doi: 10 . 21468 / SciPostPhys . 12 . 6 . 188. arXiv: 2108.04253 [hep-ph]. REFERENCES 20

    2022 arXiv

  20. [28]

    Self-supervised anomaly detection for new physics

    Barry M. Dillon, Radha Mastandrea, and Benjamin Nachman. “Self-supervised anomaly detection for new physics”. In: Phys. Rev. D 106.5 (2022), p. 056005. doi: 10.1103/PhysRevD.106.056005. arXiv: 2205.10380 [hep-ph]

    2022 arXiv

  21. [29]

    Anomalies, representations, and self-supervision

    Barry M. Dillon et al. “Anomalies, representations, and self-supervision”. In: SciPost Phys. Core7 (2024), p. 056. doi: 10.21468/SciPostPhysCore.7.3.056. arXiv: 2301.04660 [hep-ph]

    2024 arXiv

  22. [30]

    A normalized autoencoder for LHC triggers

    Barry M. Dillon et al. “A normalized autoencoder for LHC triggers”. In: SciPost Phys. Core 6 (2023), p. 074. doi: 10.21468/SciPostPhysCore.6.4.074. arXiv: 2206.14225 [hep-ph]

    2023

  23. [31]

    IRC-Safe Graph Autoencoder for Unsupervised Anomaly Detection

    Oliver Atkinson et al. “IRC-Safe Graph Autoencoder for Unsupervised Anomaly Detection”. In: Front. Artif. Intell. 5 (2022), p. 943135. doi: 10 . 3389 / frai . 2022.943135. arXiv: 2204.12231 [hep-ph]

    2022

  24. [32]

    Anomaly Awareness

    Charanjit Kaur Khosa and Veronica Sanz. “Anomaly Awareness”. In: SciPost Physics 15.2 (Aug. 2023). issn: 2542-4653. doi: 10.21468/scipostphys.15.2

    2023 doi

  25. [33]

    Uncovering latent jet substructure

    Barry M. Dillon, Darius A. Faroughy, and Jernej F. Kamenik. “Uncovering latent jet substructure”. In: Phys. Rev. D 100 (2019), p. 056002. doi: 10 . 1103 / PhysRevD.100.056002. arXiv: 1904.04200 [hep-ph]

    2019 arXiv

  26. [34]

    Learning the latent structure of collider events

    B. M. Dillon et al. “Learning the latent structure of collider events”. In: JHEP 10 (2020), p. 206. doi: 10.1007/JHEP10(2020)206. arXiv: 2005.12319 [hep-ph]

    2020 arXiv

  27. [35]

    Better Latent Spaces for Better Autoencoders

    Barry M. Dillon et al. “Better Latent Spaces for Better Autoencoders”. In: SciPost Phys. 11 (2021), p. 061. doi: 10.21468/SciPostPhys.11.3.061 . arXiv: 2104. 08291 [hep-ph]

    2021 doi

  28. [36]

    Creating simple, interpretable anomaly detectors for new physics in jet substructure

    Layne Bradshaw, Spencer Chang, and Bryan Ostdiek. “Creating simple, interpretable anomaly detectors for new physics in jet substructure”. In: Phys. Rev. D 106.3 (2022), p. 035014. doi: 10.1103/PhysRevD.106.035014 . arXiv: 2203.01343 [hep-ph]

    2022 arXiv

  29. [37]

    Generator Based Inference (GBI)

    Chi Lung Cheng et al. “Generator Based Inference (GBI)”. In: (May 2025). arXiv: 2506.00119 [hep-ph]

    2025 arXiv

  30. [38]

    Lorentz group equivariant autoencoders

    Zichun Hao et al. “Lorentz group equivariant autoencoders”. In: The European Physical Journal C 83.6 (June 2023). issn: 1434-6052. doi: 10 . 1140 / epjc / s10052-023-11633-5 . url: http://dx.doi.org/10.1140/epjc/s10052-023- 11633-5

    2023 doi

  31. [53]

    url: http://dx.doi.org/10.21468/SciPostPhys.15.2.053

    doi

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.