Pith. sign in

REVIEW 4 major objections 5 minor 3 cited by

Generator Based Inference (GBI)

T0 review · 4 major / 5 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 Data-driven background generators turn anomaly scores into calibrated physics parameters, down to $0.1\sigma$ signals.

desk verdict Useful extension of anomaly-detection toolkit with honest validation, but the 0.1σ sensitivity claim is narrower than the abstract implies. read the letter →

arxiv 2506.00119 v1 pith:NOE3MWIX submitted 2025-05-30 hep-ph cs.LGhep-ex

Chi Lung Cheng , Ranit Das , Runze Li , Radha Mastandrea , Vinicius Mikuni , Benjamin Nachman , David Shih , Gup Singh This is my paper · ORCID
classification hep-phcs.LGhep-ex
keywords GeneratorBasedInferenceresonantanomalydetectionsimulation-baseddata-drivenbackgroundestimationnormalizingflowsflowmatchinglikelihoodratioparameter
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

The paper proposes Generator Based Inference (GBI), a general way to do unbinned, high-dimensional parameter estimation when the generator used to model the data is learned from data rather than from a physics simulator. In the resonant anomaly detection setup studied here, background densities learned from sidebands are inserted into a mixture likelihood in the signal region, so the output of an anomaly search is no longer just a two-sample p-value but an estimate of the signal fraction $\mu$ and, for the hybrid GBI-PAWS method, of the parent and daughter masses $m_X,m_Y$. The paper reports that this added interpretability costs no sensitivity: GBI-PAWS on the LHC Olympics benchmark finds signals at injections as low as $0.1\sigma$, and the confidence intervals from profile likelihoods, Fisher information, and pseudoexperiments agree. The broad payoff is that many data-driven background estimates used across physics could be promoted to full likelihoods, extending the reach of simulation-based inference to settings where no reliable background simulation exists.

What carries the argument

The central object is the mixture likelihood $p_D(x|m)=\mu p_S(x|m)+(1-\mu)p_B(x|m)$, where the background density $p_B$ is a data-driven model learned from sideband events: an ensemble of twenty normalizing flows for R-ANODE and a conditional flow-matching model for GBI-PAWS. The signal density is either nonparametric, a second normalizing flow $f(x)$ fit with $p_B$ frozen, or parameterized through a pretrained classifier $g(x,\theta)$ whose odds are converted to $\Lambda_{\mathrm{FS}}=p_S(x|\theta)/p_B(x)$. The weakly supervised ratio $\Lambda_{\mathrm{WS}}=\mu\Lambda_{\mathrm{FS}}+(1-\mu)$ is then maximized as $\sum_i \log \Lambda_{\mathrm{WS}}(x_i|\theta,\mu)$, which is the single-dataset GBI loss. Confidence intervals are computed from the profile log-likelihood drop, from the Fisher information matrix, and from bootstrap pseudoexperiments; the paper checks that all three agree and cover correctly.

What would settle it

Run the GBI-PAWS and R-ANODE inference on pseudoexperiments generated from a known background density, inject a $0.1\sigma$ signal, and check that the 68% and 95% intervals for $\mu$ and the fitted masses cover at the claimed rates; if the low-injection fit instead lands on the $\{220,75\}$ GeV background artifact, or the R-ANODE bias persists when the background flow is trained on true signal-region background, the sensitivity and calibration claims are falsified.

Watch

Extended reading notes

Core claim

On the paper's own terms, the discovery is that statistical outputs of anomaly detection can be made directly interpretable by training the signal and background densities and maximizing a single mixture likelihood. For R-ANODE, a normalizing flow $f(x)$ is fit in the signal region with the sideband-learned $p_B(x)$ frozen, and $\mu$ is extracted by scanning and maximizing $\sum_i \log(\mu f(x_i)+(1-\mu)p_B(x_i))$; the learned signal density shows peaks at the true masses once $\mu$ is significantly nonzero. For GBI-PAWS, a parameterized classifier pretrained on simulated signal templates and sideband-learned background produces the supervised likelihood ratio $\Lambda_{\mathrm{FS}}=p_S(x|\theta)/p_B(x)$, and the fit maximizes $\sum_i \log(\mu\Lambda_{\mathrm{FS}}(x_i|\theta)+1-\mu)$ over $\theta=(m_X,m_Y)$ and $\mu$ using only the data. Compared with the original two-sample PAWS loss, this single-dataset formulation removes the need for a reference sample and, together with ensembling, improves sensitivity by roughly a factor of five, so the benchmark $W'$ signal is detected at $0.1\sigma$ injection. The paper states that this is a new state of the art for anomaly detection sensitivity on the LHC Olympics benchmark, with R-ANODE covering signals starting near $1\sigma$ and GBI-PAWS extending to $0.1\sigma$.

Load-bearing premise

The method assumes that the background distribution learned from the sidebands is exactly the background distribution inside the signal region, so any mismatch in the interpolation is counted by the fit as signal.

Editorial extensions

If this is right

  • A resonant anomaly search can report a signal fraction $\mu$ with calibrated confidence intervals, turning an anomaly detector into a measuring instrument.
  • The GBI-PAWS fit uses only the data itself, no separate reference sample; the paper attributes about a factor-of-five sensitivity gain to this change plus ensembling.
  • GBI-PAWS reaches $0.1\sigma$ injected signal on the LHC Olympics benchmark while R-ANODE covers signals from about $1\sigma$, so the two methods are complementary in breadth versus depth.
  • Because the background generator is abstract, the same recipe can promote other data-driven background estimates to unbinned high-dimensional likelihoods.
  • Confidence intervals from profile likelihood, Fisher information, and bootstrap pseudoexperiments agree across the mass grid, supporting the use of profile-likelihood scans in practice.

Reading between the lines

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

  • Beyond the paper: the GBI recipe should transfer to non-resonant searches or ABCD-style control regions, wherever a background model can be trained away from the signal-enriched region.
  • Beyond the paper: the $0.1\sigma$ sensitivity is demonstrated on one benchmark background with a penalty steering the fit away from a low-mass artifact; on other backgrounds, the achievable floor may be set by how signal-like the sideband interpolation error looks.
  • Beyond the paper: the depth-breadth tradeoff between R-ANODE and GBI-PAWS suggests a natural test, scanning GBI-PAWS over signal templates deliberately absent from the pretraining grid to quantify how much sensitivity is borrowed from the prior over $\theta$.
  • Beyond the paper: because the weakest point is sideband-to-signal-region interpolation, a practical extension would be to train a discriminator between sideband and signal-region background-only pseudoexperiments to estimate and subtract this bias.
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

4 major / 5 minor

Summary. The paper introduces Generator Based Inference (GBI), a framework that generalizes simulation-based inference to settings where the generator is data-driven, and applies it to resonant anomaly detection. Two methods are studied: R-ANODE, which learns a non-parametric signal density while fixing a sideband-learned background density, and GBI-PAWS, which pre-trains a parameterized classifier on simulated signal models and a data-driven background and then fits the signal fraction and mass parameters with a single likelihood. On the LHC Olympics benchmark with a W' -> XY signal at (mX,mY)=(100,500) GeV, the authors report that GBI-PAWS can detect anomalies from an injected signal fraction around 0.1 sigma and provide confidence intervals on the signal fraction and mass parameters. The paper includes coverage checks via bootstrap pseudoexperiments, comparisons of three uncertainty quantification methods, and public code and data.

Significance. If the claims hold, GBI is a useful conceptual unification: it extends likelihood-based, unbinned inference to data-driven background generators and gives anomaly detection outputs a direct statistical interpretation in terms of signal strength and physical parameters. The paper's concrete strengths are the explicit pseudoexperiment coverage validation, the agreement among likelihood-ratio and Fisher-information uncertainties, and the release of code and additional signal samples. These make the parameter-estimation part of the work reproducible. However, the headline sensitivity claims rest on assumptions that the paper itself shows are only partially controlled: the sideband-interpolated background density is the dominant source of positive bias at low signal injection for R-ANODE, and the GBI-PAWS sensitivity below about 0.1 sigma relies on a hand-added penalty term whose effect is not ablated. The broader 'new state-of-the-art' claim also goes beyond the comparisons actually presented. The framework contribution is solid, but the sensitivity and state-of-the-art statements need additional support.

major comments (4)
  1. [Sec. IV, Fig. 1] The claim that GBI-PAWS 'is able to find signals that start above about 0.1 sigma' is not supported for signals in the mass region affected by the background artifact. The paper states that the model learned a spurious low-mass peak at {mX,mY}={220,75} GeV and that an exponential penalty below 85 GeV was added to steer the fit away from it. Since the benchmark signal (100,500) GeV lies above that threshold, the reported 0.1-sigma sensitivity is, as presented, sensitivity of GBI-PAWS plus a penalty tuned to avoid this particular artifact. No ablation of the penalty threshold, no scan of signal masses below 85 GeV, and no demonstration that the penalty does not remove real low-mass signals are provided. This should be addressed before the sensitivity claim can be taken as a general property of the method.
  2. [Appendix A and Sec. II, Eq. (5)] The load-bearing assumption that the sideband-learned background density p_B(x|m) interpolates accurately into the signal region is shown in Appendix A to be violated in exactly the low-signal regime of the sensitivity claims. For R-ANODE, Figure 5 demonstrates that the positive bias at low signal injection is dominated by interpolation error and disappears only in the unphysical configurations where p_B is trained on signal-region background or where the data are generated from p_B. Because the same sideband-interpolated p_B enters the GBI-PAWS likelihood through Eq. (5), the spurious {220,75} GeV peak described in Sec. IV is a concrete manifestation of this mismodeling. The paper does not quantify the residual mismodeling relative to the 0.1-sigma signal, so the central sensitivity claim lacks a validated background-modeling uncertainty.
  3. [Abstract and Sec. IV] The abstract's claim that 'the performance on the LHCO community benchmark dataset establishes a new state-of-the-art for anomaly detection sensitivity' is not substantiated by the comparisons in the paper. Figure 1 compares GBI-PAWS only with R-ANODE and with truth; there is no quantitative comparison at matched luminosity or matched significance metric against other published anomaly-detection methods on the same benchmark (e.g., CATHODE, ANODE, or other LHC Olympics entries). The statement 'new state-of-the-art' therefore overstates what the presented results demonstrate and should either be supported with a systematic benchmark comparison or softened to a claim about improvement relative to the specific baselines used.
  4. [Sec. IV, Fig. 2] The coverage validation is described inconsistently and incompletely. The text states that the bootstrap uses 1000 pseudoexperiments, while the top panel of Fig. 2 reports '70/100' and '96/100' bootstrap coverage, implying 100 pseudoexperiments. With 100 pseudoexperiments, a 70% coverage rate for a nominal 68% interval is within binomial fluctuations, but the figure and text should agree on the number used. More importantly, the coverage validation is performed at 0.3% injection only, not at the 0.1-sigma sensitivity boundary where the penalty term is active, so the reported coverage does not validate the uncertainty estimates in the regime of the headline claim.
minor comments (5)
  1. [Sec. IV] The phrase '0.1 sigma' is not defined. State explicitly whether it refers to S/sqrt(B), a one-sided significance from a profile likelihood, or some other convention, and give the corresponding injected signal fraction.
  2. [Sec. II.B, Eq. (5)] Equation (5) is described as 'maximize the likelihood (ratio)', but the expression is the sum of log-likelihood ratios. The wording should be made precise to avoid confusion between the likelihood and the likelihood ratio.
  3. [Appendix A] The sentence 'we use model f(x) with 1/10 number of parameters than used in Ref. [11]' is grammatically awkward and should read 'one tenth the number of parameters used in Ref. [11]'.
  4. [Sec. IV] The paper says the GBI-PAWS fit 'sometimes learned a wrong mass value at low signal injection due to a background peak around {mX,mY}={220,75} GeV'. It would be helpful to report how often this occurs and how the exponential penalty changes the distribution of fitted mass values, even if a full ablation is left to future work.
  5. [Sec. III] The definition of the sideband as 'm not in [3.3,3.7] TeV' and the use of a parametric fit to p(m) should be stated more explicitly: the parametric fit is used to reweight or sample sideband events into the signal region, and this step is another potential source of interpolation error that is not separately validated.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: GBI is a standard likelihood-based extension of prior anomaly-detection methods, with limitations explicitly disclosed.

full rationale

GBI’s derivation chain is self-contained: the background density p_B(x|m) is learned from sideband data (Sec. II), the signal likelihood ratio in GBI-PAWS is trained with simulated signal and data-driven background via Eq. 2, and the weak-supervision likelihood ratio Eq. 4 is a standard algebraic construction from the classifier output; the final estimate maximizes Eq. 5. R-ANODE maximizes Eq. 1 directly. Neither step defines the target parameters (µ, m_X, m_Y) in terms of themselves. The claims of interpretable parameter estimation are tested against injected signals and pseudoexperiments (Figs. 2-3). The paper explicitly discloses the two main limitations: Appendix A shows that R-ANODE low-injection bias is dominated by sideband interpolation error of p_B, and Sec. IV describes an exponential penalty below 85 GeV to suppress a background artifact in GBI-PAWS. These are acknowledged robustness caveats, not circular reductions: the penalty does not encode the claimed (100,500) GeV signal, and the interpolation error is identified as a bias source rather than hidden. Self-citations to Refs. [11,12] are normal extensions of the authors’ prior methods and are not used as unverified proof of the new results. Hence no significant circularity.

Assumptions & free parameters 3 free parameters · 5 assumptions · 0 invented entities

The central claim rests on fitted signal parameters, a hand-chosen penalty threshold, and standard statistical and generative modeling assumptions. No new physical entities are introduced.

free parameters (3)
  • signal fraction μ = 0.268% (best fit at 0.3% injection)
    Primary parameter estimated by maximizing the GBI likelihood (Eq. 5 for PAWS, Eq. 1 for R-ANODE).
  • signal masses mX, mY (GBI-PAWS) = ~100, ~500 GeV at truth (100, 500) GeV
    Physical parameters of the W' to XY signal model fitted from data in the profile likelihood.
  • Exponential penalty threshold = 85 GeV
    Hand-chosen threshold added to the loss to steer fits away from a background artifact that mimics low-mass signals at low injection.
assumptions (5)
  • domain assumption Sideband region contains negligible signal contamination
    Used in Section II to estimate p_B from sidebands; the authors note they found no impact from signal leakage and drop it from the workflow.
  • standard math Wilks' theorem applies for profile likelihood confidence intervals
    Used in Section II (Ref [17]) to convert log-likelihood decreases into confidence regions with chi-squared thresholds.
  • standard math The trained classifier provides an optimal likelihood ratio estimate
    Equation (3) uses the relation g* approx kappa p_S/(p_S + p_B), which holds for a properly calibrated binary classifier.
  • domain assumption The signal model grid (W' to XY, mX, mY < 600 GeV, 50 GeV increments) covers the relevant new physics
    Section III defines the pre-training grid for GBI-PAWS; sensitivity to off-grid models is cited from Ref [12] but not re-established here.
  • domain assumption Generative models (normalizing flows, flow matching) converge to the target densities
    Section II assumes these models provide adequate density estimates and samples for the likelihood construction.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Generator Based Inference (GBI)." pith.science (2026). https://pith.science/paper/NOE3MWIX

@misc{pith2026250600119,
  author       = {Pith},
  title        = {Pith review of: Generator Based Inference (GBI)},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/NOE3MWIX}},
  note         = {Machine review of arXiv:2506.00119}
}
read the original abstract

Statistical inference in physics is often based on samples from a generator (sometimes referred to as a ``forward model") that emulate experimental data and depend on parameters of the underlying theory. Modern machine learning has supercharged this workflow to enable high-dimensional and unbinned analyses to utilize much more information than ever before. We propose a general framework for describing the integration of machine learning with generators called Generator Based Inference (GBI). A well-studied special case of this setup is Simulation Based Inference (SBI) where the generator is a physics-based simulator. In this work, we examine other methods within the GBI toolkit that use data-driven methods to build the generator. In particular, we focus on resonant anomaly detection, where the generator describing the background is learned from sidebands. We show how to perform machine learning-based parameter estimation in this context with data-derived generators. This transforms the statistical outputs of anomaly detection to be directly interpretable and the performance on the LHCO community benchmark dataset establishes a new state-of-the-art for anomaly detection sensitivity.

Figures

Figures reproduced from arXiv: 2506.00119 by the authors.

Figure 1
Figure 1. FIG. 1. Median estimated signal fraction ˆµ [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 3
Figure 3. shows that the GBI-PAWS works across mass val￾ues. In particular, we find that the method is effective across the entire mass grid and the three uncertainty quantification methods agree. We observe a slight bias only for the (300,300) GeV mass point, but the shift is a few percent and is likely practically irrelevant for the interpretability task. Signal Masses (mX, mY) [GeV] 0.96 0.98 1 1.02 1.04 m Y Bootstrap Like… view at source ↗
Figure 2
Figure 2. FIG. 2. Top: Inferred anomaly mass properties ( [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: FIG. 4. R-ANODE likelihood fitting at two different true [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]
Figure 5
Figure 5. Figure 5: FIG. 5. Median estimated signal fraction ˆµ [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]
Figure 7
Figure 7. Figure 7: FIG. 7. The inferred physical properties of the anomalies for [PITH_FULL_IMAGE:figures/full_fig_p009_7.png]
Figure 6
Figure 6. Figure 6: FIG. 6. Median estimated signal fraction ˆµ [PITH_FULL_IMAGE:figures/full_fig_p009_6.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 3 Pith papers

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

  1. Theory-informed neural networks for particle physics

    hep-ph 2025-07 conditional novelty 7.0 of 10

    A Deep Q-Network using matrix-element rewards reconstructs parton assignments in collider events, enabling theory-based tagging and anomaly detection without labels.

  2. Look everywhere effects in anomaly detection

    hep-ph 2025-12 conditional novelty 6.0 of 10

    Weakly supervised anomaly detectors that train and test on the same data produce badly miscalibrated p-values; independent test sets are calibrated but insensitive, while k-fold cross-validation is a workable middle ground.

  3. Toward an event-level analysis of hadron structure using differential programming

    hep-ph 2025-07 conditional novelty 4.0 of 10

    LOITS is a differentiable sampling method, demonstrated in a GAN closure test, that maps sampled events back to the parameters of a target density for event-level inference.

Reference graph

Works this paper leans on

29 extracted references · 7 canonical work pages · cited by 3 Pith papers

  1. [1]

    Cranmer, J

    K. Cranmer, J. Brehmer, and G. Louppe,The frontier of simulation-based inference,Proceedings of the National Academy of Sciences117(May, 2020) 30055–30062

    work page 2020

  2. [2]

    Arratia et al.,Presenting Unbinned Differential Cross Section Results,arXiv:2109.13243

    M. Arratia et al.,Presenting Unbinned Differential Cross Section Results,arXiv:2109.13243

    arXiv

  3. [3]

    Huetsch et al.,The landscape of unfolding with machine learning,SciPost Phys.18(2025), no

    N. Huetsch et al.,The landscape of unfolding with machine learning,SciPost Phys.18(2025), no. 2 070, [arXiv:2404.18807]. [4]A TLASCollaboration, G. Aad et al.,An implementation of neural simulation-based inference for parameter estimation in ATLAS,arXiv:2412.01600. [5]A TLASCollaboration, G. Aad et al.,Measurement of off-shell Higgs boson production in th...

    arXiv 2025

  4. [6]

    Kasieczka et al.,The LHC Olympics 2020 a community challenge for anomaly detection in high energy physics,Rept

    G. Kasieczka et al.,The LHC Olympics 2020 a community challenge for anomaly detection in high energy physics,Rept. Prog. Phys.84(2021), no. 12 124201, [arXiv:2101.08320]

    arXiv 2021

  5. [7]

    Aarrestad et al.,The Dark Machines Anomaly Score Challenge: Benchmark Data and Model Independent Event Classification for the Large Hadron Collider, SciPost Phys.12(2022), no

    T. Aarrestad et al.,The Dark Machines Anomaly Score Challenge: Benchmark Data and Model Independent Event Classification for the Large Hadron Collider, SciPost Phys.12(2022), no. 1 043, [arXiv:2105.14027]

    arXiv 2022

  6. [8]

    Karagiorgi, G

    G. Karagiorgi, G. Kasieczka, S. Kravitz, B. Nachman, and D. Shih,Machine Learning in the Search for New Fundamental Physics,arXiv:2112.03769

    arXiv

  7. [9]

    Golling, G

    T. Golling, G. Kasieczka, C. Krause, R. Mastandrea, B. Nachman, J. A. Raine, D. Sengupta, D. Shih, and M. Sommerhalder,The interplay of machine learning-based resonant anomaly detection methods, Eur. Phys. J. C84(2024), no. 3 241, [arXiv:2307.11157]

    arXiv 2024

  8. [10]

    Belis, P

    V. Belis, P. Odagiu, and T. K. Aarrestad,Machine learning for anomaly detection in particle physics,Rev. Phys.12(2024) 100091, [arXiv:2312.14190]

    arXiv 2024

Show all 29 references
  1. [11]

    R. Das, G. Kasieczka, and D. Shih,Residual ANODE, arXiv:2312.11629

    arXiv

  2. [12]

    C. L. Cheng, G. Singh, and B. Nachman,Incorporating Physical Priors into Weakly-Supervised Anomaly Detection,arXiv:2405.08889

    arXiv

  3. [13]

    Nachman and D

    B. Nachman and D. Shih,Anomaly Detection with Density Estimation,Phys. Rev. D101(2020) 075042, [arXiv:2001.04990]

    2020 arXiv

  4. [14]

    Hallin, J

    A. Hallin, J. Isaacson, G. Kasieczka, C. Krause, B. Nachman, T. Quadfasel, M. Schlaffer, D. Shih, and M. Sommerhalder,Classifying anomalies through outer density estimation (CATHODE),Phys. Rev. D106 (2022), no. 5 055006, [arXiv:2109.00546]

    2022 arXiv

  5. [15]

    D. J. Rezende and S. Mohamed,Variational inference with normalizing flows, 2016

    2016

  6. [16]

    Lipman, R

    Y. Lipman, R. T. Chen, H. Ben-Hamu, M. Nickel, and M. Le,Flow matching for generative modeling,arXiv preprint arXiv:2210.02747(2022)

    2022 arXiv

  7. [17]

    S. S. Wilks,The Large-Sample Distribution of the Likelihood Ratio for Testing Composite Hypotheses, Annals Math. Statist.9(1938), no. 1 60–62

    1938

  8. [18]

    Cranmer, J

    K. Cranmer, J. Pavez, and G. Louppe,Approximating Likelihood Ratios with Calibrated Discriminative Classifiers,arXiv:1506.02169

    arXiv

  9. [19]

    Baldi, K

    P. Baldi, K. Cranmer, T. Faucett, P. Sadowski, and D. Whiteson,Parameterized neural networks for high-energy physics,Eur. Phys. J.C76(2016), no. 5 235, [arXiv:1601.07913]

    2016 arXiv

  10. [20]

    Wald,Tests of statistical hypotheses concerning several parameters when the number of observations is large,Transactions of the American Mathematical Society54(1943), no

    A. Wald,Tests of statistical hypotheses concerning several parameters when the number of observations is large,Transactions of the American Mathematical Society54(1943), no. 3 426–482

    1943

  11. [21]

    Kasieczka, B

    G. Kasieczka, B. Nachman, and D. Shih,Official Datasets for LHC Olympics 2020 Anomaly Detection Challenge (Version v6) [Data set]., 2019. https://doi.org/10.5281/zenodo.4536624

    2020 doi

  12. [22]

    Sjostrand, S

    T. Sjostrand, S. Mrenna, and P. Z. Skands,PYTHIA 6.4 Physics and Manual,JHEP05(2006) 026, [hep-ph/0603175]

    2006 arXiv

  13. [23]

    Sj¨ ostrand, S

    T. Sj¨ ostrand, 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.191(2015) 159–177, [arXiv:1410.3012]. [24]DELPHES 3Collaboration, J. de Favereau, C. Delaere, P...

    2015 arXiv

  14. [25]

    Mertens,New features in Delphes 3,J

    A. Mertens,New features in Delphes 3,J. Phys. Conf. Ser.608(2015), no. 1 012045

    2015

  15. [26]

    Cacciari and G

    M. Cacciari and G. P. Salam,Dispelling theN 3 myth for thek t jet-finder,Phys. Lett.B641(2006) 57–61, 7 [hep-ph/0512210]

    2006 arXiv

  16. [27]

    Cacciari, G

    M. Cacciari, G. P. Salam, and G. Soyez,FastJet User Manual,Eur. Phys. J. C72(2012) 1896, [arXiv:1111.6097]

    2012 arXiv

  17. [28]

    Cacciari, G

    M. Cacciari, G. P. Salam, and G. Soyez,The anti-k t jet clustering algorithm,JHEP04(2008) 063, [arXiv:0802.1189]

    2008 arXiv

  18. [29]

    Thaler and K

    J. Thaler and K. Van Tilburg,Identifying Boosted Objects with N-subjettiness,JHEP03(2011) 015, [arXiv:1011.2268]

    2011 arXiv

  19. [30]

    Thaler and K

    J. Thaler and K. Van Tilburg,Maximizing Boosted Top Identification by Minimizing N-subjettiness,JHEP02 (2012) 093, [arXiv:1108.2701]. 8 Appendix A: Model Bias in R-ANODE As shown in Fig. 1, at low signal strength (S/ √ B <4), the estimated signal fraction ˆµtends to bias towar...

    2012 arXiv

  20. [31]

    Directly usep B to gen- erate backgrounds in SR data

    Pure background events in SR 3. Directly usep B to gen- erate backgrounds in SR data. 9 Appendix B: R-ANODE Performance Across Signal Models 0.03 0.07 0.17 0.35 0.7 1.75 3.49 17.46 S/ B 0.01 0.02 0.05 0.1 0.2 0.5 1 5 Signal Injection (%) 0.01 0.02 0.05 0.1 0.2 0.5 1 5 (%) Lumi...

  21. [300]

    6, and the learned physical properties at (300, 300) is shown in Fig

    is compared with previous result at (100, 500) in Fig. 6, and the learned physical properties at (300, 300) is shown in Fig. 7. We find that the performance at signal mass (300, 300) to be slightly better than performance at signal mass (100, 500), which could be due to the si...

Pith tools

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