Pith. sign in

REVIEW 3 major objections 7 minor 300 references

A signed neural control variate from two normalizing flows compresses event-weight ranges and removes negative weights, cutting the cost of LO and NLO collider predictions.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · grok-4.5

2026-07-30 18:05 UTC pith:PMZVBPKR

load-bearing objection Solid methods paper: signed/nested NCVs really do cut weight range and negative fractions on LO/NLO benchmarks; the cost claim is post-training only. the 3 major comments →

arxiv 2607.23591 v1 pith:PMZVBPKR submitted 2026-07-26 hep-ph

Neural Control Variates at LO and NLO

Theo Heimel , Tilman Plehn , Rebecca Revelli , Sophia Vent , Ramon Winterhalder This is my paper
classification hep-ph
keywords neural control variatesneural importance samplingnegative weightsMonte Carlo event generationNLO QCDnormalizing flowsphase-space integrationdipole subtraction
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

Monte Carlo event generators for the LHC are slowed by wide ranges of event weights and by negative weights that appear once real-emission subtraction is introduced. This paper shows that a signed neural control variate—built from two normalizing flows, one for the positive part of the integrand and one for the negative—can be subtracted from the phase-space integrand so that most of the cross section is sampled with unit weights while the residual stays small and positive. Combined with neural importance sampling, the method improves unweighting efficiency by a factor of about 2.7 for leading-order top-pair production with two gluons, and at next-to-leading order it reduces relative variance by roughly an order of magnitude while cutting the negative-weight fraction by 85–90 percent for top-pair-plus-gluon and three-jet production. The conditional form of the control variate acts as a trainable finite subtraction that sits on top of ordinary dipole subtraction, so expensive matrix-element evaluations only handle a flattened residual. A sympathetic reader cares because weight fluctuations and negative weights are the practical bottlenecks that limit high-precision, high-luminosity event generation.

Core claim

A signed neural control variate built from two normalizing flows absorbs the bulk of a phase-space integral into a trivially sampled term with unit weights, leaving a residual that is small, nearly constant, and free of negative weights. When that residual is sampled by neural importance sampling, the combined setup significantly reduces the computational cost of both leading-order and next-to-leading-order predictions. At NLO the conditional control variate functions as a learned finite subtraction that complements standard dipole schemes after the soft and collinear divergences have been regularized.

What carries the argument

The signed neural control variate c_θ(x) = C₊ p₊(x) − C₋ p₋(x), each p± a normalizing flow and each C± a learned non-negative normalization whose difference is the known integral; nested conditional and unconditional versions handle the Born-plus-real structure of NLO so the residual can be sampled with positive weights.

Load-bearing premise

The softplus negativity barrier, the chosen loss normalizations, and mixture densities that include a periodically re-adapted grid on the negative residual are enough to keep that residual non-negative and low-variance on realistic multi-channel NLO phase spaces without heavy hyper-parameter tuning.

What would settle it

Train the same pipeline on an independent higher-multiplicity NLO process with a larger negative-weight fraction; if the residual negative contribution stays large or the order-of-magnitude variance gain collapses, the central claim does not hold.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • LO unweighting efficiency for multi-gluon processes can rise by a factor of roughly 2.7 on top of neural importance sampling alone.
  • NLO negative-weight fractions can fall by 85–90 percent while relative variance drops by about an order of magnitude versus standard adaptive sampling.
  • One-third to two-thirds of an NLO cross section can be generated from the control-variate flows without evaluating loop or real-emission matrix elements.
  • The same construction works for both massive and massless final states once infrared divergences are regularized by any standard subtraction scheme.
  • Differential distributions remain faithful bin-by-bin because the control variate is an exact rewriting of the integral.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The nested construction could be stacked for NNLO by applying successive conditional control variates to successive radiation variables.
  • Unit-weight control-variate events may simplify downstream experimental reweighting and uncertainty propagation.
  • Processes whose negatives come mainly from over-subtraction may gain more from the conditional control variate than from further tuning of the physics dipoles alone.
  • Pairing the method with amplitude surrogates would confine expensive or surrogate evaluations to the residual only, compounding the saving.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 7 minor

Summary. The authors construct neural control variates (NCVs) for Monte Carlo integration and event generation, built from two normalizing flows with learned scalar normalizations (Eq. 3), so that the control variate has a known integral and its compensating term can be sampled directly with unit-weight events (Eq. 5). A nested variant (Eqs. 9–14) treats Born-like and radiation variables separately, acting at NLO as a trainable finite subtraction on top of Catani–Seymour dipoles. The method is benchmarked on two toy integrands, on LO gg→tt̄gg (factor ~2.7 unweighting-efficiency gain over MadNIS, Table 2), and on NLO e⁺e⁻→tt̄g and e⁺e⁻→qq̄g, where the residual negative-weight contribution is reduced by ~90% and ~85% and the relative variance by factors of ~6–10 against Vegas/MadNIS (Tables 3–4). The identities in §2 are standard and correctly specialized, the comparisons use independent baselines on identical integrands, and differential distributions are validated bin-by-bin against MadNIS/Vegas (Figs. 2, 4).

Significance. If the results hold, this is a useful and general addition to the neural-importance-sampling toolbox: the signed two-flow control variate is a genuinely new construction in this context, the nested conditional variant mapping onto the Born-plus-real structure of subtracted NLO cross sections is elegant, and the construction is exact — the unbiasedness of the estimator in Eqs. (4)–(5) is an identity, not a fit. The benchmarks are honest: fixed, published hyperparameters ("we do not optimize the hyperparameters"), independent baselines on the same integrands, explicit sign-floor analysis (Eqs. 36–37, 45) showing MadNIS already saturates the sign-limited variance floor and that only integrand modification can go below it. The trainable-subtraction viewpoint for NLO is conceptually appealing and subtraction-scheme-independent. The two caveats below concern how the headline claims — "avoid negative weights" and "significantly reduces the computational cost" — are supported, not the correctness of the method.

major comments (3)
  1. [§4.3–4.4, Tables 3–4, Eq. (5)] The 'sign removal' claim is established only for the residual sample, but the final unweighted event sample includes the trivially sampled negative-flow events, which carry weight −1 (Eq. 5, 'their events carry unit weight ±1'). From Tables 3–4 one can reconstruct the negative-weight fraction of the combined unweighted sample: for e+e−→tt̄g 1-NCV, C−,θ/σ=21.3% and σ−res/σ=0.4% give a negative-event fraction of roughly (21.3+0.4)/(75.6+21.3+46.1+0.4)≈15%, versus ≈4% for MadNIS; for e+e−→qq̄g 1-NCV with C−,θ/σ=84.1%, the combined negative fraction is ≈32% versus ≈11% for MadNIS. Correspondingly, the Kish factor of the combined sample ((σ/σabs)² ≈ 0.49 and ≈0.14) is substantially worse than the MadNIS baseline (≈0.92 and ≈0.79). Since negative-weight events cost in downstream simulation regardless of whether they originate from the residual or the trivial term, the abstract's 'avoid negativ
  2. [§4.3 (Eq. 114 and following), §5, Tables 6–7] The abstract and §5 claim that the method 'significantly reduces the computational cost of LO and NLO predictions', and §4.3 argues that 54–65% of the tt̄g cross section is absorbed into the trivially sampled Cθ term, saving loop-amplitude evaluations. All quoted gains (RV_σ, εuw, σ−res/σ) are inference metrics conditional on trained networks. The training itself is expensive: the losses of Eqs. (15)/(33) require f (and at NLO the full B,V,I,R,S,J combination) on mixture samples each iteration; Tables 6–7 give 10⁴ iterations × 1024 batch (LO) and 3.6×10⁴ × 8192 (NLO), i.e. O(10⁷) and O(3×10⁸) integrand evaluations respectively, plus the VEGAS re-adaptation of g_neg to the negative residual every 250 steps (§4.2). No total evaluation counts, wall-clock, or break-even sample size relative to MadNIS/Vegas (which have their own, smaller, training costs) are reported. The demonstrated stateme
  3. [§4.3–4.4, Eq. (112)] The NLO benchmarks are shown for a single collider energy and a single Durham resolution (√s=1 TeV, ycut=10⁻³, Eq. 112), with the jet boundary placed directly in the subtraction-affected region. The negative-weight sources the NCV targets (measurement-function mismatch, over-subtraction away from singular limits) depend strongly on ycut and on whether α-restricted dipoles are used (the paper itself notes the α-restriction option, §4.4). The repeated disclaimer that hyperparameters are not optimized leaves open whether the gains survive re-tuning per setup or transfer across ycut values. This does not need to become a scan, but at minimum the paper should state whether the Table 7 settings were re-used unchanged between tt̄g and qq̄g, and comment on expected sensitivity to ycut; a second ycut value for one process would substantially strengthen the NLO claim.
minor comments (7)
  1. [§2.1, after Eq. (13)] Typo: 'does not generate an variance' → 'any variance'.
  2. [§4.3, paragraph after Eq. (114)] Typo: 'where all all methods agree'. Also in Figure 4 the ratio panels show only NCV/MadNIS; including the Vegas ratio would make the comparison complete.
  3. [Table 2] The caption formatting 'Metric λneg =10 λneg =0 only NIS' is garbled. The row Neff/N(res)=0.000 for λneg=0 is consistent with the 13.7% negative residual but should perhaps be quoted with an uncertainty or as <10⁻³.
  4. [§2.2, discussion after Eq. (18)] The treatment of Z_NIS and Z_neg as 'conventions rather than derived quantities' is acceptable, but since Z_neg is degenerate with βmax (only βmax/Z_neg is meaningful), it would help to state already in §2.2 that all quoted βmax values refer to the normalization choices of Eqs. (44)/(113), as is done for the toy model.
  5. [§3.2, Eq. (55)] ε_cuts^p is assumed to have negligible statistical uncertainty (after Eq. 55); please state the sample size used for this estimate, since RV_σ(res) values such as 0.078 in Table 2 inherit this assumption.
  6. [§2.1, after Eq. (5)] The claim that the NCV events are 'unweighted with perfect efficiency' (after Eq. 5) holds for the positive flow; the negative-flow events carry weight −1. A half-sentence anticipating the discussion requested in the major comment would avoid confusion here.
  7. [§1, citations of Ref. [70]] The classical combined control-variate + importance-sampling reference [70] is appropriately cited; it would be fair to add one sentence on what the neural two-flow construction adds over the Vegas-based combination of [70], since the integration-side idea overlaps.

Circularity Check

0 steps flagged

No significant circularity: unbiased NCV estimator is standard control-variate algebra; reported gains are empirical residual metrics against independent baselines, not forced by definition or self-citation.

full rationale

The paper’s load-bearing identities—∫c_θ = C_+ − C_− (Eqs. 3–4), the residual estimator σ = ⟨(f−c_θ)/g⟩ + C_θ (Eq. 5), and the nested conditional/unconditional split (Eqs. 9–14, 103–107)—are the classical control-variate construction: any function with a known integral may be subtracted without bias. That algebra does not force small residual variance, positivity, or the quoted factors (f_uw_gain ≃ 2.7, RV reductions ~10×, σ⁻_res/σ drops of 85–90%). Those are measured post-training on independent evaluation samples against Vegas, MadNIS, and MG5aMC on the same integrands (Tables 1–4, Figs. 2, 4–5), with differential distributions required to match truth. Self-citations (MadNIS, MadSpace) supply sampling infrastructure and baselines, not a uniqueness premise that makes the NCV result tautological. The Müller et al. NCV citation and the Shyamsundar IS+CV citation are external method references. Training losses (Eqs. 15, 33) fit flows to f± and residuals; they do not redefine the target cross section as the fit. End-to-end training-cost amortization is unreported, but that is a completeness issue, not circularity. Score 0 is appropriate.

Axiom & Free-Parameter Ledger

4 free parameters · 4 axioms · 2 invented entities

The work rests on standard Monte-Carlo identities, the classical control-variate theorem, normalizing-flow universal approximation, and the Catani–Seymour dipole subtraction. The only free choices that affect the quoted gains are the loss weights λ_int, λ_neg, λ_cut, the barrier schedule β(t), and the network widths/depths; these are listed as free parameters because the paper explicitly did not optimize them yet still claims significant improvements.

free parameters (4)
  • λ_int, λ_neg, λ_cut = O(1) values (0.5–10)
    Relative weights of integral, negativity and cut-penalty losses; chosen by hand per section (Tables 5–7) and not scanned.
  • β_max and annealing schedule = β_max = 10–30
    Softplus barrier strength that enforces residual positivity; annealed over a fixed fraction of training steps.
  • flow architecture (layers, bins, hidden dim) = 3–4 layers, 10–32 bins, 64–256 hidden
    Capacity of the NCV and NIS normalizing flows; fixed per experiment in Appendix A.
  • mixture coefficients and g_neg re-adaptation interval = equal mixture; re-adapt every 250 steps
    How training samples are drawn from the joint of NIS, NCV and negative-region VEGAS grids.
axioms (4)
  • standard math Classical control-variate identity: ∫(f−c)+∫c = ∫f whenever ∫c is known exactly
    Eqs. (2)–(4); foundation of the whole method.
  • domain assumption Normalizing flows can approximate the positive and negative parts of a square-integrable density on the unit hypercube to arbitrary accuracy
    Implicit in the claim that the residual can be driven near zero (Sec. 2.1).
  • domain assumption Catani–Seymour dipole subtraction renders the real-emission integrand integrable and cancels virtual poles
    Sec. 4.1; the NCV is applied after this regularization.
  • domain assumption Infrared-safe measurement functions satisfy J(Φ_{n+1})→J(Φ̃_n) in soft/collinear limits
    Eq. (73); required for the NLO cross section to be finite before any NCV is added.
invented entities (2)
  • signed two-flow neural control variate (c_θ = C_+ p_+ − C_− p_−) no independent evidence
    purpose: Simultaneously compress positive weight range and lift negative regions while keeping the integral of c known analytically
    Core construction of Sec. 2.1; not present as a signed pair in the cited Müller et al. work.
  • nested conditional/unconditional NCV for Born-plus-real structure no independent evidence
    purpose: Exploit the factorization f = b(x_B)+r(x_B,x_R) so that the conditional NCV acts as a learnable finite subtraction
    Eqs. (9)–(14) and Sec. 4.2; the paper’s main architectural novelty for NLO.

pith-pipeline@v1.2.0-grok45-kimik3 · 30546 in / 3135 out tokens · 63181 ms · 2026-07-30T18:05:35.075885+00:00 · methodology

0 comments
read the original abstract

We employ neural control variates to minimize the range of event weights and avoid negative weights for phase-space integration and event generation. A signed control variate, built from two normalizing flows, fulfills both tasks. Combined with neural importance sampling, it significantly reduces the computational cost of LO and NLO predictions. For the NLO case, our conditional neural control variate can be viewed as a trainable subtraction term, complementing the established physics subtraction schemes for enhanced sampling performance.

Figures

Figures reproduced from arXiv: 2607.23591 by Ramon Winterhalder, Rebecca Revelli, Sophia Vent, Theo Heimel, Tilman Plehn.

Figure 1
Figure 1. Figure 1: Nested NCV for Toy-I. Left: real-emission integrand [PITH_FULL_IMAGE:figures/full_fig_p009_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Top transverse momentum distributions pT,t for t¯tgg production using an NCV trained without (left) and with (right) Lneg,θ to penalize negative weights. In [PITH_FULL_IMAGE:figures/full_fig_p014_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Representative Feynman diagrams for the Born, virtual, and real-emission [PITH_FULL_IMAGE:figures/full_fig_p017_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Selected differential distributions comparing all sampling methods for [PITH_FULL_IMAGE:figures/full_fig_p023_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Normalized weight distributions comparing all sampling methods for [PITH_FULL_IMAGE:figures/full_fig_p025_5.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

300 extracted references · 7 canonical work pages

  1. [1]

    One Generator, Any Process: LLM-Conditioning for the LHC

    Bahl, Henning and Plehn, Tilman and Schiller, Daniel and Sivagnanalingam, Thanush. One Generator, Any Process: LLM-Conditioning for the LHC. 2026. arXiv:2606.23791

    Pith/arXiv arXiv 2026

  2. [2]

    MadNIS at NLO

    De Crescenzo, Giovanni and Villadamigo, Javier Mari \ n o and Elmer, Nina and Heimel, Theo and Plehn, Tilman and Winterhalder, Ramon and Zaro, Marco. MadNIS at NLO. 2026. arXiv:2603.22407

    Pith/arXiv arXiv 2026

  3. [3]

    Explicit or Implicit? Encoding Physics at the Precision Frontier

    Breso-Pla, Victor and Greif, Kevin and Mikuni, Vinicius and Nachman, Benjamin and Plehn, Tilman and Wamorkar, Tanvi and Whiteson, Daniel. Explicit or Implicit? Encoding Physics at the Precision Frontier. 2026. arXiv:2603.08802

    arXiv 2026

  4. [4]

    The Latent Information Geometry of Jet Classification

    Kuntz, Rebecca Maria and Plehn, Tilman and Sch. The Latent Information Geometry of Jet Classification. 2026. arXiv:2603.02310

    Pith/arXiv arXiv 2026

  5. [5]

    Unfolding without Iterations, Adversaries, or Surrogates

    Ore, Ayodele and Plehn, Tilman. Unfolding without Iterations, Adversaries, or Surrogates. 2026. arXiv:2602.24282

    arXiv 2026

  6. [6]

    MadAgents

    Plehn, Tilman and Schiller, Daniel and Schmal, Nikita. MadAgents. 2026. arXiv:2601.21015

    Pith/arXiv arXiv 2026

  7. [7]

    How to Trust Learned Loop Amplitudes

    Bahl, Henning and Braun, Jens and Heinrich, Gudrun and Plehn, Tilman and Revelli, Rebecca. How to Trust Learned Loop Amplitudes. 2026. arXiv:2601.00950

    Pith/arXiv arXiv 2026

  8. [8]

    Economical Jet Taggers -- Equivariant, Slim, and Quantized

    Petitjean, Antoine and Plehn, Tilman and Spinner, Jonas and K. Economical Jet Taggers -- Equivariant, Slim, and Quantized. 2025. arXiv:2512.17011

    arXiv 2025

  9. [9]

    Generative Unfolding of Jets and Their Substructure

    Petitjean, Antoine and Butter, Anja and Greif, Kevin and Palacios Schweitzer, Sofia and Plehn, Tilman and Spinner, Jonas and Whiteson, Daniel. Generative Unfolding of Jets and Their Substructure. 2025. arXiv:2510.19906

    arXiv 2025

  10. [10]

    Forecasting Generative Amplification

    Bahl, Henning and Diefenbacher, Sascha and Elmer, Nina and Plehn, Tilman and Spinner, Jonas. Forecasting Generative Amplification. SciPost Phys. 2026. doi:10.21468/SciPostPhys.20.5.150. arXiv:2509.08048

    Pith/arXiv arXiv 2026

  11. [11]

    FASTColor -- Full-color Amplitude Surrogate Toolkit for QCD

    Villadamigo, Javier Mari \ n o and Frederix, Rikkert and Plehn, Tilman and Vitos, Timea and Winterhalder, Ramon. FASTColor -- Full-color Amplitude Surrogate Toolkit for QCD. 2025. arXiv:2509.07068

    Pith/arXiv arXiv 2025

  12. [12]

    Unbinning global LHC analyses

    Bahl, Henning and Plehn, Tilman and Schmal, Nikita. Unbinning global LHC analyses. 2025. arXiv:2509.05409

    Pith/arXiv arXiv 2025

  13. [13]

    Iterative HOMER with uncertainties

    Butter, Anja and others. Iterative HOMER with uncertainties. SciPost Phys. 2026. doi:10.21468/SciPostPhys.20.2.042. arXiv:2509.03592

    arXiv 2026

  14. [14]

    Towards Precise Simulations and Inference for the Neutron EDM

    Degenkolb, Skyler and Favaro, Luigi and Fierlinger, Peter and Franz, Jennifer and Manasawala, Husain and Plehn, Tilman. Towards Precise Simulations and Inference for the Neutron EDM. 2025. arXiv:2509.02791

    Pith/arXiv arXiv 2025

  15. [15]

    Amplitude Uncertainties Everywhere All at Once

    Bahl, Henning and Elmer, Nina and Plehn, Tilman and Winterhalder, Ramon. Amplitude Uncertainties Everywhere All at Once. SciPost Phys. 2026. doi:10.21468/SciPostPhys.20.3.083. arXiv:2509.00155

    Pith/arXiv arXiv 2026

  16. [16]

    and Lippmann, Peter and Pitz, Sebastian and Plehn, Tilman and Qu, Huilin and Spinner, Jonas

    Favaro, Luigi and Gerhartz, Gerrit and Hamprecht, Fred A. and Lippmann, Peter and Pitz, Sebastian and Plehn, Tilman and Qu, Huilin and Spinner, Jonas. Lorentz-Equivariance without Limitations. 2025. arXiv:2508.14898

    arXiv 2025

  17. [17]

    The Physics Behind ML-based Quark-Gluon Taggers

    Vent, Sophia and Winterhalder, Ramon and Plehn, Tilman. The Physics Behind ML-based Quark-Gluon Taggers. SciPost Phys. 2026. doi:10.21468/SciPostPhys.20.3.084. arXiv:2507.21214

    Pith/arXiv arXiv 2026

  18. [18]

    Simulation-Prior Independent Neural Unfolding Procedure

    Butter, Anja and Heimel, Theo and Huetsch, Nathan and Kagan, Michael and Plehn, Tilman. Simulation-Prior Independent Neural Unfolding Procedure. 2025. arXiv:2507.15084

    Pith/arXiv arXiv 2025

  19. [19]

    CP -analyses with symbolic regression

    Bahl, Henning and Fuchs, Elina and Menen, Marco and Plehn, Tilman. CP -analyses with symbolic regression. SciPost Phys. 2026. doi:10.21468/SciPostPhys.20.2.040. arXiv:2507.05858

    Pith/arXiv arXiv 2026

  20. [20]

    Large Language Models -- the Future of Fundamental Physics?

    Heneka, Caroline and Nieser, Florian and Ore, Ayodele and Plehn, Tilman and Schiller, Daniel. Large Language Models -- the Future of Fundamental Physics?. SciPost Phys. 2026. doi:10.21468/SciPostPhys.20.3.070. arXiv:2506.14757

    arXiv 2026

  21. [21]

    Lorentz Local Canonicalization: How to Make Any Network Lorentz-Equivariant

    Spinner, Jonas and Favaro, Luigi and Lippmann, Peter and Pitz, Sebastian and Gerhartz, Gerrit and Plehn, Tilman and Hamprecht, Fred A. Lorentz Local Canonicalization: How to Make Any Network Lorentz-Equivariant. 2025. arXiv:2505.20280

    arXiv 2025

  22. [22]

    How to unfold top decays

    Favaro, Luigi and Kogler, Roman and Paasch, Alexander and Palacios Schweitzer, Sofia and Plehn, Tilman and Schwarz, Dennis. How to unfold top decays. SciPost Phys. Core. 2025. doi:10.21468/SciPostPhysCore.8.3.053. arXiv:2501.12363

    Pith/arXiv arXiv 2025

  23. [23]

    Accurate surrogate amplitudes with calibrated uncertainties

    Bahl, Henning and Elmer, Nina and Favaro, Luigi and Haussmann, Manuel and Plehn, Tilman and Winterhalder, Ramon. Accurate surrogate amplitudes with calibrated uncertainties. SciPost Phys. Core. 2025. doi:10.21468/SciPostPhysCore.8.4.073. arXiv:2412.12069

    arXiv 2025

  24. [24]

    Extrapolating jet radiation with autoregressive transformers

    Butter, Anja and Charton, Fran c ois and Villadamigo, Javier Mari \ n o and Ore, Ayodele and Plehn, Tilman and Spinner, Jonas. Extrapolating jet radiation with autoregressive transformers. SciPost Phys. 2026. doi:10.21468/SciPostPhys.20.1.004. arXiv:2412.12074

    arXiv 2026

  25. [25]

    Precision calibration of calorimeter signals in the ATLAS experiment using an uncertainty-aware neural network

    Aad, Georges and others. Precision calibration of calorimeter signals in the ATLAS experiment using an uncertainty-aware neural network. SciPost Phys. 2025. doi:10.21468/SciPostPhys.19.6.155. arXiv:2412.04370

    arXiv 2025

  26. [26]

    Generative unfolding with distribution mapping

    Butter, Anja and Diefenbacher, Sascha and Huetsch, Nathan and Mikuni, Vinicius and Nachman, Benjamin and Palacios Schweitzer, Sofia and Plehn, Tilman. Generative unfolding with distribution mapping. SciPost Phys. 2025. doi:10.21468/SciPostPhys.18.6.200. arXiv:2411.02495

    Pith/arXiv arXiv 2025

  27. [27]

    A Lorentz-equivariant transformer for all of the LHC

    Brehmer, Johann and Bres \'o , V \' ctor and de Haan, Pim and Plehn, Tilman and Qu, Huilin and Spinner, Jonas and Thaler, Jesse. A Lorentz-equivariant transformer for all of the LHC. SciPost Phys. 2025. doi:10.21468/SciPostPhys.19.4.108. arXiv:2411.00446

    arXiv 2025

  28. [28]

    Profile Likelihoods on ML-Steroids

    Heimel, Theo and Plehn, Tilman and Schmal, Nikita. Profile Likelihoods on ML-Steroids. 2024. arXiv:2411.00942

    Pith/arXiv arXiv 2024

  29. [29]

    Lorentz-Equivariant Geometric Algebra Transformers for High-Energy Physics

    Spinner, Jonas and Bres \'o , Victor and de Haan, Pim and Plehn, Tilman and Thaler, Jesse and Brehmer, Johann. Lorentz-Equivariant Geometric Algebra Transformers for High-Energy Physics. 38th conference on Neural Information Processing Systems. 2024. arXiv:2405.14806

    Pith/arXiv arXiv 2024

  30. [30]

    CaloChallenge 2022: a community challenge for fast calorimeter simulation

    Amram, Oz and others. CaloChallenge 2022: a community challenge for fast calorimeter simulation. Rept. Prog. Phys. 2025. doi:10.1088/1361-6633/ae1304. arXiv:2410.21611

    arXiv 2022

  31. [31]

    SKATR: A self-supervised summary transformer for SKA

    Ore, Ayodele and Heneka, Caroline and Plehn, Tilman. SKATR: A self-supervised summary transformer for SKA. SciPost Phys. 2025. doi:10.21468/SciPostPhys.18.5.155. arXiv:2410.18899

    Pith/arXiv arXiv 2025

  32. [32]

    Advancing tools for simulation-based inference

    Bahl, Henning and Bres \'o -Pla, V \' ctor and De Crescenzo, Giovanni and Plehn, Tilman. Advancing tools for simulation-based inference. SciPost Phys. Core. 2025. doi:10.21468/SciPostPhysCore.8.3.060. arXiv:2410.07315

    arXiv 2025

  33. [33]

    Differentiable MadNIS-Lite

    Heimel, Theo and Mattelaer, Olivier and Plehn, Tilman and Winterhalder, Ramon. Differentiable MadNIS-Lite. SciPost Phys. 2025. doi:10.21468/SciPostPhys.18.1.017. arXiv:2408.01486

    Pith/arXiv arXiv 2025

  34. [34]

    Constraining the Higgs potential with neural simulation-based inference for di-Higgs production

    Mastandrea, Radha and Nachman, Benjamin and Plehn, Tilman. Constraining the Higgs potential with neural simulation-based inference for di-Higgs production. Phys. Rev. D. 2024. doi:10.1103/PhysRevD.110.056004. arXiv:2405.15847

    Pith/arXiv arXiv 2024

  35. [35]

    CaloDREAM Detector response emulation via attentive flow matching

    Favaro, Luigi and Ore, Ayodele and Schweitzer, Sofia Palacios and Plehn, Tilman. CaloDREAM Detector response emulation via attentive flow matching. SciPost Phys. 2025. doi:10.21468/SciPostPhys.18.3.088. arXiv:2405.09629

    Pith/arXiv arXiv 2025

  36. [36]

    The landscape of unfolding with machine learning

    Huetsch, Nathan and others. The landscape of unfolding with machine learning. SciPost Phys. 2025. doi:10.21468/SciPostPhys.18.2.070. arXiv:2404.18807

    Pith/arXiv arXiv 2025

  37. [37]

    PINNferring the Hubble Function with Uncertainties

    R. PINNferring the Hubble Function with Uncertainties. 2024. arXiv:2403.13899

    Pith/arXiv arXiv 2024

  38. [38]

    A Global View of the EDM Landscape

    Degenkolb, Skyler and Elmer, Nina and Modak, Tanmoy and M. A Global View of the EDM Landscape. SciPost Phys. 2026. doi:10.21468/SciPostPhys.20.6.151. arXiv:2403.02052

    Pith/arXiv arXiv 2026

  39. [39]

    Optimal, fast, and robust inference of reionization-era cosmology with the 21cmPIE-INN

    Schosser, Benedikt and Heneka, Caroline and Plehn, Tilman. Optimal, fast, and robust inference of reionization-era cosmology with the 21cmPIE-INN. SciPost Phys. Core. 2025. doi:10.21468/SciPostPhysCore.8.2.037. arXiv:2401.04174

    Pith/arXiv arXiv 2025

  40. [40]

    Staying on top of SMEFT-likelihood analyses

    Elmer, Nina and Madigan, Maeve and Plehn, Tilman and Schmal, Nikita. Staying on top of SMEFT-likelihood analyses. SciPost Phys. 2025. doi:10.21468/SciPostPhys.18.3.108. arXiv:2312.12502

    Pith/arXiv arXiv 2025

  41. [41]

    Normalizing flows for high-dimensional detector simulations

    Ernst, Florian and Favaro, Luigi and Krause, Claudius and Plehn, Tilman and Shih, David. Normalizing flows for high-dimensional detector simulations. SciPost Phys. 2025. doi:10.21468/SciPostPhys.18.3.081. arXiv:2312.09290

    Pith/arXiv arXiv 2025

  42. [42]

    Semi-visible jets, energy-based models, and self-supervision

    Favaro, Luigi and Kr. Semi-visible jets, energy-based models, and self-supervision. SciPost Phys. 2025. doi:10.21468/SciPostPhys.18.2.042. arXiv:2312.03067

    Pith/arXiv arXiv 2025

  43. [43]

    Kicking it off(-shell) with direct diffusion

    Butter, Anja and Jezo, Tomas and Klasen, Michael and Kuschick, Mathias and Palacios Schweitzer, Sofia and Plehn, Tilman. Kicking it off(-shell) with direct diffusion. SciPost Phys. Core. 2024. doi:10.21468/SciPostPhysCore.7.3.064. arXiv:2311.17175

    Pith/arXiv arXiv 2024

  44. [44]

    The MadNIS reloaded

    Heimel, Theo and Huetsch, Nathan and Maltoni, Fabio and Mattelaer, Olivier and Plehn, Tilman and Winterhalder, Ramon. The MadNIS reloaded. SciPost Phys. 2024. doi:10.21468/SciPostPhys.17.1.023. arXiv:2311.01548

    Pith/arXiv arXiv 2024

  45. [45]

    Precision-machine learning for the matrix element method

    Heimel, Theo and Huetsch, Nathan and Winterhalder, Ramon and Plehn, Tilman and Butter, Anja. Precision-machine learning for the matrix element method. SciPost Phys. 2024. doi:10.21468/SciPostPhys.17.5.129. arXiv:2310.07752

    Pith/arXiv arXiv 2024

  46. [46]

    Returning CP-observables to the frames they belong

    Ackerschott, Jona and Barman, Rahool Kumar and Gon c alves, Dorival and Heimel, Theo and Plehn, Tilman. Returning CP-observables to the frames they belong. SciPost Phys. 2024. doi:10.21468/SciPostPhys.17.1.001. arXiv:2308.00027

    Pith/arXiv arXiv 2024

  47. [47]

    How to understand limitations of generative networks

    Das, Ranit and Favaro, Luigi and Heimel, Theo and Krause, Claudius and Plehn, Tilman and Shih, David. How to understand limitations of generative networks. SciPost Phys. 2024. doi:10.21468/SciPostPhys.16.1.031. arXiv:2305.16774

    Pith/arXiv arXiv 2024

  48. [48]

    Jet diffusion versus JetGPT Modern networks for the LHC

    Butter, Anja and Huetsch, Nathan and Palacios Schweitzer, Sofia and Plehn, Tilman and Sorrenson, Peter and Spinner, Jonas. Jet diffusion versus JetGPT Modern networks for the LHC. SciPost Phys. Core. 2025. doi:10.21468/SciPostPhysCore.8.1.026. arXiv:2305.10475

    Pith/arXiv arXiv 2025

  49. [49]

    and Favaro, Luigi and Feiden, Friedrich and Modak, Tanmoy and Plehn, Tilman

    Dillon, Barry M. and Favaro, Luigi and Feiden, Friedrich and Modak, Tanmoy and Plehn, Tilman. Anomalies, representations, and self-supervision. SciPost Phys. Core. 2024. doi:10.21468/SciPostPhysCore.7.3.056. arXiv:2301.04660

    Pith/arXiv arXiv 2024

  50. [50]

    New Theory Paradigms at the LHC

    M. New Theory Paradigms at the LHC. 2023. doi:10.1142/9789811280184_0001

    work page doi:10.1142/9789811280184_0001 2023

  51. [51]

    and Plehn, Tilman and Vogel, Lorenz

    Butter, Anja and Dillon, Barry M. and Plehn, Tilman and Vogel, Lorenz. Performance versus resilience in modern quark-gluon tagging. SciPost Phys. Core. 2023. doi:10.21468/SciPostPhysCore.6.4.085. arXiv:2212.10493

    Pith/arXiv arXiv 2023

  52. [52]

    MadNIS - Neural multi-channel importance sampling

    Heimel, Theo and Winterhalder, Ramon and Butter, Anja and Isaacson, Joshua and Krause, Claudius and Maltoni, Fabio and Mattelaer, Olivier and Plehn, Tilman. MadNIS - Neural multi-channel importance sampling. SciPost Phys. 2023. doi:10.21468/SciPostPhys.15.4.141. arXiv:2212.06172

    Pith/arXiv arXiv 2023

  53. [53]

    Modern Machine Learning for LHC Physicists

    Plehn, Tilman and Butter, Anja and Dillon, Barry and Heimel, Theo and Krause, Claudius and Winterhalder, Ramon. Modern Machine Learning for LHC Physicists. 2022. arXiv:2211.01421

    Pith/arXiv arXiv 2022

  54. [54]

    Ghosh, Aishik and Nachman, Benjamin and Plehn, Tilman and Shire, Lily and Tait, Tim M. P. and Whiteson, Daniel. Statistical patterns of theory uncertainties. SciPost Phys. Core. 2023. doi:10.21468/SciPostPhysCore.6.2.045. arXiv:2210.15167

    Pith/arXiv arXiv 2023

  55. [55]

    Cornering extended Starobinsky inflation with CMB and SKA

    Modak, Tanmoy and R. Cornering extended Starobinsky inflation with CMB and SKA. SciPost Phys. 2023. doi:10.21468/SciPostPhys.15.2.047. arXiv:2210.05698

    Pith/arXiv arXiv 2023

  56. [56]

    and others

    Maltoni, F. and others. TF07 Snowmass Report: Theory of Collider Phenomena. 2022. doi:10.2172/1898811. arXiv:2210.02591

    Pith/arXiv arXiv 2022

  57. [57]

    Two invertible networks for the matrix element method

    Butter, Anja and Heimel, Theo and Martini, Till and Peitzsch, Sascha and Plehn, Tilman. Two invertible networks for the matrix element method. SciPost Phys. 2023. doi:10.21468/SciPostPhys.15.3.094. arXiv:2210.00019

    Pith/arXiv arXiv 2023

  58. [58]

    To profile or to marginalize - A SMEFT case study

    Brivio, Ilaria and Bruggisser, Sebastian and Elmer, Nina and Geoffray, Emma and Luchmann, Michel and Plehn, Tilman. To profile or to marginalize - A SMEFT case study. SciPost Phys. 2024. doi:10.21468/SciPostPhys.16.1.035. arXiv:2208.08454

    Pith/arXiv arXiv 2024

  59. [59]

    Hazma meets HERWIG4DM: precision gamma-ray, neutrino, and positron spectra for light dark matter

    Coogan, Adam and Morrison, Logan and Plehn, Tilman and Profumo, Stefano and Reimitz, Peter. Hazma meets HERWIG4DM: precision gamma-ray, neutrino, and positron spectra for light dark matter. JCAP. 2022. doi:10.1088/1475-7516/2022/11/033. arXiv:2207.07634

    Pith/arXiv arXiv 2022

  60. [60]

    Loop amplitudes from precision networks

    Badger, Simon and Butter, Anja and Luchmann, Michel and Pitz, Sebastian and Plehn, Tilman. Loop amplitudes from precision networks. SciPost Phys. Core. 2023. doi:10.21468/SciPostPhysCore.6.2.034. arXiv:2206.14831

    Pith/arXiv arXiv 2023

  61. [61]

    and Favaro, Luigi and Plehn, Tilman and Sorrenson, Peter and Kr

    Dillon, Barry M. and Favaro, Luigi and Plehn, Tilman and Sorrenson, Peter and Kr. A normalized autoencoder for LHC triggers. SciPost Phys. Core. 2023. doi:10.21468/SciPostPhysCore.6.4.074. arXiv:2206.14225

    arXiv 2023

  62. [62]

    Campbell, J. M. and others. Event generators for high-energy physics experiments. SciPost Phys. 2024. doi:10.21468/SciPostPhys.16.5.130. arXiv:2203.11110

    Pith/arXiv arXiv 2024

  63. [63]

    Theory, phenomenology, and experimental avenues for dark showers: a Snowmass 2021 report

    Albouy, Guillaume and others. Theory, phenomenology, and experimental avenues for dark showers: a Snowmass 2021 report. Eur. Phys. J. C. 2022. doi:10.1140/epjc/s10052-022-11048-8. arXiv:2203.09503

    Pith/arXiv arXiv 2021

  64. [64]

    Jets and Jet Substructure at Future Colliders

    Bonilla, Johan and others. Jets and Jet Substructure at Future Colliders. Front. in Phys. 2022. doi:10.3389/fphy.2022.897719. arXiv:2203.07462

    arXiv 2022

  65. [65]

    Machine learning and LHC event generation

    Badger, Simon and others. Machine learning and LHC event generation. SciPost Phys. 2023. doi:10.21468/SciPostPhys.14.4.079. arXiv:2203.07460

    Pith/arXiv arXiv 2023

  66. [66]

    Ephemeral Learning - Augmenting Triggers with Online-Trained Normalizing Flows

    Butter, Anja and Diefenbacher, Sascha and Kasieczka, Gregor and Nachman, Benjamin and Plehn, Tilman and Shih, David and Winterhalder, Ramon. Ephemeral Learning - Augmenting Triggers with Online-Trained Normalizing Flows. SciPost Phys. 2022. doi:10.21468/SciPostPhys.13.4.087. arXiv:2202.09375

    Pith/arXiv arXiv 2022

  67. [67]

    Calomplification the power of generative calorimeter models

    Bieringer, Sebastian and Butter, Anja and Diefenbacher, Sascha and Eren, Engin and Gaede, Frank and Hundhausen, Daniel and Kasieczka, Gregor and Nachman, Benjamin and Plehn, Tilman and Trabs, Mathias. Calomplification the power of generative calorimeter models. JINST. 2022. doi:10.1088/1748-0221/17/09/P09028. arXiv:2202.07352

    Pith/arXiv arXiv 2022

  68. [68]

    and Finke, Thorben and Kr

    Buss, Thorsten and Dillon, Barry M. and Finke, Thorben and Kr. What's anomalous in LHC jets?. SciPost Phys. 2023. doi:10.21468/SciPostPhys.15.4.168. arXiv:2202.00686

    Pith/arXiv arXiv 2023

  69. [69]

    and Kerner, Matthias and Butter, Anja and Heinrich, Gudrun and Plehn, Tilman

    Winterhalder, Ramon and Magerya, Vitaly and Villa, Emilio and Jones, Stephen P. and Kerner, Matthias and Butter, Anja and Heinrich, Gudrun and Plehn, Tilman. Targeting multi-loop integrals with neural networks. SciPost Phys. 2022. doi:10.21468/SciPostPhys.12.4.129. arXiv:2112.09145

    Pith/arXiv arXiv 2022

  70. [70]

    Probing the Inflaton Potential with SKA

    Modak, Tanmoy and Plehn, Tilman and R. Probing the Inflaton Potential with SKA. SciPost Phys. Core. 2022. doi:10.21468/SciPostPhysCore.5.3.037. arXiv:2112.09148

    Pith/arXiv arXiv 2022

  71. [71]

    Generative networks for precision enthusiasts

    Butter, Anja and Heimel, Theo and Hummerich, Sander and Krebs, Tobias and Plehn, Tilman and Rousselot, Armand and Vent, Sophia. Generative networks for precision enthusiasts. SciPost Phys. 2023. doi:10.21468/SciPostPhys.14.4.078. arXiv:2110.13632

    Pith/arXiv arXiv 2023

  72. [72]

    Back to the formula - LHC edition

    Butter, Anja and Plehn, Tilman and Soybelman, Nathalie and Brehmer, Johann. Back to the formula - LHC edition. SciPost Phys. 2024. doi:10.21468/SciPostPhys.16.1.037. arXiv:2109.10414

    Pith/arXiv arXiv 2024

  73. [73]

    and Kasieczka, Gregor and Olischlager, Hans and Plehn, Tilman and Sorrenson, Peter and Vogel, Lorenz

    Dillon, Barry M. and Kasieczka, Gregor and Olischlager, Hans and Plehn, Tilman and Sorrenson, Peter and Vogel, Lorenz. Symmetries, safety, and self-supervision. SciPost Phys. 2022. doi:10.21468/SciPostPhys.12.6.188. arXiv:2108.04253

    Pith/arXiv arXiv 2022

  74. [74]

    From models to SMEFT and back?

    Brivio, Ilaria and Bruggisser, Sebastian and Geoffray, Emma and Killian, Wolfgang and Kr. From models to SMEFT and back?. SciPost Phys. 2022. doi:10.21468/SciPostPhys.12.1.036. arXiv:2108.01094

    Pith/arXiv arXiv 2022

  75. [75]

    Unsupervised hadronic SUEP at the LHC

    Barron, Jared and Curtin, David and Kasieczka, Gregor and Plehn, Tilman and Spourdalakis, Aris. Unsupervised hadronic SUEP at the LHC. JHEP. 2021. doi:10.1007/JHEP12(2021)129. arXiv:2107.12379

    Pith/arXiv arXiv 2021

  76. [76]

    Shared Data and Algorithms for Deep Learning in Fundamental Physics

    Benato, Lisa and others. Shared Data and Algorithms for Deep Learning in Fundamental Physics. Comput. Softw. Big Sci. 2022. doi:10.1007/s41781-022-00082-6. arXiv:2107.00656

    Pith/arXiv arXiv 2022

  77. [77]

    and Plehn, Tilman and Sauer, Christof and Sorrenson, Peter

    Dillon, Barry M. and Plehn, Tilman and Sauer, Christof and Sorrenson, Peter. Better Latent Spaces for Better Autoencoders. SciPost Phys. 2021. doi:10.21468/SciPostPhys.11.3.061. arXiv:2104.08291

    Pith/arXiv arXiv 2021

  78. [78]

    Understanding Event-Generation Networks via Uncertainties

    Bellagente, Marco and Haussmann, Manuel and Luchmann, Michel and Plehn, Tilman. Understanding Event-Generation Networks via Uncertainties. SciPost Phys. 2022. doi:10.21468/SciPostPhys.13.1.003. arXiv:2104.04543

    Pith/arXiv arXiv 2022

  79. [79]

    and Keilbach, Fabian and Plehn, Tilman and Kasieczka, Gregor and Whiteson, Daniel

    Baldi, Pierre and Blecher, Lukas and Butter, Anja and Collado, Julian and Howard, Jessica N. and Keilbach, Fabian and Plehn, Tilman and Kasieczka, Gregor and Whiteson, Daniel. How to GAN Higher Jet Resolution. SciPost Phys. 2022. doi:10.21468/SciPostPhys.13.3.064. arXiv:2012.11944

    Pith/arXiv arXiv 2022

  80. [80]

    Measuring QCD Splittings with Invertible Networks

    Bieringer, Sebastian and Butter, Anja and Heimel, Theo and H. Measuring QCD Splittings with Invertible Networks. SciPost Phys. 2021. doi:10.21468/SciPostPhys.10.6.126. arXiv:2012.09873

    Pith/arXiv arXiv 2021

Showing first 80 references.