Pith. sign in

REVIEW 6 minor 13 references

Searches for heavy neutral leptons with machine learning at the CMS experiment

T0 review · 0 major / 6 minor · reviewed 2026-08-11 · deepseek-v4-flash

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

Pith's one-line read CMS searches for heavy neutral leptons find no signal and set new exclusion limits.

desk verdict A faithful, internally consistent proceedings summary of two already-published CMS HNL searches; it adds no new physics but should not be judged as a standalone research paper. read the letter →

arxiv 2412.06298 v1 pith:M55ELXBI submitted 2024-12-09 hep-ex

Joscha Knolle (for the CMS Collaboration) This is my paper · ORCID
classification hep-ex PACS 14.60.Pq14.60.St
keywords heavyneutralleptonsCMStrileptonsearchdisplaceddileptonmachinelearningboosteddecisiontreesparticleflownetworksexclusionlimits
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

Two CMS searches for heavy neutral leptons (HNLs) in 138 fb⁻¹ of 13 TeV proton-proton collision data are presented, one for promptly decaying HNLs in the 10 GeV to 1.5 TeV mass range and one for long-lived HNLs of 1 to 20 GeV with displaced decays. Both searches use machine-learned discriminators to suppress standard model backgrounds, and both observe no significant deviation from the background expectation. The results are expressed as 95% confidence-level exclusion limits on the coupling strength $|V_{\ell N}|^2$ as a function of HNL mass, and in several coupling scenarios they improve on previous limits. The notable new territory is the first probe of exclusive tau-neutrino couplings for HNL masses above the W boson mass, which had not been covered before.

What carries the argument

The machinery is the machine-learned discriminator paired with a data-driven background estimate. In the prompt search, boosted decision trees (BDTs) with up to 43 kinematic input variables are trained separately for mass ranges, coupling scenarios, and final-state flavour combinations, and the BDT score distributions enter maximum-likelihood fits. In the displaced search, particle flow networks (PFNs) — deep networks that act on unordered sets of up to 50 reconstructed particles plus event-level variables — produce a score that, together with the lepton–secondary-vertex invariant mass $m(l,\mathrm{SV})$, defines signal and sideband regions for the ABCD method. These discriminators are what separate a small HNL signal from large standard model backgrounds and turn event counts into exclusion limits.

What would settle it

If a future experiment or a reprocessing of CMS data found a resonant excess in the trilepton or displaced-dilepton spectrum at a mass and mixing already excluded by these curves, the exclusion would be wrong. A direct check is a closure test: in a background-dominated control sample, compute the ABCD prediction from the two sidebands and compare it with the observed yield in the corresponding signal-like box; a significant mismatch would invalidate the displaced search's background estimate and its limits.

Watch

Extended reading notes

Core claim

The central claim is that, with 138 fb⁻¹ of 13 TeV data, CMS finds no evidence of heavy neutral leptons and therefore excludes HNL couplings that previous searches had left open. The prompt trilepton analysis, based on boosted decision tree scores in maximum-likelihood fits, yields the most stringent limits over a wide mass range. The displaced dilepton analysis, based on particle flow network scores and an ABCD background estimate, provides competitive limits that are particularly important for long-lived HNLs with masses of 10 to 20 GeV. For a Majorana HNL coupling exclusively to tau neutrinos, the prompt search reaches above the W mass for the first time, extending beyond the earlier DELPHI bound.

Load-bearing premise

Both searches assume HNLs are produced through a Drell–Yan W boson decaying to a lepton and the HNL, with the production rate and lifetime set by the same mixing parameter $|V_{\ell N}|^2$; the displaced search further assumes its ABCD method can extrapolate the background from the sidebands, defined by the particle-flow-network score and the lepton–secondary-vertex mass, into the signal region.

Editorial extensions

If this is right

  • If the limits are correct, HNLs with masses from about 1 GeV to 1.5 TeV are ruled out at 95% confidence for the tested mixing strengths in the electron- and muon-coupling scenarios.
  • The prompt trilepton channel becomes the most stringent constraint over a wide mass interval, superseding the earlier same-sign and displaced-lepton CMS searches shown for comparison.
  • The first tau-exclusive exclusion above the W boson mass closes a region of parameter space that had been open since the DELPHI experiment.
  • The displaced dilepton channel provides the strongest constraint for long-lived HNL scenarios at 10 to 20 GeV, where prompt and displaced searches overlap.

Reading between the lines

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

  • The same PFN-plus-ABCD recipe could be transplanted to other long-lived particle searches, such as hidden-sector scalars or dark photons, by re-training the discriminator on the alternative signal model; the background machinery would not need to change.
  • A direct extension of the tau-exclusive result is to push the same search to lower masses and mixings with Run 3 data, since the BDT training is already separated by coupling scenario.
  • The displaced search's validity hinges on the ABCD assumption; a public closure check in control samples enriched in top-quark and diboson production would make the claim testable at the level of the background model.
  • Cross-interpreting the two searches at their boundary (masses near 10 to 20 GeV) could yield a combined HNL limit that is stronger than either channel alone, an analysis the paper does not perform.
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

0 major / 6 minor

Summary. This proceedings contribution by J. Knolle on behalf of the CMS Collaboration summarizes two CMS searches for heavy neutral leptons (HNLs) using 138 fb^-1 of 13 TeV proton-proton collision data. The prompt trilepton search covers HNL masses from 10 GeV to 1.5 TeV and uses boosted decision trees with up to 43 kinematic variables; the displaced dilepton search covers 1-20 GeV and uses particle-flow networks together with an ABCD method. The paper reports that no significant deviations from the standard model background are observed and presents 95% CL exclusion limits on the HNL mixing parameter |V_lN|^2 for Dirac and Majorana HNLs under different coupling scenarios. It further states that the prompt search gives the most stringent limits over a wide mass range and probes exclusive tau-neutrino couplings above the W mass for the first time. The manuscript is a faithful, high-level summary of Refs. [5] and [6] rather than a new standalone analysis.

Significance. If taken at face value, the summarized results exclude a substantial region of HNL parameter space, and the claimed extension of exclusive tau-coupling limits above the W mass is of phenomenological interest. The paper's value lies in providing a concise, readable entry point to two detailed CMS analyses, and it correctly attributes all results to the cited references. I considered the stress-test concern that the reported limits depend on Drell-Yan W->lN production and on the closure of the ABCD sideband extrapolation; in my reading this does not identify a defect in the present manuscript, because the paper makes no independent statistical claim and explicitly relies on Refs. [5] and [6] for the supporting evidence. The absence of systematic tables, closure tests, and full statistical machinery is appropriate for a conference proceedings contribution.

minor comments (6)
  1. [References, Ref. [6]] Section 1 refers to the second search as "published in Ref. [6]", but the reference list describes Ref. [6] as "Submitted to JHEP". Please correct the wording or update the reference status so that the two statements are consistent.
  2. [Section 3] The text says the PFN score is "one of two variables" in the ABCD method but never names the second variable in the body; it appears only in the figure label as m(l,SV). Please state explicitly that the second variable is the invariant mass m(l,SV) and define it in the text.
  3. [Fig. 2 (right)] The legend entries in Fig. 2 (right) include unexplained abbreviations such as "VDY" and "HNL3/HNL5/HNL6/HNL10". Please expand these labels or explain them in the caption so that the background and signal samples are identifiable without consulting Ref. [6].
  4. [Fig. 3] The upper panels of Fig. 3 show many overlapping curves from different CMS searches, and the luminosity label "35.9 138fb-1" is ambiguous. Please clarify which data sets correspond to which curves and distinguish the observed and expected curves of the prompt trilepton search more clearly.
  5. [Throughout (Sections 1-4)] The body text contains numerous missing spaces between words, for example "13TeV", "ppcollision", "trackercanbereconstructed", and "HNLdecay". If this reflects the compiled PDF rather than an extraction artifact, the LaTeX source should be fixed so that the proceedings text is properly typeset.
  6. [Abstract and Section 4] The phrase "results exceed previous limits" is ambiguous: exclusion limits can be described as stronger, more stringent, or surpassing previous limits. Please rephrase to avoid a reading in which the limits are numerically larger rather than more constraining.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the reported limits are experimental outputs from collision data, not derivations from their own inputs.

full rationale

No circular step is present. This manuscript is a proceedings summary of two peer-reviewed CMS analyses; every quantitative claim (the 95% CL exclusion limits on |V_lN|^2, the mass ranges, and the use of BDT/PFN scores) is an experimental result constructed from collision data and simulated signal samples, not a quantity defined in terms of the very result it is supposed to establish. The background estimates are data-driven (ABCD sideband extrapolation in the displaced search and maximum-likelihood fits in the prompt search), and any underlying modeling assumptions, such as Drell-Yan W->lN production or ABCD closure, are external physics and statistical assumptions of the referenced analyses, not tautological relations. The heavy use of CMS self-citations is normal collaboration practice and is not load-bearing in a circular sense: Refs. [5] and [6] are cited as the sources of the results, and the limits stand on the referenced data analysis rather than on the citations themselves. No equation or argument in this summary defines an output in terms of the input that it claims to predict, so there is no derivation chain that reduces to its own inputs.

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

The scanned HNL mass and mixing angle are model parameters being excluded, not fitted degrees of freedom, so they are not free parameters in the ledger. The HNL itself is not an invented entity of this paper; it is a pre-existing theoretical proposal that the search targets. The ledger records only the physical and methodological assumptions inherited from the underlying analyses.

assumptions (3)
  • domain assumption HNLs are produced in pp collisions through Drell-Yan W -> l N, with cross-section and decay length set by the mixing parameter |V_lN|^2.
    Invoked in Sec. 1 and Fig. 1 to define the signal signatures and to translate event counts into coupling limits.
  • domain assumption For m_N < 20 GeV the HNL can travel far enough for a displaced secondary vertex reconstructed by the CMS tracker; for m_N > 20 GeV decay is prompt.
    This lifetime/detector assumption separates the two searches and sets the mass ranges discussed in Secs. 1 and 3.
  • domain assumption The ML classifiers and the ABCD sideband method give unbiased background estimates in the signal regions.
    The prompt limits rely on BDT-score fits and the displaced search relies on ABCD extrapolation (Secs. 2 and 3); closure is not demonstrated in this proceedings.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Searches for heavy neutral leptons with machine learning at the CMS experiment." pith.science (2026). https://pith.science/paper/M55ELXBI

@misc{pith2026241206298,
  author       = {Pith},
  title        = {Pith review of: Searches for heavy neutral leptons with machine learning at the CMS experiment},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/M55ELXBI}},
  note         = {Machine review of arXiv:2412.06298}
}
read the original abstract

Two recent searches for heavy neutral leptons (HNLs) performed with proton-proton collision data recorded at 13 TeV by the CMS experiment are presented. A prompt search in the trilepton final state analyses events with exactly three charged leptons originating from the primary proton-proton interaction vertex, targeting HNL masses between 10 GeV and 1.5 TeV. A displaced search in the dilepton final state analyses events with exactly one prompt charged lepton and a second nonprompt charged lepton associated with a jet and a secondary vertex, targeting HNL masses between 1 and 20 GeV. In both searches, machine-learning methods are applied to separate the HNL signal from the standard model background. Exclusion limits are set on the HNL coupling strength as a function of the HNL mass, covering different mass ranges and HNL scenarios. In several cases, the results exceed previous limits.

Figures

Figures reproduced from arXiv: 2412.06298 by the authors.

Figure 1
Figure 1. Examples of leading-order Feynman diagrams for production and decay of an HNL (indicated with the symbol N) resulting in a final state with three charged leptons (left) or two charged leptons and two quarks (right). Here, the HNL is a Majorana particle, and thus the charged leptons from the HNL production and decay vertex can have the same charge. In this contribution, two HNL searches are presented. The first searc… view at source ↗
Figure 2
Figure 2. Left: Distribution of the BDT score trained with the 10–40 GeV HNL mass range and muon neutrino couplings for the combined µµµ and eµµ categories of the prompt search in the trilepton final state, from Ref. [5]. Right: Distribution of the high-mass PFN score for the combined ee and µe categories of the displaced search in the dilepton final state, from Ref. [6]. 3. Displaced search in dilepton final state Events are… view at source ↗
Figure 3
Figure 3. Exclusion limits at 95% CL on the HNL mixing parameter as a function of the HNL mass for a Majorana HNL with exclusive electron (upper left), muon (upper right), or tau (lower, from Ref. [5]) neutrino coupling. The solid (dashed) curve indicates the observed (expected) exclusion. References [1] C. Antel et al., Eur. Phys. J. C 83 (2023) 1122, doi:10.1140/epjc/s10052-023-12168-5. [2] CMS Collaboration, JINST 3 (2008)… view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

13 extracted references · 3 canonical work pages

  1. [5]

    CMS Collaboration, JHEP 06(2024) 123,doi:10.1007/JHEP06(2024)123

    doi:10.1007/jhep06(2024)123 2024

  2. [6]

    Submitted toJHEP

    CMS Collaboration, 2024.arXiv:2407.10717. Submitted toJHEP

    arXiv 2024

  3. [1]

    Antel et al.,Eur

    C. Antel et al.,Eur. Phys. J. C83(2023) 1122,doi:10.1140/epjc/s10052-023-12168-5

    doi:10.1140/epjc/s10052-023-12168-5 2023

  4. [2]

    CMS Collaboration, JINST 3 (2008) S08004,doi:10.1088/1748-0221/3/08/S08004

    doi:10.1088/1748-0221/3/08/s08004 2008

  5. [3]

    CMS Collaboration, JINST 19(2024) P05064,doi:10.1088/1748-0221/19/05/P05064

    doi:10.1088/1748-0221/19/05/p05064 2024

  6. [4]

    Accepted byPhys

    CMS Collaboration, 2024.arXiv:2405.17605. Accepted byPhys. Rept

    arXiv 2024

  7. [7]

    CMS Collaboration, JHEP 01(2019) 122,doi:10.1007/JHEP01(2019)122

    doi:10.1007/jhep01(2019)122 2019

  8. [8]

    CMS Collaboration, JHEP 07(2022) 081,doi:10.1007/JHEP07(2022)081

    doi:10.1007/jhep07(2022)081 2022

Show all 13 references
  1. [9]

    CMS Collaboration, Phys. Rev. Lett.131 (2023) 011803, doi:10.1103/PhysRevLett.131.011803

    2023 doi

  2. [10]

    CMS Collaboration, JHEP 03(2024) 105,doi:10.1007/JHEP03(2024)105

    2024 doi

  3. [11]

    CMS Collaboration, Phys. Rev. D110 (2024) 012004,doi:10.1103/PhysRevD.110.012004

    2024 doi

  4. [12]

    CMS Collaboration, JHEP 06(2024) 183,doi:10.1007/JHEP06(2024)183

    2024 doi

  5. [13]

    DELPHI Collaboration, Z. Phys. C74(1997) 57,doi:10.1007/s002880050370. 4

    1997 doi

Pith tools

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