Development and performance of a transformer-based lepton isolation algorithm with the ATLAS experiment

arXiv:2611.08194 · hep-ex · Submitted 2026-11-05 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: High Energy Physics: Experiment. The day in high-energy experiment: detectors, searches and limits.

Kai: We're looking at the paper: "Development and performance of a transformer-based lepton isolation algorithm with the ATLAS experiment".

Mira: Here's the gist of the paper. The Prompt Lepton Isolation Tagger (PLIT) is a transformer-based algorithm developed to distinguish prompt leptons (electrons and muons) originating from electroweak gauge boson decays from non-prompt leptons arising from hadron decays and misidentified leptons.

Kai: PLIT leverages information from associated charged-particle tracks, combined with lepton properties, utilizing a transformer neural network to identify correlations that enhance discrimination.

Mira: Its performance was evaluated using simulated topquark pair production events against existing lepton isolation algorithms, demonstrating superior prompt lepton efficiency while maintaining high rejection of non-prompt leptons.

Kai: PLIT is specifically designed to separate prompt leptons from those predominantly produced in semileptonic decays of heavy-flavor hadrons.

Paper summary: Kai: So we’re looking at this paper now, "Development and performance of a transformer-based lepton isolation algorithm with the ATLAS experiment." It's about this new tool called PLIT, which is a transformer based algorithm designed to separate prompt leptons from non-prompt ones.

Mira: Right. The core idea here is using a transformer neural network to look at the tracks around a lepton candidate along with its properties to find correlations that help discriminate between the two sources of leptons. It's being used in measurements where you have lots of leptons, and you need to reject those from heavy-flavor hadron decays.

Lev: From my side, I'm interested in how much real hardware overhead this thing adds. The paper says they trained it on substantially larger simulated topquark pair production events compared to previous methods, which is a big data effort for a machine learning model like this.

Kai: Exactly. They used these larger samples to improve the discrimination between prompt leptons and those coming from semileptonic decays of heavy-flavor hadrons. This is directly aimed at reducing that single largest systematic uncertainty in the ATLAS measurement of the topquark pair production process, which was that non-prompt lepton background.

Mira: The paper claims PLIT demonstrates improved prompt lepton efficiency while maintaining a high rejection of those non-prompt leptons when compared to existing algorithms. It seems to be tackling this problem by looking at how tracks are associated with the lepton candidate within a cone of less than zero point four radians in radius.

Lev: That track information is key, and I wonder what the computational cost really is when you have up to forty tracks being fed into this transformer network for every candidate. We need to make sure that overhead doesn't swamp the actual physics signal we are looking for on real hardware.

Kai: The architecture itself is based on the GN2 transformer architecture that ATLAS uses for heavy-flavor jet tagging, but they adapted it specifically for lepton isolation. This means they’re using a well-established structure but applying it to this new task of lepton separation.

Paper summary: Mira: And the output of this network is a probability, p prompt, and then they build their final discriminant by taking the logarithm of the ratio between that probability and one minus that probability. That way you get their final PLIT discriminant, D PLIT = (p prompt / (one - p prompt)) <ref:2611.08194#pg2>.

Lev: So they’re moving from a BDT approach to something based on a transformer architecture for the isolation task. That shift in methodology suggests they are looking for more complex, non-linear correlations between the lepton and its surrounding activity.

Kai: And the performance claims are quite specific. For muons at low transverse momentum, they say the rejection is one point five to two times larger than PLIV for any given efficiency level.

Mira: And for electrons, they report a rejection gain of a factor of two or more across both p T ranges. Specifically, at low efficiency, that electron rejection can reach a factor of five for endcap electrons.

Lev: That’s interesting because the performance gains are varying by lepton type and by kinematic region. It shows the model isn't universally better everywhere; it has different strengths depending on what you are looking at.

Kai: And in terms of efficiency for the same rejection level, they show relative improvements of ten percent for muons, fifty percent for barrel electrons, and a hundred percent improvement for endcap electrons at low p T.

Mira: These relative efficiency gains are significant because they show that PLIT doesn't just reject more background; it also lets you keep more of the actual prompt signal. That is important when you are trying to measure things precisely.

Lev: The paper does mention that operating points were defined and their efficiencies calibrated using data with the standard tag-and-probe method applied to Z to l plus l minus events, which gives them those data-to-simulation correction factors, or SFs.

Kai: That calibration process is crucial for translating the simulation results into what we can actually expect in real detector data. And it seems the dominant uncertainty they found comes from modeling the hadronic activity in the simulation itself—specifically the difference between nominal Powheg plus Pythia and an alternative Sherpa sample.

Paper summary: Mira: That MC modeling issue is a known hurdle, but they state that all other systematic uncertainties, like template shape or QCD multijet background description, stay below one point zero percent across the whole transverse momentum range.

Lev: So the main remaining challenge for this algorithm’s performance seems to be how well we can model those underlying hadronic interactions in our simulations. That points toward future work focusing on better modeling of those specific processes.

Kai: This paper presents the development, performance, and calibration of PLIT, which is a transformer based lepton isolation algorithm designed to distinguish prompt leptons from non-prompt ones produced in hadron decays using information from associated charged-particle tracks and lepton properties.

Mira: The authors are showing that this approach leads to improved prompt lepton efficiency while maintaining high rejection of non-prompt leptons when compared with existing algorithms, using simulated topquark pair production events.

Lev: Ultimately, they’re showing how to use a transformer network adapted from jet tagging to solve the specific problem of lepton isolation in a way that improves background suppression in multilepton final states.

Kai: Looking at the title and authors, it seems like this work is focused on developing a new technique to handle the significant non-prompt lepton background that plagues measurements at the LHC when you're looking for prompt leptons.

Mira: It suggests a direction where leveraging transformer networks on track information can provide a more powerful way to isolate genuine electroweak boson decays from the much larger background from heavy-flavor hadron decays.

Lev: For running this on real hardware, it means we need to figure out how efficiently we can extract those forty tracks and feed them into that network without introducing too much latency or requiring excessive memory.

Kai: So, the implication is that if this method holds up in data as well as in simulation, it gives us a better way to measure the properties of top quark pair production events by cleaning up the final state lepton sample.

Conclusion: Kai: So we're looking at this paper about PLIT, which is a transformer based algorithm from the ATLAS experiment for separating prompt leptons from non-prompt ones, and it’s been really interesting to look at how they built and measured this stuff.

Mira: Yeah. The authors are showing that by using a neural network architecture adapted from jet tagging to isolate leptons, they can get better results than what we had before on simulated top quark pair production events.

Lev: From my side, it’s interesting because they trained this thing on much larger simulated samples than previous methods, which is a big data effort for training a transformer like this.

Kai: Right. The paper shows that even though the architecture is based on something used for jet tagging, they successfully adapted it for lepton isolation using track information around the candidate lepton.

Mira: It means they are leveraging those forty tracks within a small radius, less than zero point four radians, to find correlations that help discriminate between prompt and non-prompt leptons.

Lev: And from an engineering view, the network outputs a probability for a prompt lepton, and they build their final discriminant by taking the logarithm of the ratio between that probability and one minus that probability.

Kai: The performance numbers are pretty detailed too. They show gains for muons at low transverse momentum and significant rejection improvements for electrons across different kinematic regions.

Mira: It’s also important because they calibrated these operating points using data from Z to lepton pairs, which gives them correction factors to translate simulation into what we actually see in the detector.

Lev: The main caveat they flag is modeling the hadronic activity in their simulations, specifically how Powheg plus Pythia compares to a different sample like Sherpa.

Kai: So what this paper really means is that using these transformer networks on track information gives us a new way to clean up the final state leptons from top quark measurements by suppressing those non-prompt background sources.

Mira: It suggests we can push the limits of what we can measure in processes like top quark pair production if we can get the simulation modeling of hadronic activity right.

Lev: And it opens up a path for using transformer networks in other areas where complex correlations between many particles are needed, which is something I’m always interested in exploring for real hardware.

The ATLAS Collaboration

EUROPEAN ORGANISATION FOR NUCLEAR RESEARCH (CERN)

hep-ex

Submitted: 2026-11-05

Updated: 2026-11-05

Comments: 42 pages in total, author list starting page 26, 14 figures, 2 tables, submitted to EPJC. All figures including auxiliary figures are available at https://atlas.web.cern.ch/Atlas/GROUPS/PHYSICS/PAPERS/MUON-2024-01

License: http://creativecommons.org/licenses/by/4.0/

Importance score: 92/100

The gist: The Prompt Lepton Isolation Tagger (PLIT) is a transformer-based algorithm developed to distinguish prompt leptons (electrons and muons) originating from electroweak gauge boson decays from…

Key concepts

Prompt Lepton Isolation Tagger (PLIT)
A transformer-based algorithm developed to distinguish prompt leptons from non-prompt ones originating from hadron decays. It uses associated charged-particle tracks and lepton properties within a neural network to find correlations that improve discrimination.
Transformer Neural Network
The core machine learning architecture used in PLIT, adapted from the GN2 transformer used for heavy-flavor jet tagging. This network processes kinematic and detector data of the lepton along with information from up to 40 tracks within a small cone around the candidate lepton.
Discriminant ($D_{ ext{PLIT}}$)
The final output score calculated by PLIT, defined as the logarithm of the ratio between the probability that a lepton is prompt and the probability that it is not prompt. This value is used to classify whether a detected lepton originates from an electroweak decay or a hadronic decay.
Data-to-Simulation Efficiency Correction Factors (SFs)
Calibration factors derived by applying standard tag-and-probe methods to $Z o l^+l^-$ events. These factors are used to correct the efficiency measured in simulated data, helping to map simulation results accurately onto real experimental measurements.

Terminology

Summary

The Prompt Lepton Isolation Tagger (PLIT) is a transformer-based algorithm developed to distinguish prompt leptons (electrons and muons) originating from electroweak gauge boson decays from non-prompt leptons arising from hadron decays and misidentified leptons. p. 0

Where the paper says it

EUROPEAN ORGANISATION FOR NUCLEAR RESEARCH (CERN) Submitted to: EPJC CERN-EP-2026-292 6th November 2026 Development and performance of a transformer-based lepton isolation algorithm with the ATLAS experiment The ATLAS Collaboration The Prompt Lepton Isolation Tagger (PLIT) is a transformer-based algorithm designed to distinguish prompt leptons (electrons and muons) produced in the primary decays of electroweak gauge bosons from non-prompt leptons produced in hadron decays and from misidentified leptons. It is primarily used in measurements targeting multilepton final states with important backgrounds from non-prompt and misidentified leptons.

PLIT leverages information from associated charged-particle tracks, combined with lepton properties, utilizing a transformer neural network to identify correlations that enhance discrimination. p. 0
Where the paper says it

It is primarily used in measurements targeting multilepton final states with important backgrounds from non-prompt and misidentified leptons. PLIT exploits information from charged-particle tracks associated with the lepton candidate, together with lepton properties, using a transformer neural network to identify correlations between the tracks and improve the discrimination.

Its performance was evaluated using simulated topquark pair production events against existing lepton isolation algorithms, demonstrating superior prompt lepton efficiency while maintaining high rejection of non-prompt leptons. p. 0
Where the paper says it

PLIT exploits information from charged-particle tracks associated with the lepton candidate, together with lepton properties, using a transformer neural network to identify correlations between the tracks and improve the discrimination. Its performance is evaluated using simulated topquark pair production events and compared with that of previously used lepton isolation algorithms, demonstrating improved prompt lepton efficiency while maintaining high rejection of non-prompt leptons.

PLIT is specifically designed to separate prompt leptons from those predominantly produced in semileptonic decays of heavy-flavor hadrons. p. 2

Where the paper says it

This paper presents the Prompt Lepton Isolation Tagger (PLIT), developed using a transformer neural network architecture [9] based on the GN2 jet flavour-tagging algorithm developed by ATLAS [10]. PLIT is designed to distinguish prompt leptons from non-prompt leptons produced predominantly in semileptonic decays of heavy-flavour hadrons.

The training utilized substantially larger simulated t bar event samples compared to previous methods, and PLIT integrates conventional isolation observables with detailed track information reconstructed within a cone around the lepton candidate, employing the same underlying track representation as the GN2 algorithm. p. 2
Where the paper says it

The training is performed using simulated 𝑡𝑡¯ events, with substantially larger simulated samples available than for the training of the previous PLIV algorithm. PLIT combines conventional isolation observables with detailed information from charged-particle tracks reconstructed within a cone around the lepton candidate, using the same underlying track representation as the GN2 algorithm.

A significant advantage is that PLIT does not rely on jet reconstruction or jet flavor-tagging tools. p. 2
Where the paper says it

PLIT combines conventional isolation observables with detailed information from charged-particle tracks reconstructed within a cone around the lepton candidate, using the same underlying track representation as the GN2 algorithm. As a consequence, PLIT does not depend on jet reconstruction or jet flavour-tagging tools.

The architecture of PLIT is based on the GN2 transformer architecture used by ATLAS for heavy-flavor jet tagging, adapted for lepton isolation. p. 6

Where the paper says it

5 The PLIT transformer neural network: architecture and training The PLIT architecture is based on the GN2 transformer architecture developed in ATLAS for heavy-flavour jet tagging [10], adapted to the problem of lepton isolation. The Salt framework [48] is used to train the transformer.

The network processes kinematic and detector properties of the lepton alongside information from up to 40 tracks within a cone of delta R < 0. p. 6
Where the paper says it

The Salt framework [48] is used to train the transformer. The algorithm processes the kinematic and detector-related properties of the lepton together with information from tracks reconstructed in its vicinity: for each lepton candidate, up to 40 tracks within a cone of Δ𝑅 < 0.

4 around the candidate lepton. p. 1
Where the paper says it

Both the promptness and isolation properties can be used to achieve a good separation. Conventional isolation requirements use variables based on the scalar sum of the transverse momentum of charged-particle tracks and the transverse energy deposited in the calorimeter within a cone around the lepton candidate [2, 3].

The network outputs a prompt lepton probability, p prompt, from which the PLIT discriminant is constructed as D PLIT = (p prompt / (1 - p prompt)). p. 6
Where the paper says it

The primary objective of PLIT is the classification of leptons into prompt and non-prompt categories. The network outputs the prompt lepton probability, denoted by 𝑝prompt, from which the PLIT discriminant is constructed as 𝐷PLIT = log 𝑝prompt 1 − 𝑝prompt .

Performance comparisons show that PLIT outperforms the previous PLIV algorithm in both p T ranges and across the entire efficiency range. p. 2

Where the paper says it

PLIT is designed to distinguish prompt leptons from non-prompt leptons produced predominantly in semileptonic decays of heavy-flavour hadrons. The training is performed using simulated 𝑡𝑡¯ events, with substantially larger simulated samples available than for the training of the previous PLIV algorithm.

For muons, at low p T, the rejection is 1. p. 2
Where the paper says it

Given the large amount of non-prompt lepton background at low transverse momentum, PLIV operated at low prompt lepton efficiency (below 50%) in that region of phase space. Improving the prompt lepton efficiency while maintaining the same level of non-prompt lepton rejection would significantly increase the signal acceptance in multilepton measurements.

5 to 2 times larger than PLIV for any given efficiency. p. 9
Where the paper says it

Muons satisfying the Medium identification OP and electrons satisfying the Loose identification OP are used for training and performance evaluation, with the looser electron selection providing increased training sample size. Samples are ten times larger than the ones used previously to train PLIV; they contain 75 million muons and 68 million electrons for training, and 15 million muons and 10 million electrons for validation and testing, with a proportion of 80% of prompt and 20% of non-prompt leptons.

For electrons, the rejection gain is a factor of 2 or more across both p T ranges, reaching up to a factor of 5 for endcap electrons at low efficiency. p. 5
Where the paper says it

The resulting dataset corresponds to a total integrated luminosity of 166 ± 3 fb−1 for muons and 165 ± 3 fb−1 for electrons [45–47]. 2 The ℎdamp parameter is a resummation damping factor and one of the parameters that controls the matching of Powheg matrix elements to the parton shower and thus effectively regulates the high-𝑝T radiation against which the hard-process system recoils.

In terms of prompt lepton efficiency for the same rejection level, PLIT achieves relative improvements: 10% for muons, 50% for barrel electrons, and 100% for endcap electrons at low p T. p. 0
Where the paper says it

Its performance is evaluated using simulated topquark pair production events and compared with that of previously used lepton isolation algorithms, demonstrating improved prompt lepton efficiency while maintaining high rejection of non-prompt leptons. This article presents the development, performance, and calibration of PLIT for electrons and muons in proton–proton collisions at a centre-of-mass energy of √ 𝑠 = 13.

Operating points (OPs) were defined, and their prompt lepton efficiencies were calibrated in data using the standard tag-and-probe method applied to Z to l+l- events. p. 0

Where the paper says it

6 TeV recorded by the ATLAS experiment at the Large Hadron Collider. Operating points are defined for both electrons and muons, and their efficiencies are calibrated using 𝑍 → 𝑙 + 𝑙 − events in Run 3 data, providing data-to-simulation correction factors for physics analyses.

This calibration yields data-to-simulation efficiency correction factors (SFs). p. 2
Where the paper says it

The performance of the algorithm in simulated events, including the definition and evaluation of the OPs, is described in Section 6. The calibration of the algorithm and the derivation of data-to-simulation efficiency correction factors are presented in Section 7.

The dominant uncertainty for both leptons stems from the modeling of hadronic activity in the simulation (MC modeling), specifically the difference between nominal Powheg+Pythia and an alternative Sherpa sample. All other systematic uncertainties, including template shape, QCD multijet background description, and fit range choice (lambda-SC), remain sub-dominant below 0. p. 2
Where the paper says it

In the measurements for which PLIT is intended, light hadron decays, misidentified lepton candidates, and photon conversions misidentified as electrons, constitute subdominant sources of background alongside the dominant contribution from heavy-flavour hadron decays; these contributions are collectively referred to as “non-prompt leptons”. In the ATLAS measurement of the 𝑡𝑡𝐻¯ process in multilepton final states using LHC Run 1 data [4], the uncertainty on the non-prompt lepton background was the single largest systematic uncertainty.

1% over the full p T range. p. 8
Where the paper says it

ReLU activations and cross-entropy loss functions are used throughout the architecture. The final loss is constructed as a weighted combination of the three task losses, enabling simultaneous optimisation of the full network.

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Improve lepton discrimination by using a transformer neural network architecture based on GN2 jet flavour tagging to distinguish prompt leptons from non-prompt leptons produced predominantly in semileptonic decays of heavy-flavour hadrons.

  2. Enhance rejection power by exploiting information from charged-particle tracks reconstructed within a cone around the lepton candidate, as PLIT combines conventional isolation observables with detailed information from charged-particle tracks reconstructed within a cone around the lepton candidate.

  3. Enable implicit vertexing information learning by using auxiliary objectives during training that classify each track according to its production origin and predict whether pairs of tracks originate from a common vertex.

  4. Increase discrimination power by constructing the PLIT discriminant using DPLIT = log(pprompt / (1 - pprompt)), which is more practical to use than pprompt because the algorithm typically operates at values of pprompt that are very close to 1.

  5. Optimize operating points by defining two PLIT OPs for each conventional PLIV OP: one matching the PLIV efficiency and another matching the PLIV rejection, as described in section 6.2.

  6. Calibrate prompt lepton efficiency using a data-to-simulation correction factor derived from the tag-and-probe method applied to Z plus Z minus events to provide unbiased measurements for simulation scaling factors (SFs).

Abstract

The Prompt Lepton Isolation Tagger (PLIT) is a transformer-based algorithm designed to distinguish prompt leptons (electrons and muons) produced in the primary decays of electroweak gauge bosons from non-prompt leptons produced in hadron decays and from misidentified leptons. It is primarily used in measurements targeting multilepton final states with important backgrounds from non-prompt and misidentified leptons. PLIT exploits information from charged-particle tracks associated with the lepton candidate, together with lepton properties, using a transformer neural network to identify correlations between the tracks and improve the discrimination. Its performance is evaluated using simulated top-quark pair production events and compared with that of previously used lepton isolation algorithms, demonstrating improved prompt lepton efficiency while maintaining high rejection of non-prompt leptons. This article presents the development, performance, and calibration of PLIT for electrons and muons in proton--proton collisions at a centre-of-mass energy of sqrt s =13.6 TeV recorded by the ATLAS experiment at the Large Hadron Collider. Operating points are defined for both electrons and muons, and their efficiencies are calibrated using Z to l+l- events in Run 3 data, providing data-to-simulation correction factors for physics analyses.

Sources

Related papers