A machine-learning method for the estimation of the background from jets misidentified as hadronic tau lepton decays
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: "A machine-learning method for the estimation of the background from jets misidentified as hadronic tau lepton decays".
Paper summary: Kai: So we're looking at this paper, "A machine-learning method for the estimation of the background from jets misidentified as hadronic tau lepton decays." The main thing here is that when you look for tau leptons, a big problem comes up: jets from quarks and gluons sometimes get mistaken for these hadronic tau decays.
Mira: Right. It claims they developed something called the MUltivariate Fake-Factor INference approach, which is a machine-learning method to estimate this background directly from the data we observe. They use boosted decision tree reweighting to learn correction factors that depend on multiple variables at once, instead of using simpler low-dimensional parametrizations.
Lev: That’s interesting because learning those multidimensional factors directly from the data should handle complex correlations better than just fitting a simple curve to a few variables one <ref:2610.13634#pg1>.
Kai: Exactly. They show performance studies using proton-proton collision data from the CMS experiment at CERN's LHC, specifically at ten point six TeV with an integrated luminosity of sixty-two point four femtobarns.
Mira: And what they found is that compared to those low-dimensional parametrisation methods, this MUFFIN approach models key kinematic observables much better, especially in regions with high transverse momentum or p T.
Lev: So the practical implication there is a reduction of up to seventy percent in the systematic uncertainty related to these jets being misidentified as hadronic tau leptons one <ref:2610.13634#pg1>.
Kai: It moves away from those simple parametrizations because it replaces them with per-object event weights that capture how all those kinematic and event-level variables are correlated.
Mira: That dependence on strict factorisation assumptions is what they want to minimize by learning these per-object factors directly from the data one <ref:2610.13634#pg1>.
Lev: From an engineering standpoint, that sounds like it should be robust when you run it on real hardware, but we need to see how much overhead that training process actually lands on the computing resources.
Kai: The authors show this works for all analysis channels involving hadronic tau lepton final states, with specific demonstrations in both the fully hadronic tau-h tau-h channel and the semileptonic tau mu plus hadronic channel.
Mira: It’s also useful because they claim it improves modeling across all distributions compared to the established Fake Factor method.
Lev: Can you tell us what makes that improvement at high p T regions so significant for someone actually trying to look for new physics?
Kai: Because at the high end of the tau-h p T spectrum, MUFFIN captures the kinematic dependencies of misidentification rates more accurately, which is crucial for those new physics searches.
Mira: And a specific advantage they point out is that MUFFIN learns the full multidimensional dependence of misidentification probability from data, incorporating variables like tau-h eta and decay mode as continuous inputs into the boosted decision tree.
Lev: That avoids the limitation of the standard Fake Factor method where you can't use observables that weren't directly used when deriving those correction factors one <ref:2610.13634#pg1>.
Kai: So they manage to incorporate those correlated variables as continuous inputs, which allows for a consistent application across different analysis categories and phase space regions.
Mira: The total relative uncertainty in the jet to tau-h background yield in representative inclusive selections is reported as six percent for the tau-h tau-h QCD multijet and five percent for the tau mu plus hadronic W plus jets or tt one <ref:2610.13634#pg1>.
Lev: Six and five percent uncertainties are respectable, but they note that the extrapolation uncertainty is identified as the single largest source of systematic uncertainty in both cases.
Kai: Looking ahead, they mention that training on larger combined data sets would be a natural path to increase statistical precision while keeping it robust against evolving experimental conditions.
Mira: And since the method is independent of specific tau-h identification algorithms, it should stay ready for future algorithm developments and apply to other data-driven background estimation problems involving non-prompt objects beyond tau candidates one <ref:2610.13634#pg1>.
Lev: So, for someone building real hardware, it sounds like the main task moving forward is figuring out how much more training data you need before you can trust that uncertainty budget.
Conclusion: Kai: So we're wrapping up this look at "A machine-learning method for the estimation of the background from jets misidentified as hadronic tau lepton decays."
Mira: It’s about using a machine learning tool, MUFFIN, to estimate that jet background—where quarks and gluons look like tau leptons—directly from observed data.
Lev: The authors are trying to tackle this by replacing old methods with something that learns the misidentification probability across many different kinematic spaces at once.
Kai: They achieved this using boosted decision tree reweighting, which means they’re not fitting simple curves; they’re learning complex factors that depend on all the variables simultaneously.
Mira: The main point is that this approach gives much better modeling of key observables, especially when the transverse momentum of those jets is quite high.
Lev: And that improved modeling leads to a reduction in systematic uncertainty for those backgrounds by up to seventy percent compared to older methods.
Kai: It’s about making sure we get a cleaner picture of the real tau signals we’re looking for in the data at CERN.
Mira: They show this works across all channels, whether it's the fully hadronic tau-h tau-h channel or the semileptonic one with a muon.
Lev: The method avoids relying on those strict factorisation assumptions that plague some of the simpler techniques.
Kai: It’s a robust tool for precision measurements involving tau leptons in searches for things beyond our standard model physics.
Mira: They highlight that it handles variables you might not have used when you first derived the correction factors, like the eta or decay mode of the tau.
Lev: The uncertainty in those background yields they report is around five to six percent, which is pretty solid for a measurement this complex.
Kai: But they point out that extrapolation uncertainty remains their biggest systematic hurdle in both of those main scenarios.
Mira: So, while the machine learning approach looks strong on the data modeling front, the biggest remaining challenge for them is understanding how far that estimate can reliably go when you move outside what they actually trained on.
Lev: That makes sense; you’ve learned a lot from the proton-proton collisions at one energy, but real hardware has its own quirks.
Kai: It shows that machine learning can be a powerful way to handle this kind of messy background estimation when traditional fitting methods just can't keep up with the complexity.
Mira: So we see a new way to handle this specific type of fake background by treating it as a complex pattern recognition problem rather than just a simple counting problem.
Lev: And that opens the door for applying similar approaches to other data-driven problems involving non-prompt objects, not just tau leptons.
European Organization for Nuclear Research (CERN) · CMS Collaboration · Journal of Instrumentation
hep-ex, physics.ins-det
Submitted: 2026-10-09
Updated: 2026-11-16
Comments: Replaced with the published version. Added the journal reference and the DOI. All the figures and tables can be found at http://cms-results.web.cern.ch/cms-results/public-results/publications/TAU-25-001 (CMS Public Pages)
Journal ref: JINST 21 (2026) P11012
DOI: 10.1088/1748-0221/21/11/P11012
License: http://creativecommons.org/licenses/by/4.0/
Importance score: 92/100
The gist: Backgrounds arising from quark- and gluon-initiated jets that are misidentified as hadronically decaying tau leptons present a significant experimental challenge in analyses involving tau leptons.
Key concepts
- MUltivariate Fake-Factor INference (MUFFIN)
- A machine-learning technique that estimates the background from jets misidentified as hadronic tau leptons. It uses boosted decision tree reweighting to learn correction factors directly from observed events, modeling how jet properties relate to tau misidentification across various kinematic regions.
- Boosted Decision Tree Reweighting
- A machine-learning technique used within MUFFIN to create per-object correction factors. It learns multidimensional, object-specific correction factors by training on data. This allows the method to capture complex correlations between multiple kinematic and event variables without needing simple low-dimensional models.
- Fake Factor (FF) Method
- An established method used previously to estimate the background from misidentified jets. MUFFIN is compared against this method, showing that MUFFIN provides improved modeling of key observables, especially at high transverse momentum ($p_T$), leading to a significant reduction in systematic uncertainty.
- Per-Object Event Weights
- MUFFIN replaces traditional low-dimensional parametrizations with per-object weights. These weights capture correlations between multiple kinematic and event variables for each individual object. This approach diminishes the reliance on strict factorisation assumptions and provides a more flexible way to estimate background.
Terminology
Summary
Backgrounds arising from quark- and gluon-initiated jets that are misidentified as hadronically decaying tau leptons present a significant experimental challenge in analyses involving tau leptons. p. 0 EUROPEAN ORGANIZATION FOR NUCLEAR RESEARCH (CERN)
CERN-EP-2026-243
2026/10/12
CMS-TAU-25-001
A machine-learning method for the estimation of the
background from jets misidentified as hadronic tau lepton
decays
The CMS Collaboration*
Abstract
Backgrounds arising from quark- and gluon-initiated jets misidentified as hadronically decaying tau leptons constitute a major experimental challenge for analyses
involving tau leptons. The MUltivariate Fake-Factor INference (MUFFIN) approach,
a machine-learning-based method for estimating this background directly from observed events, is presented. EUROPEAN ORGANIZATION FOR NUCLEAR RESEARCH (CERN)
CERN-EP-2026-243
2026/10/12
CMS-TAU-25-001
A machine-learning method for the estimation of the
background from jets misidentified as hadronic tau lepton
decays
The CMS Collaboration*
Abstract
Backgrounds arising from quark- and gluon-initiated jets misidentified as hadronically decaying tau leptons constitute a major experimental challenge for analyses
involving tau leptons. The MUltivariate Fake-Factor INference (MUFFIN) approach,
a machine-learning-based method for estimating this background directly from observed events, is presented. The MUltivariate Fake-Factor INference (MUFFIN) approach,
a machine-learning-based method for estimating this background directly from observed events, is presented. Boosted decision tree reweighting is used to learn multidimensional, per-object correction factors directly from observed data, thereby modelling the jet to hadronic tau lepton misidentification probability across the kinematic
and topological phase space without relying on low-dimensional parametrisations. The MUltivariate Fake-Factor INference (MUFFIN) approach,
a machine-learning-based method for estimating this background directly from observed events, is presented. Boosted decision tree reweighting is used to learn multidimensional, per-object correction factors directly from observed data, thereby modelling the jet to hadronic tau lepton misidentification probability across the kinematic
and topological phase space without relying on low-dimensional parametrisations.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Performance studies were conducted using proton-proton collision data collected by the CMS experiment at CERN's LHC at sqrt s = 13.6 TeV, corresponding to an integrated luminosity of 62. p. 0 Performance studies are presented using proton-proton collision data collected by the
CMS experiment at the CERN LHC at √
s = 13. 6 TeV, corresponding to an integrated
luminosity of 62. 6 TeV, corresponding to an integrated
luminosity of 62. Compared to low-dimensional parametrisation methods, improved modelling of key kinematic observables is achieved, particularly in regions of
high pT
. Compared to low-dimensional parametrisation methods, improved modelling of key kinematic observables is achieved, particularly in regions of
high pT
. The improved modelling leads to a reduction of up to 70% in the systematic
uncertainty associated with jets being misidentified as hadronically decaying tau leptons.Where the paper says it
Where the paper says it
Where the paper says it
The MUFFIN method fundamentally differs from low-dimensional parametrisation techniques by replacing them with per-object event weights that capture correlations between multiple kinematic and event-level variables, thereby diminishing the dependence on strict factorisation assumptions. p. 3 2
The MUFFIN method replaces explicit low-dimensional parametrisations with per-object event
weights that capture correlations between multiple kinematic and event-level variables, reducing the reliance on factorisation assumptions. At the same time, this novel procedure retains
the fully data-driven nature of the background estimation strategy and provides a more unified
and reproducible background estimation workflow, developed and demonstrated here using
proton-proton (pp) collision data collected with the CMS detector. At the same time, this novel procedure retains
the fully data-driven nature of the background estimation strategy and provides a more unified
and reproducible background estimation workflow, developed and demonstrated here using
proton-proton (pp) collision data collected with the CMS detector. It is applicable to all analysis
channels involving τh final states, and is demonstrated explicitly in the fully hadronic τh
τh
and
in the semileptonic τµ
τh
channel, where τµ denotes a τ lepton decaying to a muon. The improved modelling leads to a reduction of up to 70% in the systematic
uncertainty associated with jets being misidentified as hadronically decaying tau leptons. The MUFFIN method provides a robust and scalable method for background
estimation in precision measurements and searches for beyond-the-standard-model
physics involving tau lepton final states.Where the paper says it
Where the paper says it
Where the paper says it
A notable strength of MUFFIN is its ability to provide improved modeling of all distributions relative to the FF method. p. 0 The improved modelling leads to a reduction of up to 70% in the systematic
uncertainty associated with jets being misidentified as hadronically decaying tau leptons. The MUFFIN method provides a robust and scalable method for background
estimation in precision measurements and searches for beyond-the-standard-model
physics involving tau lepton final states. Compared to low-dimensional parametrisation methods, improved modelling of key kinematic observables is achieved, particularly in regions of
high pT
. The improved modelling leads to a reduction of up to 70% in the systematic
uncertainty associated with jets being misidentified as hadronically decaying tau leptons. The MUltivariate Fake-Factor INference (MUFFIN) approach,
a machine-learning-based method for estimating this background directly from observed events, is presented. Boosted decision tree reweighting is used to learn multidimensional, per-object correction factors directly from observed data, thereby modelling the jet to hadronic tau lepton misidentification probability across the kinematic
and topological phase space without relying on low-dimensional parametrisations. In particular, FF
s are typically parametrised using fixed binning in a limited number of kinematic variables, which constrains the achievable parametrisation and leads to both statistical
limitations from sparsely populated bins and systematic uncertainties associated with missing
variable dependencies. The method assumes that the misidentification probability factorises
between these variables and a separate set of variables used to define the analysis categories,
implicitly neglecting potential correlations with other event-level or object-level quantities. The improved modelling leads to a reduction of up to 70% in the systematic
uncertainty associated with jets being misidentified as hadronically decaying tau leptons. The MUFFIN method provides a robust and scalable method for background
estimation in precision measurements and searches for beyond-the-standard-model
physics involving tau lepton final states. Compared to low-dimensional parametrisation methods, improved modelling of key kinematic observables is achieved, particularly in regions of
high pT
. The improved modelling leads to a reduction of up to 70% in the systematic
uncertainty associated with jets being misidentified as hadronically decaying tau leptons.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Looking forward, further performance enhancements are anticipated with larger datasets. p. 3 Within the solenoid volume are a silicon pixel and strip tracker, a lead tungstate
crystal electromagnetic calorimeter (ECAL), and a brass and scintillator hadron calorimeter
(HCAL), each composed of a barrel and two endcap sections. Forward calorimeters extend
the pseudorapidity (η) coverage of the barrel and endcap detectors. Consequently, improvements in the jet → τh background modelling are expected to have a larger impact on the overall background estimation in the τh
τh
channel, while
still providing benefits in the τµ
τh
channel. Throughout this paper, the FF method is used as
the reference baseline against which the MUFFIN method is compared.Where the paper says it
Where the paper says it
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
-
Multidimensional reweighting for background estimation can be achieved by learning
per-object event weights
directly from observed data using a Boosted Decision Tree (BDT) as described in Section 6.1. This allows the system to capturecorrelations between multiple kinematic and event-level variables, reducing the reliance on factorisation assumptions.
-
The improved AI system can perform background estimation by replacing
explicit low-dimensional parametrisations with per-object event weights that capture correlations between multiple kinematic and event-level variables,
which provides amore unified and reproducible background estimation workflow.
-
The method enables the AI to generalize the Fake Factor (FF) paradigm by learning
multidimensional reweighting functions directly from observed data and simulation,
leading to a reduction ofup to 70% in the systematic uncertainty associated with jets being misidentified as hadronically decaying tau leptons.
-
The system can perform background estimation in both fully hadronic and semileptonic tau channels, as it is demonstrated explicitly in these channels where
the jet → τh background constitutes a significantly larger fraction of the total background in the τh τh channel than in the τµ τh channel.
-
The AI can utilize a continuous multidimensional TF, defined by
wMUFFIN(z) ≡ Ppass(z)/Pfail(z),
which replaces binned FF (x) to modelthe jet → τh misidentification probability across the kinematic and topological phase space without relying on low-dimensional parametrisations.
-
The system can perform performance comparisons by quantifying the agreement between its prediction and the reference FF method using a
mean absolute per-bin Poisson pull, ⟨z⟩,
which serves as a metric for improved modelling of observables liketotal transverse mass [12, 69], mtotal T = qm2T(τ1, τ2) + m2T(τ1,⃗p miss T) + m2T(τ2,⃗p miss T),
andvisible invariant mass mvis.
-
The system can provide a data-driven estimate of residual modelling uncertainty by evaluating the non-closure r(z) ≡ Nobs(z)/Npred(z) in validation regions (VRs), where the
Total uncertainty = q((1 - r) squared + 2r)
combines statistical and observed deviation from unity. -
The AI can reduce systematic uncertainties by learning
full multidimensional dependence of the misidentification probability from data, incorporating η, decay mode, and other correlated observables as continuous inputs to the BDT,
which avoids the limitation of FF method in treating observables not directly used in TF derivation. -
The system can provide improved modelling in high-pT regions critical for new physics searches by capturing
the kinematic dependencies of the misidentification rates
better than low-dimensional parametrisation methods, as demonstrated by improved performance across key observables likepT h T,
mtotal T,
andR(lead h,sublead h).
Abstract
Backgrounds arising from quark- and gluon-initiated jets misidentified as hadronically decaying tau leptons constitute a major experimental challenge for analyses involving tau leptons. The MUltivariate Fake-Factor INference (MUFFIN) approach, a machine-learning-based method for estimating this background directly from observed events, is presented. Boosted decision tree reweighting is used to learn multidimensional, per-object correction factors directly from observed data, thereby modelling the jet to hadronic tau lepton misidentification probability across the kinematic and topological phase space without relying on low-dimensional parametrisations. Performance studies are presented using proton-proton collision data collected by the CMS experiment at the CERN LHC at sqrt s = 13.6 TeV, corresponding to an integrated luminosity of 62.4 fb-1. Compared to low-dimensional parametrisation methods, improved modelling of key kinematic observables is achieved, particularly in regions of high p T. The improved modelling leads to a reduction of up to 70% in the systematic uncertainty associated with jets being misidentified as hadronically decaying tau leptons. The MUFFIN method provides a robust and scalable method for background estimation in precision measurements and searches for beyond-the-standard-model physics involving tau lepton final states.
Sources
- Observation of the Higgs boson decay to a pair of tau leptons with the CMS detector
- Cross-section measurements of the Higgs boson decaying into a pair of tau-leptons in proton-proton collisions at $\sqrt{s}=13$ TeV with the ATLAS detector
- Search for nonresonant Higgs boson pair production in final state with two bottom quarks and two tau leptons in proton-proton collisions at $\sqrt{s}$ = 13 TeV
- Analysis of the CP structure of the Yukawa coupling between the Higgs boson and $\tau$ leptons in proton-proton collisions at $\sqrt{s} =$ 13 TeV
- Measurements of Higgs boson production in the decay channel with a pair of $\tau$ leptons in proton-proton collisions at $\sqrt{s}$ = 13 TeV
- Constraints on anomalous Higgs boson couplings to vector bosons and fermions from the production of Higgs bosons using the $\tau\tau$ final state
- Measurement of the $CP$ properties of Higgs boson interactions with $\tau$-leptons with the ATLAS detector
- Search for the non-resonant production of Higgs boson pairs via gluon fusion and vector-boson fusion in the $b\bar{b}\tau^+\tau^-$ final state in proton-proton collisions at $\sqrt{s} = 13$ TeV with the ATLAS detector
- Differential cross-section measurements of Higgs boson production in the $H\to\tau^+\tau^-$ decay channel in $pp$ collisions at $\sqrt{s}=13$ TeV with the ATLAS detector
- Probing the Higgs boson CP properties in vector-boson fusion production in the $H\rightarrow\tau^+\tau^-$ channel with the ATLAS detector
- Search for heavy neutral resonances decaying to tau lepton pairs in proton-proton collisions at $\sqrt{s}$ = 13 TeV
- Searches for additional Higgs bosons and for vector leptoquarks in $\tau\tau$ final states in proton-proton collisions at $\sqrt{s}$ = 13 TeV
- Search for heavy Higgs bosons decaying into two tau leptons with the ATLAS detector using $pp$ collisions at $\sqrt{s}=13$ TeV
- Search for a third-generation leptoquark coupled to a $\tau$ lepton and a b quark through single, pair, and nonresonant production in proton-proton collisions at $\sqrt{s}$ = 13 TeV
- Search for pair-produced scalar and vector leptoquarks decaying into third-generation quarks and first- or second-generation leptons in pp collisions with the ATLAS detector
- Search for singly and pair-produced leptoquarks coupling to third-generation fermions in proton-proton collisions at $\sqrt{s} =$ 13 TeV
- Search for the pair production of long-lived supersymmetric partners of the tau lepton in proton-proton collisions at $\sqrt{s}$ = 13 TeV
- Search for direct stau production in events with two hadronic $\tau$-leptons in $\sqrt{s} = 13$ TeV $pp$ collisions with the ATLAS detector
- Search for lepton-flavour violation in high-mass dilepton final states using 139 $\mathrm{fb}^{-1}$ of $pp$ collisions at $\sqrt{s}=13$ TeV with the ATLAS detector
- Estimation of backgrounds from jets misidentified as $\tau$-leptons using the Universal Fake Factor method with the ATLAS detector
Related papers
- DrSai: an agentic AI system toward autonomous physics analysis workflows at BESIII
- CANTON- mu Proposal: A Next-Generation Muon g-2 Measurement at Sub-0.1 ppm Precision
- Search for dark photon dark matter near 1,meV using a superconducting mixer
- ANUBIS detector and its sensitivity to neutral long-lived particles
- How to scale your HEP ML models: A recipe for robust architecture comparisons at scale
- Search for the 16 O(ppp) to 13 C pi+ pi+ e+ Decay Mode in Super-Kamiokande Using Machine Learning Techniques