"Transforming" LHCb: self-supervised maps of heavy-flavour decays
Marko Stamenkovic, Greg Landsberg
hep-ex · cs.LG
Quick verdict
Self-supervised transformer learns LHCb heavy-flavour structure from public data; credible proof of principle, but sideband overlap weakens anomaly claims.
Summary
The authors train a ~1M-parameter set transformer on simulated LHCb Open Data dijets (735,348 events, anti-kT R=0.5, secondary-vertex-tagged jets) to learn heavy-flavour decay environments without labels: PID inputs are masked and up to four constituents deleted, and the model predicts the missing multiplicity, kinematics, charge, identity, and topology. Deletion detection yields an inclusive paired AUC of 0.72 (0.87 for four removed particles). Linear probes on the frozen embeddings beat a random-weight encoder on five flavour/charge tasks, recover 97-100% of fully supervised AUC, and reach opposite-jet tagging power of 10.6% (b) and 31.1% (c) — which the authors explicitly caution is not comparable to LHCb's exclusive-decay taggers. Removing reconstructed D0 candidates from b-jets raises the missingness score more than matched-track removal (paired excess 2.49 [2.17, 2.83]); an independently trained model on 2017 collision data (1.03 fb-1) reproduces the response for four high-statistics charm modes exceeding both sidebands, and simulation-trained probes separate charge in real data (Ptag 54.5% for B+ to J/psi K+). Main limitations: signal and sideband distributions overlap substantially, beauty-mode statistics are tiny (28-66 candidates), the anomaly section is exploratory, and no code release is stated.
Review complete
Assessment at a glance
4
risk flags identified
● high 0
● medium 2
● low 2
5 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- A self-supervised transformer trained without flavour or exclusive-decay labels learns LHCb heavy-flavour decay structure, beating a random-weight encoder on all five flavour/charge tasks and recovering 97-100% of fully supervised AUC (Sec. 4.2).
- Opposite-jet tagging power reaches 10.6% (b) and 31.1% (c) in simulated inclusive samples, conditional on jet selection and explicitly not comparable to LHCb's published exclusive-decay taggers (Sec. 4.2.1).
- Removing reconstructed decay candidates produces systematically larger missingness responses than removing matched tracks, e.g. a paired excess of 2.49 [2.17, 2.83] for D0 in b-jets (Sec. 4.3.2).
- A model trained only on 2017 collision Open Data responds to eight reconstructed charm/beauty modes, with all four high-statistics charm modes exceeding both adjacent sidebands (Sec. 5.3).
- Simulation-trained affine charge probes transfer unchanged to collision data, separating charge in B+ and prompt D decays (Sec. 5.4).
Strengths
Exemplary honesty on the tagger comparison
Section 4.2.1 explicitly states the 10.6% opposite-jet tagging power is conditional on a pre-selected sample (100% tagging efficiency, charged-hadron content, no data calibration) and 'should not be interpreted as an improvement' over the LHCb inclusive tagger (5.4-7.8%), preventing the abstract's ~10% headline from being over-read.
Well-controlled intervention design
Reconstructed-decay removals (Sec. 4.3) are tested against mass sidebands and matched-track controls matched in multiplicity, charge, momentum, angular spread and displacement; the D0 test reports a paired within-jet excess of 2.49 [2.17, 2.83], positive for 88.5% of pairs, with AUC 0.83.
Genuine validation on real collision data
Proper ablations for the representation claim
Disciplined treatment of the anomaly survey
Weaknesses
Signal-vs-sideband separation is modest and untested statistically
Completion heads are weak in absolute terms
Beauty-mode evidence in data is thin and its explanation post hoc
Single pretraining runs, no seed variance, no code statement
Data tagging powers rest on approximate subtraction
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
This JHEP-format proof of principle (arXiv:2609.09275, Brown University) trains a ~1M-parameter set transformer (five blocks, d=96, 8 heads, pairwise-relation attention biases, learned completeness token) to learn heavy-flavour decay environments from LHCb jets without flavour or exclusive-decay labels. Pretraining corrupts each jet two ways: 0-4 constituents deleted under six mechanistic types (Table 4, 25% of views unchanged) and PID masked on 15% of survivors; heads reconstruct missing multiplicity, per-particle kinematics/charge/PID, set summaries, topology, corruption type, and masked PID (Eq. 3.1). Claims: (i) the representation encodes flavour/charge — affine probes beat a random-weight encoder on five tasks and recover 97-100% of fully supervised AUC (Sec. 4.2); (ii) opposite-jet tagging power of 10.6% (b) and 31.1% (c) in simulated inclusive samples (Sec. 4.2.1); (iii) removing reconstructed decay candidates produces larger missingness responses than matched-track removal, validated in simulation (D0 paired excess 2.49 [2.17, 2.83]) and with an independently trained model on 1.03 fb-1 of 2017 collision Open Data (Sec. 5).
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 4 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What does this paper actually do?
It trains a self-supervised transformer (~1M parameters) on LHCb jet data that learns what a complete heavy-flavour decay environment looks like, without any flavour labels: it masks particle-identification inputs and deletes up to four particles, then predicts what is missing. The learned representation supports flavour tagging and flags jets that appear to be missing particles, a probe for decays with invisible or unreconstructed components.
Is the ~10% tagging power better than LHCb's existing flavour tagger?
No direct comparison is valid. The 10.6% (b-jet) figure is measured on pre-selected simulated dijets where every jet is tagged, includes charged B hadrons that are easier to tag, and is not calibrated on real data. The published LHCb inclusive tagger reaches 5.4-7.8% under different conditions, and the authors explicitly state their result should not be read as an improvement.
Was real LHCb collision data used, or only simulation?
Both. Simulation (CERN Open Data record 4910) provides controlled tests, and a second model was trained from scratch only on 2017 proton-proton collision Open Data (record 94000, 1.03 fb-1). Simulation-trained charge probes were also applied unchanged to reconstructed B and D decays in that collision data.
Is the code available?
The paper uses public CERN Open Data and documents its inputs and hyperparameters in detail, but the provided text contains no code-release statement, so the pipeline would currently need to be reimplemented.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Numeric values that appear only inside figure panels (e.g., per-sideband mean scores and uncertainty bars in Fig. 8, exact probe points in Figs. 2-3) are not recoverable from the text extraction; only values stated in the text or captions were verified.
- The extraction runs cleanly through reference [34] on page 30 with no evidence of truncation of the body, appendices, or bibliography.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.