Likelihood-Based Unsupervised Anomaly Detection in CMS Dijet Events
Bhavishya Chebrolu, Hitesh Rasineni, Prajwal Aaryan Immadi
hep-ex · cs.LG
Quick verdict
Unsupervised neural-spline-flow scan of 378k CMS dijet events finds substructure-driven outliers; validation credible but in-sample; no new physics claimed.
Summary
This paper trains a neural spline flow density estimator on 61 reconstructed observables from 378,278 dijet events in the CMS Run2011B HT open dataset (7 TeV, ~0.5 fb−1) to learn the QCD multijet background without simulation, labels, or signal hypotheses. Events are ranked by negative log-likelihood; the top 0.1% (379 events) shows coherent deviations concentrated in jet-substructure observables (leading-jet girth KS D=0.575, N-subjettiness τ1–τ3, ECF ratio D=0.508), while detector, vertex, pileup, and trigger observables remain flat. No narrow resonance appears in the dijet mass spectrum; a global bump-hunt reports a borderline look-elsewhere-corrected p-value of 6.0×10−4. Validation includes mass-decorrelated retraining (72% overlap of selections), HT-conditioned permutation nulls, multi-seed stability (88–93% overlap, Spearman ρ>0.95), and threshold/architecture variations. The authors acknowledge in-sample training-scoring, the absence of signal-injection sensitivity studies, and make no new-physics claim, positioning the work as a proof of concept for model-independent exploration on real collider data.
Review complete
Assessment at a glance
5
risk flags identified
● high 0
● medium 2
● low 3
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 neural spline flow trained on 61 reconstructed observables from 378,278 real CMS Run2011B dijet events learns the QCD multijet background with no simulation, labels, or signal hypotheses.
- The top 0.1% of events by negative log-likelihood (379 events) deviate coherently and specifically in jet-substructure observables (girth D=0.575, N-subjettiness, ECF ratio D=0.508) rather than in any single kinematic variable.
- The deviations are not mass-sculpted: a mass-decorrelated flow yields 72% overlap of selections and no mjj excess, and the mjj spectrum shows no narrow resonance (borderline global bump-hunt p=6.0×10−4).
- The anomaly selection is stable under random seeds (88-93% overlap, ρ>0.95), architecture changes, threshold variation, and shows no correlation with detector, vertex, pileup, or trigger observables.
- No new physics is claimed; the work is a proof of concept that flow-based density estimation can flag rare, structured, multivariate deviations in real collider data.
Strengths
End-to-end demonstration on real collision data
A fully unsupervised pipeline (Sections 3-4) runs on 378,278 events from CMS Run2011B HT AOD open data (record 274, ~0.5 fb−1) with no Monte Carlo, labels, or sidebands at any stage — a setting most flow-based anomaly-detection papers only exercise on simulation or LHC Olympics benchmarks.
Validation breadth beyond the genre norm
Mass-decorrelated retraining with a quantified 72% top-0.1% overlap (Section 5.4.1), HT-conditional permutation nulls (5.4.2), multi-seed stability of 88.1-92.9% overlap with Spearman ρ≥0.955 (Table 11), architecture/threshold stress tests (Appendices D.1-D.2), sideband and detector sanity checks (B.3-B.4).
Statistical honesty about the score and the KS tests
Physically interpretable anomaly characterization
Reproducibility provisions
Weaknesses
In-sample training and scoring
Bump-hunt null underspecified
No sensitivity calibration
Unquantified reconstruction-artifact channel
Quantitative gaps in decorrelation claims
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
The paper applies a fully unsupervised, likelihood-based anomaly search to real collider data: a rational-quadratic neural spline flow (RQ-NSF) is trained on 61 reconstructed observables from 378,278 dijet events of the CMS Run2011B HT AOD open dataset (√s = 7 TeV, ~0.5 fb⁻¹ via file-level subsampling of the >10 TB record 274; Section 3.1, Appendix C). The anomaly score is −log p(x) (Eq. 3); the top 0.1% (99.9th percentile, 379 events) is the anomaly population. Central claims: (i) the flow learns the SM multijet background without simulation, labels, or sidebands; (ii) the tail population deviates coherently in jet-substructure observables (Section 5.2, Table 1); (iii) the deviations are not mass sculpting, detector artifacts, or training noise (Sections 5.3-5.4, Appendices B and D); (iv) explicitly, no new-physics claim is made (Sections 1, 6). The scope is methodological, and the authors are consistently candid about limits (in-sample scoring, no signal injections, KS p-values not discovery-calibrated) — an honesty that strengthens the manuscript.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 5 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
Did this paper discover new physics at the LHC?
No. The paper explicitly makes no new-physics claim. It flags 379 events (top 0.1%) that are statistically unusual under a learned density of CMS dijet data, concentrated in jet-substructure observables; the dijet mass spectrum shows no narrow resonance, and the global bump-hunt p-value of 6.0×10−4 is reported as borderline, not as a discovery.
What data does the analysis use?
Public CMS Run2011B HT AOD data from the CERN Open Data portal (record 274), 7 TeV proton-proton collisions recorded in 2011. The authors use a 125 GB file-level subset (~0.5 fb−1 of integrated luminosity), yielding 378,278 events after pT>30 GeV, mjj>200 GeV, at-least-two-jets, finiteness, and validated-run cuts.
How is an event defined as anomalous?
A neural spline flow (a normalizing-flow density estimator) learns the joint probability density of 61 standardized event observables; the anomaly score is the event's negative log-likelihood. Events in the top 0.1% of scores (99.9th percentile, 379 events) are labeled anomalous candidates — a percentile-based, threshold-free definition.
Is the code and the anomaly event list available?
The Declarations state that the neural spline flow training and scoring code is publicly available and that the list of anomalous events (run number, lumi-section, event number) is provided as supplementary material; however, no repository URL or DOI is given in the manuscript text.
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.
- Figure contents (Figs. 1-8) are known only through their captions; visual features of the distributions could not be assessed from the text extraction.
- The metadata abstract is cut mid-sentence, but the complete abstract is present at the start of the extracted full text, so no content was lost.
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.