PaperlayerView exact revision on arXiv

Source-linked AI summary

Likelihood-Based Unsupervised Anomaly Detection in CMS Dijet Events

Bhavishya Chebrolu, Hitesh Rasineni, Prajwal Aaryan Immadi

arXiv:2609.06686v1hep-excs.LG

TL;DR

Signal-specific searches may miss unexpected or poorly modelled signatures, motivating complementary model-independent approaches. This study trains a neural spline flow on CMS dijet data and identifies low-likelihood events; selected events show correlated jet-substructure deviations without a narrow dijet-mass resonance. The findings support an exploratory anomaly-search use, not a new-physics claim.

  • Problem

    Signal-hypothesis-specific searches may be blind to unexpected or poorly modelled signatures, motivating complementary model-independent strategies.

    Source: §1 · p. 1

  • Method

    A neural spline flow learns the dominant background from a high-dimensional set of CMS dijet observables, and events with low likelihood are subjected to validation and stress tests.

    Source: §1 · p. 3, §4 · p. 11

  • Results

    At the 99.9th-percentile score threshold, 379 events show correlated deviations particularly in jet girth, N-subjettiness, and energy correlation ratios, while spanning the accessible phase space without a narrow resonance.

    Source: §5.3 · p. 19, §6 · p. 21

  • Takeaways & Limitations

    The study demonstrates a validated, model-independent exploratory use of flow-based density estimation for rare, structured deviations in CMS collision data, without claiming new physics.

    Source: §1 · p. 3, §6 · p. 21

    The method targets broad multivariate departures rather than narrow resonances or single-variable anomalies, and the study lacks signal injections and supervised-classifier comparisons to quantify sensitivity to specific scenarios.

    Source: §6 · p. 23

Abstract

from arXiv · show

We present an unsupervised search for anomalous dijet events in proton--proton collision data using neural spline flow density estimation. A normalizing flow model is trained on a high-dimensional feature space comprising jet, dijet, and event-level observables to learn the dominant Standard Model background directly from data, without assuming a specific signal hypothesis. Events assigned low likelihood under the learned density are identified as potential anomalous events. Using this approach on a CMS Open Data dijet sample, we investigate extreme events in the tail of the anomaly-score distribution and perform an extensive validation and robustness study. This study includes feature-level statistical comparisons, mass decorrelation tests, permutation-based null tests, and evaluations of training stability. The selected anomalous events exhibit notable departures from the background-only expectation, primarily in jet-substructure observables, while remaining stable under several known sources of bias in unsupervised learning. The identified anomalies are distributed across the kinematic phase space and do not exhibit a narrow structure in the dijet invariant-mass spectrum. Instead, they show correlated deviations across multiple observables, consistent with a multivariate difference in jet substructure and event topology rather than a localized resonance. While no claim of new physics is made, this study demonstrates that neural spline flow-based density estimation can be sensitive to rare, structured deviations in collider data and may provide a model-independent exploratory tool for searches for physics beyond the Standard Model at the LHC.

1 Introduction

The paper motivates data-driven anomaly searches as a complement to signal-specific analyses that may miss unexpected signatures. It applies unsupervised neural spline flow density estimation to CMS dijet data and validates low-likelihood events as multivariate deviations rather than a specific new-physics claim.

1 Introduction

  • Specialized signal-hypothesis searches may be blind to unexpected or poorly modelled signatures, motivating complementary model-independent, data-driven strategies.The motivation includes unresolved questions such as dark matter and the origin of the electroweak scale.Source: §1 · p. 1
  • Dijet events combine broad BSM reach with a large QCD background and high-dimensional observables, making them challenging for traditional techniques.The paper frames model-independent anomaly detection as a way to search for unusual event topologies in this setting.Source: §1 · p. 2
  • Neural spline flow density estimation learns the background from CMS Open Data and selects events with anomalously low likelihood for validation.The analysis examines selection stability across model initialization, feature selection, and analysis cuts, alongside kinematic and jet-substructure distributions.Source: §1 · p. 3
  • The study applies a fully unsupervised flow framework to real CMS Run-1 AOD data without Monte Carlo simulation, injected signals, or sideband labels at training time.It uses 61 dijet and event-level observables and tests the selected events with mass-decorrelation, permutation-based null tests, and stability checks.Source: §1 · p. 3
  • The selected events show correlated deviations across observables, including multi-prong and non-QCD-like radiation patterns, rather than a specific new-physics interpretation.The paper presents the work as methodological and exploratory, not as evidence for new physics.Source: §1 · p. 3

2 Literature Review

The review places unsupervised density estimation among complementary collider anomaly-detection strategies, then relates the paper’s approach to sideband, likelihood-ratio, and flow-matching searches. It also summarizes validation lessons, jet-definition and substructure context, and benchmark experience that inform the study’s exploratory use of real collision data.

2 Literature Review

  • Anomaly detection aims to learn the Standard Model background from data and identify deviations without committing to a particular BSM model.The review contrasts this with autoencoders, which flag high reconstruction error, and weakly supervised methods that use sidebands or auxiliary labels.Source: §2 · pp. 3–4

2.1 Density-based sideband methods: CATHODE

  • CATHODE [10] trains a conditional density estimator on sidebands, samples synthetic background in the signal region, then classifies real versus synthetic events.The present analysis instead scores real events directly under one learned density, without a resonant signal-region assumption or a subsequent classifier.Source: §2.1 · p. 4

2.2 NSF likelihood-ratio searches on CMS Open Data

  • The mono-Z search [35] uses a learned likelihood-ratio score and profile-likelihood interpretation, whereas this paper uses a single background-only flow without simulated signal.The review presents the analyses as complementary: the earlier search is hypothesis-driven, while this work is fully unsupervised.Source: §2.2 · p. 5

2.3 Flow-matching background density models in the hadronic mono-Z channel

  • The hadronic mono-Z study [36] uses a conditional flow-matching continuous normalizing flow and a background negative-log-likelihood discriminant, reporting expected significances of 2.89σ, 7.62σ, and 7.41σ for three benchmarks.Removing six extra-jet kinematic features degrades expected significance by 53–71%.Source: §2.3 · p. 5

2.4 Asymptotic significance formulae and the Asimov data set

  • Asymptotic profile-likelihood formulae support calibrated inference when a signal hypothesis exists; this analysis instead uses resampling-based diagnostics because it has no signal-strength estimator.The review names Kolmogorov–Smirnov tests, HT-conditional permutation nulls, and bump-hunt pseudo-experiments as the present study’s diagnostics.Source: §2.4 · p. 6

2.5 Jet clustering and the anti-kt algorithm

  • CMS Run-1 jets use anti-kt clustering, whose rigid boundary gives the jet kinematics and substructure features a fixed geometric footprint.The review connects that stable catchment area to interpreting feature-level deviations as event properties rather than artifacts of a changing jet definition.Source: §2.5 · p. 7

2.6 Energy correlation functions and jet-substructure features

  • Energy correlation functions use particle energies and pairwise angles; their (N+1)-point correlator is sensitive to N-prong substructure without explicit subjet finding.The review notes that ECFs and N-subjettiness probe substructure through complementary mechanisms, and that their correlated deviations support an internal-topology interpretation.Source: §2.6 · pp. 7–8

2.7 The LHC Olympics benchmark for data-driven anomaly detection

  • The LHC Olympics [11] emphasized background control, diverse signal scenarios, and validation; this study applies those lessons to unblinded CMS data without ground-truth anomalies.The review also notes that concrete physical interpretation and control of systematic uncertainties remain challenges, motivating conservative interpretation and cross-checks.Source: §2.7 · pp. 8–9

3 Dataset and Feature Representation

The analysis uses a reduced CMS Run2011B high-HT AOD sample and extracts a broad set of reconstructed observables for unsupervised anomaly detection. Its feature representation combines jet and dijet kinematics with composition, substructure, event-shape, and event-context information.

3.1 Data Samples

  • The study uses CMS Run2011B HT AOD proton–proton data collected at √s = 7 TeV, recorded with high-HT triggers for inclusive jet and dijet measurements.The sample is predominantly Standard Model multijet events and uses no Monte Carlo simulation during training or evaluation.Source: §3.1 · p. 9
  • Because the full dataset exceeds 10 TB, the analysis selects 30 ROOT files from larger index groups, yielding a 125 GB working sample.The paper reports qualitative checks of key global observables across file groups and subsamples, with no significant distortions or systematic drifts within their statistical precision.Source: §3.1 · p. 9

3.2 Feature Categories and Physics Motivation

  • The analysis extracts 61 high-level observables from reconstructed CMS AOD data to describe jet kinematics, internal structure, and event topology.The representation is designed to be broadly model-independent and complementary.Source: §3.2 · pp. 10–11
  • The inputs include leading-jet kinematics, global event activity, dijet correlations, jet composition and identification, and substructure variables such as N-subjettiness and jet width.The substructure observables target multi-prong structure and complex radiation without assuming a specific physics model.Source: §3.2 · p. 10
  • Event-shape variables capture global energy-flow structure, while missing-energy and angular-correlation features account for transverse imbalance and possible unusual event conditions.Primary-vertex, pileup-energy-density, and trigger metadata provide context and help distinguish trigger-level phenomena from event-selection effects.Source: §3.2 · p. 10

4 Methodology

The analysis learns a data-driven density for CMS dijet events with a neural spline flow, then ranks low-likelihood events and tests the selection for robustness. Its workflow uses minimal validity cuts, 61 jointly modeled observables, and likelihood-based scoring without a signal hypothesis.

4.1 Data Cleaning and Event Validity Selection

  • Validity selection removes pathological or incomplete events using basic kinematic, reconstruction, and data-quality requirements, without cuts targeting jet substructure or anomaly-related observables.The final sample contains 378,278 events; the reduction from 452,055 events surviving looser kinematic selection reflects multiple cuts, not any single cut.Source: §4.1 · pp. 11–12

4.2 Feature Representation and Preprocessing

  • The input comprises 61 reconstructed jet, dijet, event, and substructure observables, retained together so the flow can learn their joint distribution and correlations.Periodic angles are sine-cosine encoded and continuous features standardized; no further feature selection or decorrelation is applied.Source: §4.2 · pp. 12–13

4.3 Neural Spline Flow Architecture and Training

  • The flow uses ten autoregressive rational-quadratic spline blocks to model the joint event density and compute exact likelihoods, trained by maximizing average log-likelihood.Training and scoring use the same cleaned sample, so absolute likelihoods have in-sample bias; interpretation relies on event ranking and explicit null and stability tests.Source: §4.3 · pp. 13–14

4.4 Likelihood-Based Anomaly Scoring

  • The anomaly score is the negative log-likelihood in standardized input space; the 99.9th percentile defines reference candidates for feature-level studies, with ranking favored over an absolute threshold.The latent-space score gives a consistent ranking (Spearman correlation > 0.99); validation includes mass-decorrelation, feature-level, permutation-null, and training-stability checks.Source: §4 · p. 11, §4.4 · pp. 14–15

5 Results

The flow captures the dense background region and selects rare events with coherent jet-substructure deviations. Mass-decorrelation, permutation, and training checks probe whether these patterns are robust.

5.1 Anomaly Score Distributions

  • The score distribution clusters at low values and rolls off smoothly into the tail; 379 events, about 0.1% of the sample, meet the 99.9th-percentile selection.Percentile ranking avoids relying on an absolute score threshold sensitive to likelihood normalization.Source: §5.1 · p. 16, §5.2 · p. 17

5.2 Anomaly Selection and Feature-Level Statistical Deviations

  • Because selection and feature tests use the same inputs, KS p-values are not calibrated for discovery claims; the tests characterize which features differ and whether deviations cohere.The interpretation rests on coherent differences across correlated features, not any single feature's significance.Source: §5.2 · p. 17
  • Zero values of τ21 or τ32 can reflect low-multiplicity reconstruction, while jet-flag, pileup, and missing-energy variables show no notable deviations.The fixed-value populations are reconstruction artifacts included in the flow's learned joint distribution; the authors say they do not affect anomaly selection.Source: §5.2 · pp. 17–18
  • D = 0.575 for leading-jet girth, the largest reported feature separation; anomalous events also differ across correlated observables, indicating a multivariate displacement.The energy-correlation ratio has D = 0.508, and the reported deviations are diagnostics rather than discovery-level significance.Source: §5.2 · pp. 17–18, Figure 4 · p. 19

5.3 Kinematic Distributions and Resonance Search

  • 8.79 was the largest sliding-window statistic (global p-value 6.0 × 10−4), but no window supports discovery-level evidence for a narrow resonance.The anomalous events span the accessible mass range, consistent with deviations across multiple observables rather than one enhanced kinematic variable.Source: §5.3 · p. 19, Figure 5 · p. 20

5.4.1 Mass Decorrelation Tests

  • 72% of top-0.1% events overlap between the main and mass-decorrelated flows; common events retain substructure deviations, arguing against mass sculpting as the main source.The mass-decorrelated flow is a targeted cross-check, not the primary analysis, because removing mass features discards potentially useful information.Source: §5.4.1 · pp. 20–21

5.4.2 Permutation Tests

  • HT-conditional score permutations yield uniformly small feature-level KS statistics, unlike the observed deviations, supporting a genuine score–kinematics correlation rather than random fluctuations.Conditioning on HT preserves its marginal distribution while disrupting the score correlations with other kinematic variables.Source: §5.4.2 · p. 21

5.4.3 Training Stability

  • Across random-seed training runs, event ordering changes only slightly and feature-level KS statistics remain stable.The authors note that varying only the random seed provides a limited exploration of model space.Source: §5.4.3 · p. 21

6 Summary and Outlook

The study uses unsupervised neural spline flows to identify low-likelihood CMS dijet events and finds correlated departures from background expectations, especially in jet substructure. Validation supports the stability of the selection, but the deviations remain exploratory and are not evidence for a specific new-physics interpretation.

6 Summary and Outlook

  • The analysis uses a neural spline flow to learn the background density from CMS Run-1 data and score events without signal samples or new-physics assumptions.The model uses 61 reconstructed jet, event-shape, and global observables.Source: §6 · p. 21
  • At the 99.9th percentile, 379 events show departures from background expectations, especially in jet girth, N-subjettiness, and energy correlation ratios.The deviations are correlated across multiple observables, placing the events in a different region of feature space.Source: §6 · p. 21
  • Validation finds the deviations stable against mass decorrelation, permutation tests, training variation, and several detector-related effects, but cannot exclude every detector contribution.The events lack narrow kinematic structures and instead show multivariate differences in event topology and jet substructure.Source: §6 · p. 21, Figure 6 · p. 22, Figure 7 · p. 23
  • The deviations are exploratory and are not interpreted as evidence for new physics.The work is a proof of concept for sensitivity to rare, structured features in collider data and for systematic validation in unsupervised analyses.Source: §6 · pp. 22–23
  • The method is naturally sensitive to broad, multivariate departures, but is not optimized for narrow resonances or anomalies confined to a few kinematic observables.The study does not include signal injections or comparisons with supervised baselines to quantify sensitivity to concrete scenarios.Source: §6 · p. 23
  • The authors propose applying the analysis to more CMS data and higher-energy runs to evaluate the statistical significance of the observed deviations.They also identify particle-level or graph inputs and complementary anomaly-detection methods as future directions.Source: §6 · p. 23

Declarations

The appendices document the dataset and input-observable inventory, then describe stability checks and feature-level comparisons. Extreme anomaly scores remain stable across training and sampling variations, while tests identify differences in kinematic and substructure features but not pileup or vertex observables.

Declarations

  • The appendix states that no funding was received, CMS Open Data are publicly available, and the neural spline flow training and scoring code is public.The code for the initial AOD-to-ROOT conversion is available upon request.Source: Declarations · p. 24

A Complete List of Input Observables

A.1 Jet Kinematics

  • The jet-kinematics features describe the global momentum scale and angular configuration of the hardest jets, while jet-ID flags use the official CMS Run-1 particle-flow identification algorithm.Jet-substructure observables examine internal energy flow and are sensitive to multi-prong patterns.Source: §A.1 · p. 25, §A.2 · p. 25, §A.3 · p. 25

A Complete List of Input Observables

  • The observable inventory also includes global event shapes, missing-energy and angular correlations, vertex and pileup quantities, jet-flavor tags, and trigger metadata.The listed tables cover these additional feature groups.Source: Table 6 · p. 26, Table 7 · p. 26, Table 8 · p. 27

B Stability of Anomaly Score Distributions

  • Across random seeds, independent subsamples, and data-taking periods, the anomaly-score distribution retains its shape and extreme-event ordering, apart from tiny adjustments.The score is the negative log-likelihood under the trained neural spline flow.Source: B Stability of Anomaly Score… · p. 27

B.1 Kolmogorov–Smirnov Tests on Input Features

  • Feature-level KS tests find marked inconsistencies in kinematic and substructure observables, but not in pileup or vertex observables.The comparisons use the 99.9th-percentile anomalies against the inclusive sample; the empirical CDFs define the KS comparison.Source: §B.1 · pp. 27–28

B.2 Multiple-Testing Considerations

The appendix treats the KS tests as exploratory diagnostics because the 61 features are numerous and correlated, without formal multiple-testing corrections. Control-region and detector checks find no localized sideband excess or unusual detector-level features that explain the selected anomalies.

B.2 Multiple-Testing Considerations

  • No formal multiple-testing correction is applied to the exploratory KS p-values because many features are correlated and the goal is to identify coherent patterns, not establish discovery-level claims for individual observables.The authors interpret correlated deviations across physically connected observables rather than isolated p-values.Source: §B.2 · p. 28
  • The observed deviations cluster in correlated feature groups rather than appearing randomly across uncorrelated observables, so individual KS p-values are treated as diagnostic indicators.The authors also report consistency across kinematics, substructure, and event shapes, with deviations persisting under mild decorrelation.Source: §B.2 · p. 28

B.3 Control Region and Sideband Checks

  • Across mass and angular sidebands, anomaly scores show no localized excess, feature-level KS statistics are uniformly small, and selected events remain distributed across the accessible kinematic range.Random samples matched in jet multiplicity and similar HT also do not show the feature-level deviations found in the analysis.Source: §B.3 · p. 28

B.4 Detector and Reconstruction Sanity Checks

  • The selected unusual events pass standard CMS Run-1 quality assessment and show no unusual detector-level features, including in jet identifiers, energy fractions, vertex and pileup flags, or trigger flags.These checks suggest detector noise or reconstruction artifacts do not explain the effects.Source: §B.4 · pp. 28–29

B.5 Summary of Validation Studies

The validation summary reports stable anomaly scores and physically organized feature-level deviations, while noting that the tests do not establish a physical mechanism. The appendix identifies the CMS 2011 Run2011B HT AOD data source and points to the reference CMS dijet analysis motivating several observables.

B.5 Summary of Validation Studies

  • The supplementary tests report stable anomaly scores, physically organized feature deviations, robustness to preprocessing and training, and no adequate detector or pileup explanation.Mass-decorrelation and random-permutation tests are also summarized as evidence that the deviations are not statistical artifacts.Source: §B.5 · p. 29
  • The validation supports statistical deviation from dominant background distributions but stops short of attributing a physical mechanism.The deviations are not presented as a definitive physical interpretation.Source: §B.5 · p. 29

C Dataset Provenance and File Selection

  • The dataset comes from CMS proton–proton collisions recorded during the 2011 LHC run, with primary input files from the Run2011B HT AOD tier.The appendix describes the dataset provenance and the choice to use a subset rather than the full available collection.Source: §C.1 · p. 29, C Dataset Provenance and File Selection · p. 29

C.1 CERN Open Data Source

  • The CERN Open Data release includes CMS AOD data with particle-flow jets, missing transverse energy, vertices, trigger data, and related reconstructed objects.The release is organized into file groups and run blocks and exceeds 10 TB.Source: §C.1 · p. 29

C.2 Reference CMS Publication

  • The cited CMS dijet reference analysis motivates the dataset and the use of dijet-level, event-shape, and substructure observables in this work.The reference is the CMS measurement of dijet angular distributions and search for quark compositeness at √s = 7 TeV.Source: §C.2 · p. 29

C.3 Partial Dataset Usage

The analysis uses a representative 125 GB fraction of the 10 TB dataset, derives 61 physics-relevant observables, and documents file selection and feature extraction for reproducibility.

C.3 Partial Dataset Usage

  • About 125 GB of the 10 TB dataset was analyzed, using all files from groups 1 and 2 and uniform ROOT-file subsampling from groups 3–5.Ten files were selected from each of groups 3–5; all selected-file events passing the common quality and kinematic selections were retained.Source: §C.3 · p. 30

C.4 File Selection Criteria

  • Files from groups 3–5 were selected for run and luminosity-section coverage and an even distribution of file indices to avoid temporal bias.The selected files and download links are listed in CERN Open Data JSON indexes and the analysis repository.Source: §C.4 · p. 30

C.5 Event Reconstruction and Branch Selection

  • The analysis remapped reconstructed CMS objects into 61 observables spanning jets, substructure, dijets, event shapes, missing energy, vertices, pileup, flavors, and triggers.The observables use publicly available descriptions of CMS Run-1 algorithms and definitions.Source: §C.5 · p. 30

C.6 Reproducibility Considerations

  • The feature-extraction macros enforce consistent event ordering, object ordering, and event selections, and avoid simulation reweighting and detector-level calibrations.The provided files, macros, and pipelines are sufficient to reproduce the processed collection from the original CERN Open Data samples.Source: §C.6 · pp. 30–31

C.7 Scope and Limitations

The study uses a subset that is small relative to the full Run-1 dataset, so it does not exclude other anomalous structures in the full data.

C.7 Scope and Limitations

  • The analyzed subset is small relative to the full Run-1 dataset, although large for standard unsupervised anomaly-detection tasks.The authors identify a larger data fraction or distributed computing as possible routes toward analyzing the full dataset.Source: §C.7 · p. 31
  • Other anomalous structures in the full dataset are not fully excluded by this subset analysis.The authors suggest analyzing a larger fraction of available data or using distributed computing in future work.Source: §C.7 · p. 31
  • A larger data fraction or distributed computing could be used in future work to analyze the full dataset.The current subset is small relative to the full Run-1 dataset.Source: §C.7 · p. 31

D Stability and Robustness of the Anomaly Selection

The appendix evaluates whether anomaly-selection findings depend on stochastic training or on how anomalous events are defined.

D Stability and Robustness of the Anomaly Selection

  • The robustness study addresses whether detected anomalies depend on random initialization, numerical correlation, or highly specified thresholds.The analysis is unsupervised and ranks events by likelihood rather than labeling them.Source: D Stability and Robustness of the… · p. 31
  • One study tests the stability of anomaly selection under randomized initialization of the neural spline flow parameters.The appendix identifies randomized initialization as one of two stability investigations.Source: D Stability and Robustness of the… · p. 31
  • A second study tests whether feature deviations remain stable as the anomaly-score threshold changes.Together, the two studies target training stochasticity and the definition of anomalous events.Source: D Stability and Robustness of the… · p. 31

D.1 Stability under Random Initialization

Repeated training and additional analysis variations yield substantial overlap among selected extreme events, supporting stability of the anomaly selection across the tested choices.

D.1 Stability under Random Initialization

  • The analysis compares anomaly scores for the same event sample across models using rank correlation, top-event overlap, and score-distribution distance.The training procedure is replicated with different random seeds while the architecture, hyperparameters, preprocessing, and dataset remain fixed.Source: §D.1 · p. 31
  • Over 88% of the top 0.1% anomalous events overlap across random initializations, rank correlations exceed 0.95, and score-distribution KS distances are small.The study holds architecture, hyperparameters, preprocessing, and dataset fixed while changing random seeds; Table 11 reports the overlap, Spearman correlation, and KS distance measures.Source: Table 11 · p. 31, §D.1 · p. 31
  • Top-0.1% event overlap remains above 70% when varying flow blocks, hidden units, training/scoring splits, and angular encoding.The tested variations include 8–12 flow blocks, 128–512 hidden units, training on 80% and scoring on 20%, and preprocessing with or without angular encoding.Source: §D.1 · p. 31

D.2 Stability under Anomaly-Score Threshold Variation

Mean feature-level KS separation increases smoothly as the anomaly-score threshold tightens across the tested percentiles, with no sharp transitions. The observed feature differences are therefore not uniquely tied to one threshold choice.

D.2 Stability under Anomaly-Score Threshold Variation

  • The outlier definition remains partly subjective because it relies on a percentile-based anomaly-score threshold.Source: §D.2 · p. 32
  • The stability check repeats feature-level KS tests between anomalies and background at four tail percentiles, calculating separations independently for each feature and their average.Source: §D.2 · p. 32
  • Mean KS separation rises gradually across the 99.5%, 99.7%, 99.9%, and 99.95% thresholds, with no sharp transitions or discontinuities.Figure 8 shows the smooth, monotonic trend over the four thresholds examined.Source: Figure 8 · p. 32, §D.2 · p. 32
  • Feature-level differences are not uniquely tied to a specific anomaly-score cut and vary smoothly across a broad, physically motivated range of anomaly definitions.The authors describe this stability as validating the findings from the analyzed anomalous-event population.Source: §D.2 · pp. 32–33
  • The appendix concludes that anomaly detection is robust to training stochasticity and analysis-level cuts, a property described as crucial when no external signal model informs selection.Source: §D.2 · p. 33

Related papers

Loading 2609.06686v1…