Many searches for invisible particles at colliders, such as dark matter or neutrinos from exotic decays, come down to one thing: finding events where the visible particles do not balance in the plane transverse to the beam. The imbalance is called missing transverse energy (MET). A neural network then separates the rare signal events from ordinary "background" events. The weak spot is that detectors also produce fake imbalance. Mismeasured jets, dead channels and extra soft activity from overlapping collisions all do this. Fake imbalance mostly lives in the tails of the MET resolution, and those tails are exactly where the purest signal selections sit. When running conditions change, the tails change shape: the spread and the correlations of the mismeasurement shift, not just its average. The standard response is to correct energy scales, or to adapt the network so that simulation and data look alike. Neither targets the width of the distribution directly.
We propose to transfer a lesson from classical speech recognition to this problem. That field spent a decade establishing that adapting only the means of a model fails whenever noise stretches or flattens the feature distribution, and that variance and covariance must be adapted too, provided overfitting is controlled when data are scarce. Our method has three steps. First, estimate a small, physically structured detector-response correction (per-object energy scale, a resolution covariance, and a tail parameter) from a few thousand background-dominated control events in the new running condition. Second, apply this correction to individual reconstructed objects and recompute MET as their negative vector sum, so that corrected events still conserve momentum. Third, propagate the fitted correction onto labelled simulation and update the frozen classifier through a small low-rank adapter. The claim we will test is specific: the correction must reach the covariance and tail level, not just the mean, to remove background mis-estimation in high-purity selections. The question of which classifier weights to adapt is secondary.
Core Claim
Under a shift in MET resolution, three things should happen, and none is visible in the area under the ROC curve. Mean-only correction should leave a background-yield bias that grows as the selection gets purer. Covariance-plus-tail correction, fit through a momentum-conserving forward model, should remove most of that bias with on the order of $10^3$ control events. Adapting the network directly on control data should either fail, because the control region has no signal labels, or erase discriminating information. The new element is not "put adapters on a physics classifier." It is to treat the scope of the detector-response correction (mean, covariance, tail) as the object of adaptation, and to fit it at the level of physics objects so that four-momentum balance holds by construction.
Where This Sits
The motivating evidence comes from acoustic modeling. Watanabe et al. show that shifting only mean vectors cannot account for changes in feature dispersion. Ren et al. show that noise and codecs alter variances and covariances, which mean-only shifts miss. Mean and covariance adaptation based on MCE linear regression and the relevance-factor study of MAP-adapted GMMs show that discriminative information lives in distribution shape. The same work also constrains the design: second-moment adaptation is unstable with little data, because covariance parameters grow quadratically with feature dimension. For that reason we restrict covariances to the 2D transverse plane, per detector region, and study where the crossover in control-sample size occurs.
The collider side supplies the failure mode and the task. The CLEO invisible Υ(1S) search (Search for invisible decays of the Υ ( 1 S ) resonance) uses a recoil-mass tag with hard hermeticity vetoes. It quantifies fake missing energy from gaps and dead cells, and corrects mismodeling only with static scale factors. Baldi, Sadowski and Whiteson introduce the HIGGS and SUSY benchmarks. There, deep networks on low-level features that include MET magnitude and direction beat shallow models, but training and testing use identically distributed simulation. DEFE reaches high discrimination on the ATLAS Higgs challenge data, also with no shift between training and test.
Two lines of work supply the closest methods, and both differ from ours. DeepTau v2.5 aligns simulation and real data by adversarial training of the full shared network. This cut jet misidentification by 30–50%, yet the authors report weaker agreement on 2022 data and leave retraining as future work. Their alignment is purely statistical and knows nothing about momentum balance. RoPE fine-tunes a summary network on 10–50 real calibration samples and models the remaining gap with optimal transport. It is shown only on low-dimensional systems, and its authors flag high-dimensional scaling as a limitation. We borrow RoPE's small-calibration-set regime but put the correction into an interpretable, momentum-conserving response model rather than into free network weights.
For updating the classifier we reuse the approach of Majumdar et al.: frozen base weights plus small adapters, so the model for the original running condition is untouched. For the future, SAML suggests that routing among several low-rank modules could handle many run periods. It also warns that such routing can collapse onto a few modules, which is why we start with one adapter per period.
Research Program
The response model acts on each reconstructed object $i$ in detector region $r$ (for example barrel versus endcap jets, and leptons). Each object's transverse momentum vector $\vec p_{T,i}$ is rescaled and smeared, and MET is recomputed from the corrected objects plus a corrected unclustered-energy term $\vec u$:
$$\vec p_{T,i}' = (1+a_r)\,\vec p_{T,i} + \epsilon_i,\qquad \epsilon_i \sim t_{\nu_r}\!\big(0,\ \Sigma_r(p_{T,i})\big),\qquad \vec E_T^{\,\text{miss}\,\prime} = -\sum_i \vec p_{T,i}' - \vec u' .$$
The symbols are as follows. $a_r$ is a scale (mean) correction. $\Sigma_r$ is a 2×2 resolution covariance whose size grows with $p_T$ in a simple parametric form. $t_{\nu_r}$ is a Student-t distribution whose degrees-of-freedom parameter $\nu_r$ controls tail heaviness. $\vec u'$ receives its own scale, covariance and tail.
This gives a nested family of corrections: mean-only ($a$), mean plus covariance ($a,\Sigma$), and full ($a,\Sigma,\nu$). All told the family has tens of parameters, not thousands. Parameters are fit so that simulated background, passed through the model, matches a target-condition control sample across the whole feature vector, not just the MET marginal. The matching loss is a sliced-Wasserstein distance, with a classifier two-sample test as a check. Because MET is recomputed from the corrected objects, a jet correction moves MET in the right direction and by the right amount. A correction applied to MET alone cannot guarantee this.
The fitted model is then applied to labelled source simulation, both signal and background, to produce a labelled target-like training set. A rank-$r$ low-rank adapter on the frozen classifier is trained on it, with a damage-control check that performance under the original condition does not degrade.
Baselines
The naive reference point is to do nothing: apply the source-trained classifier to target events unchanged. It sets the size of the problem. Every method is evaluated on the same held-out target events and working points.
The strongest discriminator on this benchmark family is the deep network of Baldi et al., reported to raise discovery significance on HIGGS from 2.5σ (shallow) to 4.9σ. On the ATLAS challenge data, DEFE reports an AUC of 0.916. These are the frozen backbones. Both numbers are for unshifted simulation. We found no reported result on Lacuna for any method on HIGGS, SUSY or the ATLAS challenge data under a controlled detector-response shift, so every shifted-condition number will be produced here.
The strongest existing adaptation method is adversarial domain alignment in the style of DeepTau v2.5, with a gradient-reversal domain head. We run it on the same backbone and give it the same target control sample and the same number of tuning trials. Its reported 30–50% misidentification reduction is on CMS tau data and does not transfer to our benchmarks as a number.
We also run three further comparisons:
- RoPE-style fine-tuning. Fine-tune the summary layers on the control sample, then apply the optimal-transport correction of RoPE.
- Density-ratio reweighting. Reweight source background to match the target control sample, then retrain.
- Retraining on corrected simulation. Fully retrain the classifier on simulation passed through our fitted correction.
Two ablations isolate the key ingredients. The first compares mean-only, mean-plus-covariance and full corrections. The second compares a MET-only correction, which applies the same family directly to the MET vector without going through objects, against the object-level, momentum-conserving version.
We count the method as beating a baseline if two things hold on held-out target data at matched calibration budget. First, its background-yield bias in the tightest selections is at least halved relative to that baseline. Second, the significance it recovers is statistically distinguishable from the baseline's across seeds.
Why not just…?
Why not just apply energy-scale factors? A scale factor is the mean-only member of our correction family, so our ablation measures it directly. The speech literature (Watanabe et al., Ren et al.) predicts it cannot repair a widened or heavier-tailed resolution. CLEO likewise relied on static scale factors for shower mismodeling. If mean-only closes most of the tail bias in our shifted benchmarks, the central claim is refuted.
Why not just fine-tune the classifier on target data? A background-dominated control region contains no labelled signal. Supervised fine-tuning would have to pair source signal with target background, which teaches the network to recognise the running condition rather than the physics. Unsupervised full-network alignment is the adversarial baseline, and it is a real threat. It can reduce the mismatch, as DeepTau v2.5 shows. However, it may do so by suppressing resolution-sensitive directions that also carry signal information. We will measure both its tail bias and its lost significance. Majumdar et al. additionally show that unconstrained fine-tuning damages the source domain: in their speech experiments, source error rose from 5.11% to 7.68%, versus 5.40% with adapters.
Why not just reweight simulation to data and retrain? Reweighting is standard practice and is a serious competitor when the shift is mild. It cannot create events in regions that simulation never populated, and heavier fake-MET tails are exactly such regions. In those regions the weights blow up. Our experiments sweep tail heaviness to locate where reweighting's variance overtakes our fitted correction. If no such point exists within realistic shifts, we will report that.
Why not just re-smear simulation by hand and retrain? That is what our method does downstream. The contribution is estimating the smearing parameters, including covariance and tails, from a small control sample while keeping momentum balance. If fully retraining on the corrected simulation matches the low-rank adapter, we will recommend full retraining. The adapter question is secondary to the claim.
Proposed Experiments
We use the HIGGS and SUSY benchmarks of Baldi et al. and the ATLAS Higgs challenge data of Ma et al.. All three contain object momenta and MET with direction, so shifts can be injected consistently: perturb objects and propagate the changes into MET. We inject four families of target conditions:
- a pure scale shift;
- an inflated, anisotropic resolution covariance;
- heavy tails;
- an out-of-family shift, namely azimuth-dependent losses that mimic a dead calorimeter region.
The out-of-family case tests what happens when the response model is wrong. Control samples range from $10^2$ to $10^5$ target background events. An honest limitation is that these are controlled perturbations of simulation, not real data. Parameter recovery is therefore checkable, but realism is not. A real-data test is the natural next step, where CMS-style $Z\to\tau\tau$ control regions like those in DeepTau v2.5 would serve.
The metrics are:
- background-yield bias (predicted versus true background in the target condition), plotted against selection purity down to background efficiencies of about $10^{-4}$;
- recovered discovery significance;
- AUC, which we expect to be uninformative;
- source-condition degradation;
- how closely the fitted $\Sigma_r$ and $\nu_r$ match the injected values.
Four checks decide the claim.
- Bias versus purity. In the covariance and tail shifts, mean-only correction should show a bias that grows with purity while the full correction stays flat. If the curves coincide, the claim fails.
- Crossover in control-sample size. Mirroring the overfitting trade-off known from speech, we expect mean-only correction to win at the smallest control samples and covariance-level correction to overtake it somewhere between $10^2$ and $10^3$ events. Locating that point is itself a practical result.
- Momentum conservation. The object-level correction should beat the MET-only correction on the jet-mismeasurement shifts, where MET-only correction cannot keep MET aligned with the jet that caused it.
- Robustness. We inject a small signal fraction into the control sample to check that the correction does not absorb real signal. We also run the out-of-family azimuthal shift and report plainly whether the low-rank adapter recovers what the response model misses.
References
- Acoustic model adaptation based on coarse/fine training of transfer vectors and its application to a speaker adaptation task
- Acoustic model adaptation for coded speech using synthetic speech
- Mean and covariance adaptation based on minimum classification error linear regression for continuous density HMMs
- Study on the Relevance Factor of Maximum a Posteriori with GMM for Language Recognition
- Search for invisible decays of the<mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline"><mml:mi>Υ</mml:mi><mml:mo stretchy="false">(</mml:mo><mml:mn>1</mml:mn><mml:mi>S</mml:mi><mml:mo stretchy="false">)</mml:mo></mml:math>resonance
- Searching for exotic particles in high-energy physics with deep learning
- Deep extreme feature extraction: New MVA method for searching particles in high energy physics
- Identification of tau leptons using a convolutional neural network with domain adaptation
- Addressing Misspecification in Simulation-based Inference through Data-driven Calibration
- Damage Control During Domain Adaptation for Transducer Based Automatic Speech Recognition
- SAML: Speaker Adaptive Mixture of LoRA Experts for End-to-End ASR