Particle physicists who hunt for rare new particles train classifiers that separate a faint signal from an overwhelming background. These classifiers have a well-known failure mode. If a classifier quietly learns to use the particle's reconstructed mass, then cutting on its score reshapes the background mass spectrum. This reshaping, called "mass sculpting," can produce a fake bump that looks like a discovery. Current methods handle it by fixing one penalty weight that trades discrimination against decorrelation and training a single model. Every such model is trained and validated only in simulation, so nobody checks whether the chosen trade-off still holds when the simulator is wrong.
We propose to train one network conditioned on a preference vector, so that it can output a tagger anywhere along the discrimination-versus-sculpting trade-off curve (the Pareto front) on demand. We then use that cheap, complete front as a measuring instrument for the gap between simulation and reality. The question is whether a working point that is Pareto-optimal under one event generator (Pythia 8) is still Pareto-optimal, or is dominated, under a different generator (Herwig++). If the front's shape and ordering move between generators, the field's practice of choosing working points inside one simulator is unsafe, and a held-out generator should be used to pick the preference. If the front is stable, that is useful reassurance that has never been measured.
Core Claim
The trade-off between background rejection and mass sculpting is a property of the simulator as well as of the model. Fixed-penalty methods hide this because they produce one point. Robust Anomaly Detection for Particle Physics Using Multi-Background Representation Learning (Gandrakota et al.) trains on several simulated backgrounds and removes mass dependence with Nuisance Randomized Distillation, which reweights events and adds an adversarial critic. They obtain one maximum-logit model with higher power (AUROC 0.914, significance improvement 2.32) and one Mahalanobis-distance model with less sculpting (Jensen-Shannon divergence, JSD, of 0.118). That is two isolated points from a fixed weighting, evaluated only on Monte Carlo. Accurate and robust methods for direct background estimation in resonant anomaly detection (Das et al.) motivates the simulator angle. They train on Pythia, evaluate on Herwig++, and absorb the drift with a systematic shift parameter $\delta_{\text{sys}}$. Their working points are fixed ad hoc at background efficiencies of $10^{-3}$ and $10^{-4}$.
Our contribution is the combination neither paper has: a full, continuously queryable front, plus a measured statement of how that front deforms across generators. The claim is falsifiable. It fails if front ordering and dominance are preserved across generators to within statistical error.
Research Program
The mechanism comes from scalarization, the practice of collapsing several objectives into one number so ordinary gradient descent applies. We follow the Pareto Set Learning setup catalogued in LibMOON. A small hypernetwork $h_\phi$ maps a preference vector $\lambda$ on the simplex to the weights (or a FiLM-style modulation) of a jet tagger. It is trained to minimize the smooth Tchebycheff scalarization from Smooth Tchebycheff Scalarization for Multi-Objective Optimization, averaged over sampled preferences:
$$\min_\phi \; \mathbb{E}_{\lambda}\left[\mu \log \sum_{i=1}^{m} \exp\!\left(\frac{\lambda_i\,(f_i(h_\phi(\lambda)) - z_i^*)}{\mu}\right)\right].$$
In this objective, $f_1$ is the classification loss, $f_2$ is a differentiable decorrelation loss between score and mass, $z^*$ is the per-objective ideal point estimated from single-objective runs, and $\mu$ is the smoothing temperature. We use distance correlation or the NuRD critic loss for $f_2$, and report JSD at evaluation. We choose the smooth Tchebycheff form over a weighted sum for two reasons. Weighted sums can only reach the convex hull of a front (Park et al.). The smooth form also needs one backward pass per step rather than per-objective gradients plus a quadratic program. Whether the physics front is actually non-convex is an empirical question we test rather than assume.
The study has three stages. First, on the hls4ml jet dataset we compare the learned front to the NuRD operating points. Second, on the LHC Olympics R&D data we train on Pythia, sweep $\lambda$, and evaluate every preference on Herwig++. Third, we test a selection rule: choose $\lambda$ by maximizing discovery significance net of $\delta_{\text{sys}}$ on a small Herwig++ calibration split, and report on a disjoint Herwig++ test split. In a further arm, we add a third objective: score stability under a nuisance perturbation of Pythia events (jet energy-scale smearing, our own construction). This tests whether robustness can be bought during training without ever touching the held-out generator.
We are explicit about one limitation. Herwig++ is still a simulator. Cross-generator shift stands in for the real reality gap; it does not replace it. The proposal measures how much a single-simulator evaluation can mislead, which is a necessary precondition for trusting any of these methods on data.
Baselines
The naive reference is a plain supervised or CWoLa-style tagger with no decorrelation, thresholded at fixed background efficiency. It shows how much sculpting and cross-generator drift occur when nothing is done. We train it on the same Pythia events with the same architecture and tuning budget.
The strongest existing in-simulation method is NuRD from Gandrakota et al., in both its maximum-logit and Mahalanobis-distance variants on hls4ml. Its reported figures are AUROC 0.914 and significance improvement 2.32 for nurd-ml, and JSD 0.118 for nurd-md. The baseline VAE in the same paper reports AUROC 0.881 and JSD 0.255. We rerun NuRD with its penalty swept over the same number of values we query from our front, using the same backbone and total GPU hours. Beating it means one of two things. Either a single preference from our network dominates both NuRD points (AUROC at least 0.914 at JSD at most 0.118), or our front's hypervolume exceeds that of the swept NuRD points at matched compute.
On the LHC Olympics data, the strongest reference is the direct cut-and-count estimator of Das et al. with CWoLa and CATHODE templates. They report over 5σ discovery significance at background efficiencies of $10^{-3}$ and $10^{-4}$, 8σ for an idealized detector, and correctly low significance on the background-only Black Box 2. We reproduce their protocol exactly and substitute our selected preference for their fixed working point. Beating it means higher Herwig++ test significance at equal or smaller $|\delta_{\text{sys}}|$, with no false discovery on Black Box 2.
The key ablation is a linear-scalarization hypernetwork trained identically, which isolates whether the Tchebycheff form matters. A second ablation selects $\lambda$ on Pythia alone, which isolates whether the held-out generator adds anything.
Why not just…?
Why not just sweep the NuRD penalty weight across many independent training runs? This is the most serious threat. With enough compute, independent runs trace the same convex part of the front, and the diagnostic part of our proposal does not require a hypernetwork. We handle it head on by running the sweep at matched compute. If the sweep matches our hypervolume and cross-generator results, we will report that the conditioned network is only a convenience. The central contribution, measuring how the front deforms between Pythia and Herwig++ and selecting working points on a held-out generator, would stand either way.
Why not just use a weighted sum instead of Tchebycheff scalarization? If the discrimination-sculpting front is convex, a weighted sum is enough and simpler. The theory in Park et al. and LibMOON only says linear weights miss concave regions; it does not say this front has any. Physics vs Distributions: Pareto Optimal Flow Matching with Physics Constraints found fixed linear weighting brittle in a related generative setting, but that is a different problem. The linear-hypernetwork ablation settles the question. We will report whether any Tchebycheff preference lies outside the convex hull of the linear front.
Why not just avoid sculpting with cut-and-count estimation, as Das et al. do? Cut-and-count removes the need to fit a mass shape. It still requires a working point, and $\delta_{\text{sys}}$ flags simulator mismatch only after the fact. Our method is complementary: it supplies the working point and tests whether choosing it under shift improves the significance that cut-and-count reports.
Why not just use multi-gradient methods such as MGDA or EPO? They can also reach non-convex regions. The smooth Tchebycheff paper reports being 36 to 44 times faster than EPO in Pareto Set Learning while matching coverage. We include EPO on hls4ml only, as a cost-quality check.
Proposed Experiments
The decisive experiment uses the LHC Olympics R&D data. We train on Pythia, query 50 preferences, and compute discrimination and sculpting for each on both Pythia and Herwig++. We report three quantities: the Kendall rank correlation of preference orderings between generators, the fraction of Pythia-Pareto points that become dominated under Herwig++, and the hypervolume loss. Bootstrap resampling over events gives the null band. A rank correlation near one and no newly dominated points falsifies the core claim.
The second check is the selection rule. We compare Herwig++-test significance and $\delta_{\text{sys}}$ for three choices: preferences chosen on Herwig++ calibration data, preferences chosen on Pythia, and the fixed working points of Das et al. In every configuration we repeat the Black Box 2 background-only test, so that no gain comes from false discoveries. We also run Black Box 1 to confirm that the known resonance is still found.
The third check, on hls4ml, places our front against the NuRD and VAE numbers above. It reports whether the Tchebycheff front contains points the linear front cannot reach.
References
- Robust anomaly detection for particle physics using multi-background representation learning
- Accurate and robust methods for direct background estimation in resonant anomaly detection
- LibMOON: A Gradient-based MultiObjective OptimizatioN Library in PyTorch
- Smooth Tchebycheff Scalarization for Multi-Objective Optimization
- Thompson Sampling for Multi-Objective Linear Contextual Bandit
- Physics vs Distributions: Pareto Optimal Flow Matching with Physics Constraints