When a very heavy particle decays at a collider, its debris flies out as tight sprays of particles called jets. Physicists rebuild the parent particle by grouping the debris back together, and the key difficulty is knowing where one parent's debris ends and the next begins. Most methods grow each jet by repeatedly merging nearby pieces. They then stop according to a fixed rule, such as a fixed radius, a radius that shrinks with momentum, or a veto when the combined mass jumps suddenly. Visual object trackers face a closely related problem: they must keep updating a target's appearance without absorbing an occluding object or a look-alike neighbour. The best trackers do not use fixed thresholds for this. They learn a confidence score that decides, step by step, whether an update should be accepted.
This note proposes to replace the hand-tuned stopping rule inside jet clustering with a small learned gate. At every merge step, the gate reads a few safe summary quantities of the two pieces being joined, together with the positions of the other hard sprays in the event. It then decides whether the merge would absorb a neighbour's debris. The gate is trained on simulation truth, is not told the masses of the particles being sought, and is restricted to inputs that keep the jets well-defined in theory. The claim is that such a gate recovers the gains of the best hand-tuned adaptive algorithms at their tuned mass point. It should also hold those gains on resonance masses it never saw during training, which is where fixed-threshold methods have to be retuned and mass-based assignment fails.
Core Claim and What Is New
Adaptive clustering already beats fixed cones. Variable-radius clustering, where the effective radius scales as $1/p_T$ (transverse momentum, the momentum perpendicular to the beam), was studied in Fat b-jet analyses using old and new clustering algorithms in new Higgs boson searches at the LHC and Fat $b$-Jet Analyses Using Old and New Clustering Algorithms in New Higgs Boson Searches at the LHC. In a heavy-Higgs to four-b-quark search, it raised significance from 10.799 to 12.897, and to nearly 14 with trimming. The mass-jump algorithm of How to decontaminate overlapping fat jets stops recombination when the merged mass jumps far above the prong masses, controlled by two thresholds $\mu$ and $\theta$. In a ten-jet top-partner final state it reached $S/\sqrt{B}=4.13$ against 3.49 for the best fixed-radius Cambridge–Aachen setup. These papers motivate the proposal, because they show that the stopping rule matters. They also expose its limit. The thresholds $\mu, \theta, R_{\min}, R_{\max}, \rho$ are static per analysis and must be retuned for each resonance mass and boost. The companion "bucket" step, which assigns subjets to parents, needs the parent masses in advance.
The tracking literature supplies the mechanism. State-Aware Tracker for Real-Time Video Object Segmentation computes a per-frame state score and uses it to block polluted masks from entering the target's memory. ARTrackV2 uses a predicted-overlap indicator to halt appearance updates under occlusion. Distractor-Aware Siamese Networks explicitly models high-scoring non-target proposals so the tracker does not drift onto them. We reuse two ideas from this work. The first is a learned gate on each update, rather than a fixed threshold. The second is explicit conditioning on nearby distractors. What we do not reuse is equally important. None of these trackers respects the physics requirement that an observable be infrared and collinear safe, meaning unchanged by adding an infinitely soft particle or splitting one particle into two collinear ones. Tracking new physics at the LHC and beyond shows why this cannot be waved away: track-only reconstruction breaks this safety and inherits hadronization-model uncertainty. Our design constraint is therefore that the gate must never be able to break it.
Concretely, clustering proceeds in Cambridge–Aachen order, which merges pieces by angular distance. A merge of pieces $i$ and $j$ is always accepted if the softer one carries momentum fraction $z_{ij}$ below a small cut $z_{\text{cut}}$. Otherwise it is accepted only if
$$g_\phi\big(x_{ij},\, \mathcal{C}_{ij}\big) > \tfrac12 ,$$
where $x_{ij}$ holds safe two-prong quantities: the two masses divided by the merged mass, $z_{ij}$, the angular separation, and the merged mass divided by $p_T$. The distractor context $\mathcal{C}_{ij}$ lists the angular offsets and momentum fractions of the other hard proto-jets above a fixed momentum threshold. The function $g_\phi$ is a small network with parameters $\phi$.
Soft merges bypass the gate, and collinear pieces are merged before any hard decision is reached. So soft or collinear perturbations should not change which hard decisions are made, which is the same logic that keeps mass-drop and mass-jump tagging safe. The gate is trained with a cross-entropy loss on truth labels: a merge is "good" if both pieces take most of their energy from the same parent parton. Crucially, no parent-mass input is given.
Research Program
The work has three parts.
The first part rebuilds the two published testbeds in one shared pipeline. The first testbed is the Type-II two-Higgs-doublet benchmark with $m_H\approx700$ GeV decaying to $hh\to4b$ at 14 TeV with pileup (MadGraph5, Pythia8, Delphes CMS card), with QCD $4b$, $t\bar t$ and $Zb\bar b$ backgrounds. The second is the 1 TeV vectorlike top partner $T\bar T\to t\bar t HH$ in a ten-jet final state, with $t\bar t$+jets and $t\bar tH$ backgrounds. Every baseline is re-run inside this pipeline, so that comparisons do not depend on differences between our simulation and the original papers'.
The second part trains the gate on a grid of heavy-resonance masses and holds some masses out entirely. This follows the logic of the GOT-10k protocol, which tests trackers only on object classes absent from training and found that doing so exposes overfitting that ordinary splits hide. Here, the "classes" are resonance masses and boost ranges.
The third part tests theoretical safety directly. We inject soft particles and apply collinear splittings to the same events and measure how much the reconstructed jets and masses move. We compare this against anti-$k_T$, mass-jump and variable-radius clustering.
The signal-to-background lens sets how we judge success. The PandaX-III study, Topological background discrimination, shows that stacking static cuts can achieve huge background rejection while keeping only about 20% of the signal. We therefore report signal efficiency and background mass sculpting alongside significance, rather than significance alone.
Baselines
The naive reference point is fixed-radius clustering: anti-$k_T$ and Cambridge–Aachen with the best radius from a small grid, followed by the same downstream selection. In the cited testbeds it scored $\Sigma=10.799$ (Fat $b$-Jet Analyses) and $S/\sqrt{B}=3.49$ (How to decontaminate overlapping fat jets). The proposal must beat it clearly, or nothing else matters.
The strongest existing method on the crowded testbed is mass-jump clustering with bucket assignment (MJ10), reported at $S/\sqrt{B}=4.13$ in How to decontaminate overlapping fat jets. On the four-b testbed, the strongest is variable-radius clustering plus trimming, reported at nearly 14 in Fat $b$-Jet Analyses. These are the right competitors because they encode the same intuition as our gate, with hand-set thresholds instead of learned ones.
For fairness, we give their thresholds ($\mu,\theta$ and $R_{\min},R_{\max},\rho$, plus trimming parameters) a grid or Bayesian search. That search uses the same number of evaluations on the same validation events as the gate's hyperparameter search, and is run separately at each mass point where retuning is permitted. The proposal counts as beating these baselines under two conditions. First, it must match them within statistical uncertainty at their tuned mass point. Second, it must exceed them at held-out masses where they must use thresholds tuned elsewhere. In both cases this must happen without worse signal efficiency or background mass sculpting.
Two ablations isolate the key ingredients. In the first, the gate drops the distractor context $\mathcal{C}_{ij}$ and sees only the pair being merged. This tests whether the tracking-style distractor awareness matters, or whether a learned mass-jump rule is enough. In the second, the gate is given the parent-mass hypothesis as an input. This measures how much generality costs at the trained mass.
Why not just…?
Why not just retune the mass-jump thresholds for each mass point? This is the most serious threat, because at any single known mass a tuned two-parameter rule may already be near optimal, and our design would only match it. The held-out-mass experiment settles this. If thresholds tuned at one mass transfer to other masses as well as the learned gate does, then the gate adds nothing, and we will say so. The file's evidence that retuning is required is qualitative: both jet papers note the sensitivity to parameters. It is not a measured transfer curve, so producing that curve is itself a contribution.
Why not just add trimming to variable-radius jets? Trimming helps. It lifted variable-radius significance from 12.897 to nearly 14 in Fat $b$-Jet Analyses. But trimming removes soft contamination inside a jet. It does not stop a jet from swallowing a neighbouring hard prong, which is the failure mode that dominates the ten-jet final state. We expect the two to be complementary, and will test the gate with and without trimming on top.
Why not skip clustering and train a classifier on all particles in the event? At a fixed signal hypothesis, such a classifier may well achieve higher raw $S/\sqrt{B}$, and we will run one as an upper reference. Its output, however, is a score rather than a set of jets with physical masses. Resonance searches at unknown mass need reconstructed objects whose mass spectra can be scanned for a bump, and a classifier trained at one mass tends to sculpt the background mass distribution. We will report that sculpting explicitly. We claim reusable, theoretically safe jets, not the best possible single-hypothesis discriminator.
Proposed Experiments
The decisive checks are four. On the ten-jet top-partner testbed at matched luminosity, we compare significance, signal efficiency, and top and Higgs mass resolution and bias against re-run MJ10 and fixed-radius Cambridge–Aachen. The same comparison is made on the four-b testbed against variable-radius clustering with and without trimming. We then repeat the comparison at resonance masses held out from training, with each baseline using thresholds tuned at the nearest training mass. We perform the soft and collinear perturbation test, and require that the gate's jet-mass shifts stay at the level of mass-jump clustering rather than of track-only or unconstrained learned methods. Finally, we scan pileup to see whether distractor conditioning keeps its advantage as soft contamination grows, which is the open problem both jet papers leave standing.
The result is negative if the no-distractor ablation matches the full gate, or if retuned thresholds match it at held-out masses. That would mean the tracking analogy supplies nothing beyond a fitted cut, and we would report it as such.
References
- Fat b-jet analyses using old and new clustering algorithms in new Higgs boson searches at the LHC
- How to decontaminate overlapping fat jets
- State-Aware Tracker for Real-Time Video Object Segmentation
- ARTrackV2: Prompting Autoregressive Tracker Where to Look and How to Describe
- Distractor-Aware Siamese Networks for Visual Object Tracking
- Tracking new physics at the LHC and beyond
- GOT-10k: A Large High-Diversity Benchmark for Generic Object Tracking in the Wild
- Topological background discrimination in the PandaX-III neutrinoless double beta decay experiment