Many tests of the theory of the strong force hinge on one question: how often does a gluon split into a heavy quark pair ($g \to b\bar{b}$ or $g \to c\bar{c}$) inside a jet? Leading-order calculations miss this mechanism, and it is a leading suspect behind measured heavy-flavor excesses. At RHIC, bottom production came out roughly double the theoretical prediction. Parton-shower simulators such as PYTHIA and HERWIG++ disagree with each other on exactly these splitting categories. The trouble is that every tool used to count splittings in real collisions is calibrated on one of those same disagreeing simulators. A measurement meant to test theory therefore quietly inherits the theory of whichever generator calibrated it.
Speaker recognition faced a similar problem and largely solved it. A speaker embedding is a fixed-length vector computed from a variable-length recording. It must identify the voice while ignoring the microphone, the room, and the words spoken. The field learned to get there by training a network to tell apart thousands of speakers, recorded under many deliberately varied channels, and then judging the system on trials that cross channels. We propose to treat the hard partonic splitting as the "speaker" and the hadronization and shower model as the "channel." We would train a jet embedding to classify hundreds of fine-grained truth-level splitting configurations under many generators and tunes. The test that matters is whether a splitting rate measured with this embedding stays unbiased on a generator it never saw.
Core Claim
A jet representation trained to separate many fine-grained hard-splitting classes across several parton-shower generators will have two properties. It will discriminate $g \to Q\bar{Q}$ jets from other jets at least as well as a single-generator transformer. And it will cut the generator dependence of the extracted splitting fraction substantially relative to both learned and analytic taggers. The new element is not a new layer. It is a change of objective and of evaluation, both borrowed from speaker recognition. The first borrowed idea is to classify many identities so the embedding is forced to carry hard-process information rather than an easy shortcut. The second is to count cross-channel robustness, not in-distribution accuracy, as the result.
Relation to Prior Work
The physics motivation comes from three heavy-flavor measurements. Heavy flavour production at HERA reports beauty cross-sections well above leading-order predictions. The PHENIX dimuon analysis (Measurements of μμ pairs from open heavy flavor and Drell-Yan in p+p collisions at √s=200 GeV) finds bottom production about double the theory. That same analysis finds that charm correlations favor PYTHIA over POWHEG and that Drell-Yan agrees with next-to-leading-order QCD. Both analyses motivate the proposal, and PHENIX's reliance on simulation-derived templates also shows how generator choice enters a measurement.
The task definition and the strongest non-learned competitor come from Disentangling heavy flavor at colliders. That paper separates $g \to Q\bar{Q}$ from non-splitting jets using SoftDrop declustering and a FlavorCone algorithm that centers cones on reconstructed heavy hadrons. It compares PYTHIA 8.212, HERWIG++ 2.7.1, VINCIA, DIRE and POWHEGBOX, and documents large differences between generators. We reuse its simulation setup and its truth definitions. We differ in replacing hand-built geometry with a learned embedding, and in measuring robustness rather than only discrimination.
Two learned jet models constrain the architecture and supply comparisons. Particle Transformer for Jet Tagging injects pairwise kinematic features into attention and summarizes the jet with a single class token. It reaches 86.1% accuracy on JETCLASS, but it is trained and tested on one simulation. Learning Symmetry-Independent Jet Representations via Jet-Based Joint Embedding Predictive Architecture (J-JEPA) pretrains without hand-made augmentations, citing simulation mismodeling as motivation. It reports a background rejection of about 70.5 on top tagging with 10% of the labels. It targets label efficiency on top jets, not generator invariance on heavy-flavor splitting.
From speech we take the method. ECAPA-TDNN uses statistics pooling, meaning the jet or utterance summary includes a per-feature mean and spread rather than one learned token. Its version is channel- and context-dependent attentive pooling. It trains on VoxCeleb2 with noise (MUSAN) and reverberation (RIR) augmentation, and reports 0.87% equal error rate on VoxCeleb1. Its ablation attributes 9.8% of the error reduction to the pooling. Attentive Statistics Pooling for Deep Speaker Embedding introduced the weighted mean-and-spread summary. Neither ever faced a nuisance as structured as a shower model, and whether their gains transfer is an open question we test rather than assume. The observation that classifying many identities forces the network to drop transient cues also comes from this literature. It is the specific ingredient we isolate.
Research Program
We would generate jet samples with the five generators used by Ilten et al. We would add several tunes per generator where available, so that the "channel" set covers a range rather than a few points. For each jet we record the truth splitting history. The identity label is a cell in a grid over splitting type ($g \to b\bar{b}$, $g \to c\bar{c}$, $Q \to Qg$, light), splitting momentum fraction $z$, opening angle, and jet transverse momentum. This yields several hundred classes.
The encoder is Particle Transformer's pairwise-biased attention, run over particles and reconstructed displaced vertices. Its class token is replaced by channel-dependent attentive statistics pooling, so the jet summary is a per-channel weighted mean and spread. It is trained with a margin-based multi-class loss across all generators at once.
A light head on the frozen embedding performs the actual tagging task, $g \to Q\bar{Q}$ versus everything else. Evaluation follows the leave-one-channel-out practice of speaker verification. We train on four generators and test on the fifth, rotating through all five. We also add a closure study that turns a tagger into a measurement. On the held-out generator, we reweight the sample to inject a known $g \to b\bar{b}$ fraction. We then extract that fraction with the tagger calibrated only on the training generators, and record the bias.
Finally, we would carry over the length-uncertainty idea of Incorporating Uncertainty from Speaker Embedding Estimation to Speaker Verification in one narrow check. Jets with few constituents are the analog of short utterances. We would test whether their embeddings are more generator-sensitive, and whether down-weighting them reduces closure bias.
Baselines
The naive reference point is the anti-$k_t$ jet with standard heavy-flavor tagging, as used by Ilten et al. It sets the floor for both discrimination and generator spread.
The strongest analytic competitor is FlavorCone combined with SoftDrop from Disentangling heavy flavor at colliders. The only reported number is a 30–40% improvement over anti-$k_t$ in separating splitting from non-splitting jets. This is the competitor most likely to be robust, because it has no fitted parameters to overfit a generator. We would rerun it on our samples with the published configuration. To beat it, the embedding must match or beat its closure bias on the held-out generator while clearly exceeding its rejection at fixed efficiency.
The strongest learned competitor is Particle Transformer trained directly on the binary task. No reported result was found on Lacuna for ParT on $g \to Q\bar{Q}$ tagging. Its only relevant anchor is 86.1% on JETCLASS, which we would reproduce to validate our implementation. We would run it in two settings, trained on one generator and trained on the pooled four-generator mix. In both it gets the same particle inputs, parameter count within 10%, the same number of training steps, and the same hyperparameter search budget as our model.
Three ablations isolate the proposal's ingredients, since the new element is the combination of objective and evaluation. The first swaps the fine-grained multi-class objective for the binary objective with everything else fixed. The second swaps attentive statistics pooling for the class token. The third swaps multi-generator training for single-generator training. We would also report the embedding's own performance on JETCLASS, to show that the pooling does not damage standard tagging.
Why not just…?
Why not just use SoftDrop and FlavorCone, which are already interpretable? They may well be the most robust option, and we treat that as a live outcome rather than a straw man. Their limitation is the one Ilten et al. document: collimated heavy hadrons and hadronization smearing blur the reconstructed splitting, which caps discrimination. If FlavorCone matches our closure bias and our rejection gain is small, then learned embeddings are not worth their opacity for this measurement. The leave-one-generator-out comparison decides this directly.
Why not just train Particle Transformer on a mix of generators? This is the most serious threat. Multi-condition training is, after all, half of what made speaker embeddings robust. The ParT-on-mixed-generators baseline and the binary-versus-fine-grained ablation exist precisely to test it. Our hypothesis is that a binary objective lets the network lean on whatever generator-specific cue separates two classes most cheaply. Hundreds of kinematic identity classes should instead force it to encode the hard splitting itself. If mixed-generator ParT closes to within the seed-to-seed spread of our model, the contribution collapses to "train on several generators." We would report that plainly.
Why not just fit templates from simulation, as PHENIX did? Template fitting is how the measurement is actually made, and our tagger feeds such a fit rather than replacing it. The issue is that templates taken from one generator transfer that generator's splitting model into the answer. The PHENIX result sits alongside a factor-of-two bottom discrepancy and a PYTHIA-versus-POWHEG preference. The closure study measures exactly this. We run the fit once with templates from the tagger and once with templates from standard kinematic variables, and compare the bias on the held-out generator.
Why not just add mean-and-spread pooling to Particle Transformer and call it done? That change is cheap, and the pooling ablation evaluates it by itself. We make no claim that it matters. The spread statistics might even amplify generator differences, if hadronization changes mostly show up as changes in radiation dispersion.
Decisive Checks
The primary result is the closure bias on the extracted $g \to b\bar{b}$ fraction, averaged over the five held-out generators. It is reported against the raw spread between generators in the true fraction, which is the systematic uncertainty a measurement currently inherits.
The proposal succeeds if two conditions hold together. First, the full model's closure bias is smaller than both FlavorCone's and mixed-generator ParT's. Second, its rejection at 50% signal efficiency beats FlavorCone.
It fails, informatively, in two ways. If the binary-objective ablation matches the full model, the identity-classification idea does nothing here. If the held-out bias remains comparable to the gap between generators, no representation trained on simulation alone can separate hard splitting from hadronization. That would point future effort toward calibration on data rather than toward architecture.
Applying the method to open-flavor data such as the PHENIX forward dimuon sample is the natural follow-up. It is contingent on the simulation result.
References
- Heavy flavour production at HERA
- Measurements of <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline"><mml:mi>μ</mml:mi><mml:mi>μ</mml:mi></mml:math> pairs from open heavy flavor and Drell-Yan in <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline"><mml:mi>p</mml:mi><mml:mo>+</mml:mo><mml:mi>p</mml:mi></mml:math> collisions at <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline"><mml:msqrt><mml:mi>s</mml:mi></mml:msqrt><mml:mo>=</mml:mo><mml:mn>200</mml:mn><mml:mtext> </mml:mtext><mml:mtext> </mml:mtext><mml:mi>GeV</mml:mi></mml:math>
- Disentangling heavy flavor at colliders
- Particle Transformer for Jet Tagging
- Learning Symmetry-Independent Jet Representations via Jet-Based Joint Embedding Predictive Architecture
- ECAPA-TDNN: Emphasized Channel Attention, Propagation and Aggregation in TDNN Based Speaker Verification
- Attentive Statistics Pooling for Deep Speaker Embedding
- Incorporating Uncertainty from Speaker Embedding Estimation to Speaker Verification