Year: 2026
Type: preprint
Abstract: Knowledge distillation for jet taggers is usually cast as compression toward a near-oracle teacher. We instead train a small reliability-gated correction layer jointly over two frozen teachers --- a GN2-type ATLAS flavour tagger (3 classes) and CMS ParT (10 classes) --- reading only their frozen penultimate embeddings and logits, so it adds negligible marginal cost. A single per-sample reliability signal reweights the distillation loss, drives a prototype-based repair of the teachers' errors, and anchors the inference-time gate that mixes teacher and student logits. Zero-initialised, the 82\,k-parameter student starts as a bit-exact copy of the teacher --- a provable invariant on the training distribution --- and departs only where the data support it. Under a statistically-grounded evaluation, with binomial, Poisson and paired uncertainties on every table, it reproduces both teachers at $\geq\!99\%$ agreement and matches their overall accuracy while redistributing it toward their weakest classes: the JetClass per-class accuracy of $H\!\to\!4q$, QCD and $H\!\to\!bb$ rises by $+0.8$, $+0.4$ and $+0.2$ percentage points ($|z|>7$), offset by comparable drops elsewhere. Oracle heads on the frozen embedding recover a budget consistent with zero in both domains, and a conformal-risk-control analysis certifies the false-accept rate. This full re-evaluation supersedes an earlier, incorrectly-evaluated set of numbers; code and models are available at https://github.com/sysu17363020/JetCoRD.
Source: arXiv (Cornell University)
Language: English
Subfield: Nuclear and High Energy Physics
Field: Physics and Astronomy
Domain: Physical Sciences
SDG: Quality education
Open Access status: green
OpenAlex ID: W7169756643