🤖 AI

RINO: Renormalization Group Distillation with No Labels

RINO introduces a novel self-supervised learning framework for high-energy physics that leverages the multi-scale renormalization group structure of jet formation to pretrain models on unlabeled QCD data, achieving superior out-of-distribution classification performance and data efficiency compared to existing methods.

Original authors: Zichun Hao, Raghav Kansal, Abhijith Gandrakota, Chang Sun, Haoyang Li, Zihan Zhao, Ngadiuba Jennifer, Javier Duarte, Maria Spiropulu

Published 2026-10-15

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Zichun Hao, Raghav Kansal, Abhijith Gandrakota, Chang Sun, Haoyang Li, Zihan Zhao, Ngadiuba Jennifer, Javier Duarte, Maria Spiropulu

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

In the high-energy physics laboratories that probe the deepest secrets of the universe, scientists smash particles together at nearly the speed of light. The goal is to catch a glimpse of rare, heavy particles that might explain how the cosmos works. However, these collisions produce a chaotic spray of thousands of smaller particles, creating a complex cloud of data that is incredibly difficult to sort through. To make sense of this, researchers rely on powerful computer programs trained to recognize specific patterns, much like a security guard learning to spot a particular face in a crowd. But there is a fundamental problem: these programs are usually trained on computer simulations. While these simulations are our best guess at what the real world looks like, they are never perfect. They miss subtle details of how particles interact and how detectors record them. When a model trained on this imperfect simulation is finally applied to real data from the collider, it often stumbles, losing a significant portion of its accuracy. This gap between the simulated training ground and the real experimental battlefield has long limited what artificial intelligence can achieve in physics.

A team of researchers has now proposed a new way to bridge this gap, bypassing the need for perfect simulations to teach the computer how to see. Instead of relying on labeled examples of rare particles, which are scarce and often simulated, they taught their system using the vast, unlabeled background of ordinary particle collisions. They focused on a specific type of particle spray called a "jet," which forms when a quark or gluon is ejected from a collision and breaks apart into a shower of other particles. These jets have a unique, multi-layered structure. Just as a tree has a trunk, branches, and twigs, a jet has a core, sub-branches, and fine details, all existing at different levels of resolution. The researchers realized that this natural hierarchy could serve as a teacher. They developed a method called RINO, which stands for Renormalization Group DIstillation with NO Labels. This approach treats the same jet at different levels of detail simultaneously, forcing the computer to understand how the object looks when viewed broadly and when examined closely, without ever needing to know what the jet actually is.

The core of this new method lies in how the computer views the data. In previous attempts to teach machines about jets without labels, researchers often used random tricks to create different versions of the same image, similar to how one might crop or rotate a photograph. But jets are not photographs; they are collections of moving points with no fixed shape. The RINO team instead used a physical process called clustering to create their views. Imagine taking a single jet and grouping its particles together, starting with the most tightly bound pairs and working outward. By stopping this grouping process at different stages, the researchers could generate a series of views of the same jet: one showing the full, fine-grained spray of particles, and others showing the same jet as a few coarse, large clumps. These different views represent the same physical object at different scales of resolution. The system then uses a "teacher" and a "student" setup. The student looks at the jet in its most detailed form, while the teacher looks at the same jet in a coarser, simplified form. The student is tasked with predicting what the teacher sees, learning to recognize the underlying structure that remains consistent regardless of how much detail is visible.

This approach proved remarkably effective at teaching the computer to recognize the true nature of the particles. When the researchers tested their system, they found that it could learn to identify rare, heavy particles, such as those from top quarks, with high accuracy even after being trained on only a tiny fraction of labeled data. In fact, the system trained with their new method using just one percent of the available labeled examples performed as well as a traditional system trained on one hundred percent of the labeled data. This suggests that the method captures the essential physics of the jets so well that it does not need vast amounts of specific examples to learn the job. Furthermore, the system showed a strong ability to generalize. When tested on a type of particle decay it had never seen before—a Higgs boson breaking into four quarks—the system performed significantly better than other methods, even without any retraining. This indicates that the system learned the general rules of how particle sprays form, rather than just memorizing specific examples.

The success of this method also revealed something important about how these computer models learn. The researchers discovered that the best performance came not from making the model completely indifferent to the scale of the view, but from allowing the model to retain a sense of scale. In other words, the system worked best when it could remember that a jet looks different at a coarse level than at a fine level, rather than trying to force all views to look identical. This insight challenges some previous assumptions in the field, where the goal was often to make the model's understanding of an object completely independent of how it was viewed. By preserving the relationship between the different scales, the model became more robust and better at distinguishing between the common background noise and the rare, interesting signals. The findings suggest that by grounding artificial intelligence in the fundamental physical structures of the data itself, scientists can build tools that are far more reliable when applied to the messy, imperfect reality of experimental physics. This work offers a promising path forward for using machine learning to discover new physics, reducing the reliance on imperfect simulations and allowing researchers to trust their tools more when they look at the real universe.

Technical Summary: RINO – Renormalization Group Distillation with NO Labels

Problem Statement

In High-Energy Physics (HEP), supervised machine learning models rely on labeled Monte Carlo (MC) simulations to train classifiers for tasks like jet tagging. However, these simulations are imperfect representations of real detector data, leading to significant domain shifts. When models trained on simulation are applied to real data, performance can degrade by 20–40%. While Self-Supervised Learning (SSL) offers a path to pretrain on unlabeled real data, existing methods often rely on inductive biases derived from computer vision (e.g., image cropping) or natural language, which do not naturally translate to the variable-length, point-cloud nature of particle jets. Furthermore, many existing physics-motivated SSL methods operate at a single physical scale, ignoring the fundamental multi-scale structure inherent in jet formation.

Methodology: The RINO Framework

The authors propose RINO (Renormalization Group DIstillation with NO Labels), an SSL framework that leverages the intrinsic multi-scale renormalization group (RG) structure of Quantum Chromodynamics (QCD) jet formation.

1. Physics-Informed View Construction

Instead of artificial data augmentations, RINO generates "views" of the same jet by varying the resolution of its description. This is achieved using the $k_T$ clustering algorithm, which approximately inverts the QCD parton shower (RG flow). By stopping the clustering sequence at different numbers of subjets ( $N$ ), the method produces a sequence of kinematic descriptions ranging from a fine-grained unclustered particle cloud to coarse-grained representations with few subjets ( $N \in \{2, 3, 4, 6, 8, 16\}$ ). These views correspond to physically meaningful coarse-graining steps.

2. Self-Distillation Architecture

RINO adopts a teacher-student self-distillation framework (inspired by DINO):

Teacher: Receives intermediate-scale views ( $N \in \{6, 8, 16\}$ ).
Student: Receives views across the full spectrum, including coarse-grained ( $N \in \{2, 3, 4\}$ ) and unclustered ($uncl.$) states.
Objective: The student is trained to match the teacher's representations via a composite loss function:
- $L_{DINO}$ : Aligns representations across the view hierarchy using Sinkhorn–Knopp centering to prevent collapse.
- $L_{iBOT}$ : A masked particle prediction task where the student predicts the teacher's representation of masked particles.
- $L_{KoLeo}$ : A differential entropy estimator to regularize embeddings for diversity on the hypersphere.
Update Mechanism: The teacher's parameters are updated via an Exponential Moving Average (EMA) of the student's parameters.

3. Training Protocol

To simulate realistic deployment conditions where signal data is scarce and background-dominated:

Pretraining: Exclusively on QCD jets from the JETCLASS dataset (excluding all signal classes like top quarks, W/Z bosons, and Higgs bosons).
Finetuning: Performed on the smaller JETNET dataset (simulating limited labeled simulation data).
Evaluation: Out-of-Distribution (OOD) transfer is evaluated on the signal classes of the larger JETCLASS dataset, which serves as the target distribution with a distribution shift relative to JETNET.

Key Contributions

First Multi-Scale SSL for Jets: RINO is the first SSL framework to align jet representations across multiple $k_T$ clustering scales, exploiting the RG flow structure of jet formation.
Realistic Controlled Comparison: The framework is pretrained solely on QCD background (mimicking real collider data) and evaluated on rare signal classes, addressing the overrepresentation of signals in prior SSL studies.
Ablation of Inductive Biases: The study isolates the contributions of multi-scale view construction versus the SSL objective (contrastive vs. self-distillation), revealing that preserving scale-dependent structure in the backbone is critical for performance.

Experimental Results

The paper evaluates RINO against supervised training and recent SSL baselines (JetCLR, MPMv1, MPMv2, OmniJet-α) using top-quark jet tagging (hadronic $t \to bqq$ ) as the primary downstream task.

OOD Performance: With a full labeled dataset for finetuning, RINO improves OOD classification accuracy on top-quark jets by 5.9% over the strongest SSL baseline (MPMv1) and by 7.8% over supervised training from scratch.
Data Efficiency: RINO matches the performance of fully supervised training on the entire JETNET dataset when finetuned on only 1% of the labeled data.
Cross-Topology Generalization: Without retraining, RINO improves the Area Under the Curve (AUC) on the unseen Higgs boson decay channel $H \to 4q$ by 4.1% over the strongest SSL baseline. It demonstrates the ability to generalize to 4-prong topologies despite being finetuned only on 3-prong top-quark jets.
Domain Shift Metrics: RINO exhibits significantly lower distribution shift between finetuning and evaluation domains, with a 3 $\times$ smaller 1-Wasserstein distance (W1) and 6–7 $\times$ smaller squared Maximum Mean Discrepancy (MMD $^2$ ) compared to baselines.
Ablation Insights:
- Replacing standard augmentations with $k_T$ views (JetCLR-scale) yields a +11.4% gain over standard JetCLR.
- Adding cross-scale self-distillation to an iBOT-only baseline yields a +14.0% gain.
- The combination of multi-scale views and self-distillation (RINO) outperforms the combination of multi-scale views and contrastive learning (JetCLR-scale) by 8.6%.
- Scale-Awareness: Analysis reveals that RINO's backbone retains scale-dependent structure (unlike scale-invariant backbones in other methods), which the projection head shields to achieve invariance. This preservation of scale information is identified as a key driver of the performance gap.

Significance and Claims

The paper claims that RINO demonstrates the viability of exploiting fundamental physical structures (RG flow) to construct SSL views, moving beyond domain-agnostic augmentations. By pretraining on background-dominated QCD data, RINO reduces reliance on imperfect signal simulations and achieves robust transfer to rare signal classes. The work highlights that preserving scale-dependent structure in the backbone while achieving invariance through the projection head is a superior strategy for OOD generalization in HEP compared to enforcing scale invariance directly on the backbone.

The authors note limitations, including that the evaluation relies on a simulation-to-simulation proxy (JETNET to JETCLASS) rather than direct real-data validation, and that the $k_T$ inductive bias is most beneficial for multi-prong substructure tasks rather than flavor-tagging tasks driven by track-level features.