Year: 2026
Type: article
Abstract: Large language models (LLMs) have demonstrated that scaling, driven by compute, can yield dramatic improvements in performance, with current state-of-the-art systems reaching up to one trillion parameters. In contrast, state-of-the-art jet flavor classifiers in High-Energy Physics (HEP) remain in the tens of millions of parameters. Scaling laws offer a framework for predicting how much HEP models benefit from systematic increases in capacity and training data. In this work, we study the behavior of increasingly larger flavor taggers, deriving scaling laws under optimal use of compute budget and limited datasets. We also identify the regimes where “double descent” emerges. The emerging scaling behavior points to a clear need for significantly larger datasets and models to saturate available compute and approach the limit of achievable performance. We therefore train a classifier on a 20 times larger dataset (7.7~billion jets) and find that the observed performance follows the scaling laws prediction. This motivates continued integrated efforts in data generation, model scaling, and inference infrastructure for large-scale models in HEP.
Author: The ATLAS collaboration
Language: English
Subfield: Nuclear and High Energy Physics
Field: Physics and Astronomy
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7137680358
