AI and the World's Hard Problems · Physics · arXiv · 2026-10-21
This is SParTA: A modular, efficient and scalable transformer architecture for jet-tagging
Jet tagging at high-energy colliders faces the quadratic memory cost of full self-attention over all jet constituents. SParTA applies quadratic interaction attention only to the N hardest constituents, ordered by descending transverse momentum, and integrates the remaining soft constituents through linear cross-attention, computing pairwise features dynamically from four-momenta. A 300K-parameter model reaches 92.02% AUC and 85.06% accuracy on quark-gluon tagging, matching the 2.14M-parameter Particle Transformer at one-seventh the parameter count and beating IAFormer by 21% in background rejection at 30% signal efficiency, while on boosted top-tagging it outperforms ParT and ParticleNet with 511K parameters.