Particle-Lund Multimodality in Jet Taggers
Researchers have developed a new multimodal architecture called PLuM that combines particle constituents with Lund plane splittings for improved jet tagging in high-energy physics. This approach processes both types of data jointly using a unified transformer, allowing for cross-attention to determine the added value of structured QCD information. The PLuM model demonstrated significant gains in tagging top-quarks and H to bb decays, suggesting that explicit hierarchical information remains complementary to raw particle representations for certain topologies. AI
IMPACT This research suggests that incorporating physics-specific structured data can enhance the performance of transformer-based models in scientific applications.