Speaker
Description
The CMS Phase-2 Level-1 Trigger (L1T) Scouting program introduces real-time software reconstruction at the full 40 MHz rate, enabling physics analyses directly at trigger level. One of the most promising applications is the reconstruction of low-transverse-momentum (soft) hadronic tau leptons, which are essential for searches for low-mass scalars ϕ → ττ but are poorly reconstructed by existing L1 algorithms optimised for higher-pT objects.
This work presents a fully PF-based tau reconstruction workflow designed specifically for the L1T Scouting environment. The chain integrates a GPU-friendly density-based clustering algorithm for PF candidates and introduces a transformer-based neural network for tau identification, capable of exploiting multi-particle correlations under extreme pileup (⟨PU⟩ = 200) and respecting the real-time constraints of the scouting farm.
The framework is evaluated using simulated low-mass ϕ → ττ signals and Minimum Bias backgrounds. Preliminary results show a significant improvement in soft-tau efficiency (down to 5 GeV) while maintaining an acceptable fake rate at 40 MHz. The approach substantially enhances acceptance for low-mass di-tau resonances and demonstrates that advanced PF and ML methods can be deployed effectively within the Phase-2 L1T Scouting paradigm.