The day in conversation for 2026-11-15

discussion

Transcript

The discussion: Kai: The AlphaHits results showing a TrackML score of 0.959 for full reconstruction is impressive, but how robust is that score when you compare it directly against the inputs used by the Particle Transformer baseline? Mira: I’m concerned about the scalability aspect; achieving that metric might rely heavily on specific detector geometries or trigger efficiencies not fully accounted for in that comparison setup, especially since we are trying to move toward truly foundation models. Lev: That leads me to thinking about the top-quark pair reconstruction work, which showed a 35% improvement in invariant mass resolution using ML methods on CMS data; what specific algorithmic change allowed that jump over traditional kinematic fitting techniques? Kai: The paper only mentions the method, but I'm interested in whether that 35% gain is maintained when moving from simulated events to real LHC data, or if that performance drops significantly when the input topology changes due to pileup. Mira: Exactly, and speaking of topology changes, I wonder if we should be focusing more on the branching fraction measurements for chi cJ decays—those first 5 sigma results on strange particle decays suggest a new handle on

More episodes

← Home