COMETA Colloquium: Jeff Krupa

Monday 26 Oct 2026, 14:00 → 16:00 Europe/Zurich

This event is organised by the COMETA COST Action, a EU-funded networking initiative that promotes knowledge sharing and cooperation across the theory, experiment, and ML communities, with the aim of improving the measurement and interpretation of multiboson processes at the LHC.

Find recordings of all COMETA Colloquia on our youtube channel @multibosons!

🔔	If you are not a member of COMETA but you would like to receive news and annoucements about this Colloquia series, subscribe to the cometa-colloquia e-Group
	To join COMETA apply at www.cost.eu/actions/CA22130/
💬	For any issues or requests, contact the organizers at cometa-colloquia-org@cern.ch

 

     

 

14:00 → 15:00 How many bits is a jet? Tokenizing LHC data for foundation models

Many foundation models for the LHC begin by turning detector data into discrete tokens, as language models do with words. The tokenization step fixes the information models see, yet it has been the subject of limited study. This talk (based on arXiv:2610.13876) treats tokenization as a rate-distortion problem, probing trade-offs between bits and physics fidelity. Using residual vector quantization on tracks, topoclusters and calorimeter cells from two simulated detectors, we find simple design rules that hold across the set-based modalities. At 4-11× compression, tokens preserve most of the information needed for jet pT regression, particle-flow reconstruction and flavour tagging. I’ll close with some of the open questions that this raises: how tokenized models scale, and whether tokens should be built hierarchically, from particles up to jets and events.

Speaker: Jeffrey Krupa (SLAC)