Finding the unexpected
Flagging unexpected signals is one of the four jobs CERN gives its planned model. Here is what that means and how it is done now.
Most searches for new physics start with a theory: if this particle exists, it should show up like that. Anomaly detection works the other way round. A model learns what ordinary collisions look like and flags the ones that don't fit, without being told what to look for. CERN lists flagging unexpected signals as one of four jobs for its Large Physics Model (CERN).
Already in the trigger
The trigger throws away nearly every collision within microseconds, so anything strange has to be noticed there or it is gone. CERN reported in May 2026 that ATLAS and CMS now run machine-learning models on chips in their first-level triggers, including anomaly detectors that keep unusual collisions no planned search asked for (CERN). That work is part of the Next Generation Triggers project, a five-year programme funded with $48 million from the Eric and Wendy Schmidt Fund (CERN Courier).
Why a foundation model helps
An anomaly detector is only as good as its picture of normal. A model pretrained on billions of collisions should have a richer picture than one trained for a single search. A March 2026 paper by Vinicius Mikuni and Benjamin Nachman tests that idea directly (arXiv 2603.23593), and their OmniLearned models report strong results on an anomaly-detection task in CMS data (arXiv 2510.24066).
The catch with simulation
If the model learns "normal" from simulated collisions, real collisions that differ from the simulation for dull reasons, a miscalibrated detector or a gap in the simulation, can look anomalous too. Telling those apart from new physics is the hard part, and why any flag from Flare would start an investigation rather than end one.