What a Large Physics Model is

A large language model is pretrained once, on a vast amount of text, until it has picked up the patterns of language. After that it can be tuned for many jobs, from translation to code, with far less data than training each from scratch. That pretrained base is called a foundation model.

A Large Physics Model, in the sense CERN uses in its October 2026 announcement, is the same idea with physics data in place of words. CERN and Mistral plan to pretrain one model on billions of simulated proton-proton collisions at the LHC, then tune it for reconstruction, particle identification, finding unexpected signals and detector design (CERN's release). Reuters quotes a CERN spokesperson putting it the same way: the idea behind generative AI, trained on collision data rather than text (Reuters).

What "pretraining" means for collisions

Text models learn by predicting a hidden word. Collision models do something similar. Published versions hide one particle in a jet and ask the model to predict it (Masked Particle Modeling), or turn particles into tokens and learn to generate them like sentences (OmniJet-α). Either way, no human has to label the data: the physics in the simulation is the teacher. CERN hasn't said which recipe Flare will use. The collider models that tried these recipes.

How it differs from ChatGPT

• It reads detector signals, not words. Nothing announced says it will have a chat interface.
• Its training data is made by simulation, not collected from the web, so the people training it know the right answer for every example.
• It is judged by physics: how precisely it measures a particle's energy, how often it mistakes one particle for another.

The other meaning

The phrase was in print before CERN used it. A January 2025 paper by 22 physicists and philosophers proposed "Large Physics Models" as AI tools built for physics research in general, including language models that help with theory and analysis, developed and checked by the community (arXiv 2501.05382). CERN's plan is narrower: one model for collider data.

Why build one at all

Reconstruction at the LHC is a chain of specialised algorithms, each written and maintained by its own team. A single model that learns the whole chain could be retrained when the detector changes instead of rewritten, run on GPUs, and share what it learns between tasks. A CMS team showed in February 2026 that one learned model could already match the hand-written particle-flow step (CERN). How reconstruction works.

What could go wrong

A model trained on simulation learns the simulation, including its mistakes. Physicists will want proof that it behaves on real collisions, and that it doesn't hide the very surprises it is meant to find. That is why papers matter more than press releases here. Training on simulation.