Speaker
Description
Foundation models are large neural networks pretrained on vast datasets and adapted to many downstream tasks with minimal task-specific training. In high-energy physics, precise Monte Carlo event generators allow the simulation of billions of events, but the enormous space of beyond-Standard-Model scenarios makes training specialized large models for each analysis computationally impractical. A foundation model approach enables a single pretrained model to generalize across many analyses. We present EveNet, a foundation model for event-level collider analysis, pretrained on 500 million simulated Standard Model proton-proton collision events. EveNet employs a shared transformer-based encoder with multiple task-specific decoder heads and is trained using a combination of self-supervised and supervised objectives. We evaluate EveNet on three representative physics analyses and test its performance on real proton-proton collision data from the CMS Open Data. With minimal fine-tuning, EveNet achieves strong performance compared to models trained individually for each task and demonstrates robust generalization to unseen signals. These results show that foundation models can significantly reduce computational and human effort for collider physics analyses at current and future experiments.