FERMIACC: Reasoning models for particle physics

Abstract: I will present a prototype architecture for using large-language models for reasoning tasks in particle physics. For illustration, I will use collider physics anomalies as the application. The architecture is a frontier reasoning model scaffolded with agents which have access to deterministic particle physics tools. The FERMIACC ingests an LHC experimental paper with an excess and develops new physics hypotheses and parameter estimates for explaining this excess. It also implements the analysis and simplified statistics. It autonomously runs collider simulation and the analysis to deterministically check its hypothesis. I will discuss some lessons we have learnt in developing this prototype and some future plans.