UrQMD-Agent: An LLM-Assisted Framework for Microscopic Tracing of Proton Azimuthal Correlations

Abstract

We present an LLM-assisted framework for microscopic transport studies that connects a general-purpose large language model (LLM) to the UrQMD transport model through a model-specific Skill and a deterministic scientific tool layer. The LLM supports literature-guided scientific planning and interpretation, while numerical calculations, event-history reconstruction, and statistical analyses are performed by deterministic tools. As a demonstration, the framework is applied to proton azimuthal-partition fluctuations in Au+Au collisions at $\sqrt{s_{NN}}=3.0$--39~GeV. In the 0--5\% centrality class, the fluctuations are statistically resolved below zero from 3 to 27~GeV and consistent with zero at 39~GeV. An exact pair-kernel decomposition shows that the negative signal is carried mainly by proton pairs containing at least one decay-fed proton. A random-label null test further distinguishes the observed pair contributions from expectations based on pair abundance alone. Reaction-history tracing shows that, for pairs of two decay-fed protons, the negative pair-kernel contribution is already present in the parent momenta evaluated at their respective formation vertices. The modular architecture is designed to accommodate different general-purpose LLMs and can be extended to other microscopic transport models through dedicated Skills and tool interfaces. This provides a pathway toward systematic microscopic analyses and comparisons across transport models within a common LLM-assisted framework.