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

summary

The gist

UrQMD-Agent: An LLM-Assisted Framework for Microscopic Tracing of Proton Azimuthal Correlations presents an LLM-assisted framework that connects a generalpurpose large language model to the UrQMD…

This episode discusses

The paper

UrQMD-Agent: An LLM-Assisted Framework for Microscopic Tracing of Proton Azimuthal Correlations · Read on arXiv

Fu-Peng Li, Yige Huang

Key Laboratory of Nuclear Physics and Ion-beam Application (MOE) & Institute of Modern Physics, Fudan University · Shanghai Research Center for Theoretical Nuclear Physics, NSFC and Fudan University · State Key Laboratory of Heavy Ion Science and Technology, Institute of Modern Physics, Chinese Academy of Sciences

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.

Transcript

Introduction to the show: ident: Nuclear Theory. The day in nuclear theory: structure, reactions and the models behind them.

Kai: We're looking at the paper: "UrQMD-Agent: An LLM-Assisted Framework for Microscopic Tracing of Proton Azimuthal Correlations".

Mira: Here's the gist of the paper. UrQMD-Agent: An LLM-Assisted Framework for Microscopic Tracing of Proton Azimuthal Correlations presents an LLM-assisted framework that connects a generalpurpose large language model to the UrQMD transport model through a model-specific Skill and a deterministic scientific tool layer The negative signal is carried mainly by proton pairs containing at least one decay-fed proton This framework enables finalstate observables to be connected to microscopic reaction histories rather than treating UrQMD solely as an event generator Framework Architecture The LLM-assisted analysis framework connects a generalpurpose large language model to a local UrQMD installation through the model-specific UrQMD Skill and a deterministic scientific tool layer Given a scientific question, the LLM consults relevant literature and model documentation and uses the UrQMD Skill to propose appropriate model settings and analysis procedures After the proposed physics configuration and computational plan are confirmed by the user, the LLM invokes the deterministic tools to prepare input files, execute UrQMD simulations, and process the resulting event records All numerical computations are performed by these tools rather than by the LLM itself, supporting the reproducibility and traceability of the analysis 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 Simulation Setup and Observables The study simulated Au+Au collisions at various energies using the UrQMD transport model in cascade mode For the azimuthal-partition analysis, 200,000 events were generated at each energy Protons were selected in the midrapidity region with y < 0.5 and 0.4 < pT < 2 GeV/c The normalized azimuthal-partition fluctuation is defined as ∆σ2 (N, w) = σ2(Nw) − NP0(1 − P0) N(N − 1), P0 = w2/2π Positive values of ∆σ2 indicate fluctuations above the independent-binomial baseline, while negative values indicate suppressed fluctuations Pair Decomposition and History Tracing The exact eventwise representation of the observable is given by ∆σ2 event(w) = 2 Σik K(w, diⱼ) Xi<j An accepted proton is classified as decay-fed (D) if it originates from a recorded decay vertex and non-decay-fed (X) otherwise The total pair contribution satisfies ∆σ2 total(w) = ∆σ2 DD(w) + ∆σ2 DX(w) + ∆σ2 XX(w) The decomposition is evaluated at w = 120◦ Reaction History Analysis and Constraints The recorded reaction histories allow the same DD pair population to be followed backward through the hadronic evolution For distinct-parent DD pairs, the negative pair-kernel correlation is already present in the parent momenta evaluated at their respective formation vertices over 3–27 GeV The tracing shows that for distinct-parent DD pairs, the negative pair-kernel contribution is already present in the parent momenta evaluated at their respective formation vertices This indicates that the final-state negative DD correlation is largely inherited from the angular correlations already present in the parent momenta at their formation vertices The UrQMD calculations exhibit negative azimuthalpartition correlations from 3 to 27 GeV in the 0–5% centrality interval, while the signal is consistent with zero at 39 GeV The results demonstrate the potential of this LLM-assisted framework for reproducible microscopic transport analyses and reaction-history tracing Future work will investigate the contributions of different reaction sources, extend the framework to other transport models for systematic comparisons, and explore applications to experimental physics analyses The numerical data supporting the results of this study, along with the source code and instructions for reproducing the main analyses, are publicly available in Ref.

55: The AI usage statement indicates that ChatGPT and DeepSeek were used to review and improve the readability and language of the manuscript The authors independently verified all derivations, numerical results, and scientific statements and take full responsibility for the content The paper concludes that tracing the DD pairs through recorded reaction history shows that their negative correlation is already present at the formation-vertex level from 3 to 27 GeV The results demonstrate the potential of this LLM-assisted framework for reproducible microscopic transport analyses and reaction-history tracing The UrQMD calculations exhibit negative azimuthalpartition correlations from 3 to 27 GeV in the 0–5% centrality interval, while the signal is consistent with zero at 39 GeV.

Mira: Page 1 --- UrQMD-Agent: An LLM-Assisted Framework for Microscopic Tracing of Proton Azimuthal Correlations Fu-Peng Li and Yige Huang Key Laboratory of Nuclear Physics and Ion-beam Application (MOE) & Institute of Modern Physics, Fudan University, Shanghai 200433, China Shanghai Research Center for Theoretical Nuclear Physics, NSFC and Fudan University, Shanghai 200438, China We present an LLM-assisted framework for microscopic transport studies that connects a generalpurpose 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 √sNN = 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 Page 2 --- II.

Kai: METHOD A.

Paper summary: Kai: So to wrap up on UrQMD-Agent: An LLM-Assisted Framework for Microscopic Tracing of Proton Azimuthal Correlations, the authors show how this framework uses an LLM for planning and interpretation while deterministic tools handle the heavy numerical lifting. They demonstrated that in Au+Au collisions between three point zero and thirty-nine gigaelectronvolts, they resolved a negative signal in proton azimuthal partition fluctuations between three and twenty-seven gigaelectronvolts.

Mira: The implication is that by tracing these decay-fed pairs back through the reaction history, we see that this negative correlation is already present at the formation vertex level from three to twenty-seven gigaelectronvolts. This means a lot of what we observe in final state measurements is inherited from earlier stages of the collision.

Lev: What this means for someone looking at experimental data is that you don't have to assume all those signals are purely final state effects; you can use reaction history tracing to test that assumption against the model.

Kai: And the modular architecture they describe suggests this isn't just a single study; it’s a pathway toward systematic microscopic analyses and comparisons across different transport models using this common LLM-assisted framework.

Mira: The paper sets up a very structured way to connect high level scientific questions to the actual simulation outputs through an agentic workflow. It’s about making microscopic transport studies more accessible by bridging the gap between complex model output and literature guidance.

Conclusion: Kai: So, we’re talking about this paper called UrQMD-Agent by Fu-Peng Li and Yige Huang, which is basically an LLM helping to trace proton correlations in heavy ion collisions using the UrQMD model.

Mira: It’s a framework that links a general language model to the transport code through a skill and some tools that do the actual math, so you use AI for planning but not for running simulations.

Lev: The core idea is moving away from just treating UrQMD as an event generator, using this structure to actually connect final state observables back to reaction histories.

Kai: They showed in Au+Au collisions at different energies that they found a negative signal in the azimuthal partition fluctuations between three and twenty-seven gigaelectronvolts, but it goes back to zero at thirty-nine gigaelectronvolts.

Mira: And the key finding is that this negative signal is carried mainly by proton pairs that have at least one decay-fed proton.

Lev: That’s interesting because reaction history tracing shows that for those specific pairs, the negative contribution was already there in the parent momenta at their formation vertices from three to twenty-seven gigaelectronvolts.

Kai: So, instead of just seeing a negative correlation at the end, you can trace it back and see where it started in the collision evolution.

Mira: It implies that a lot of what we measure in those final state particles isn't purely new physics; it’s inherited from earlier stages of the reaction.

Lev: The architecture they built is modular, meaning you could swap out different general-purpose LLMs or even plug in other transport models with dedicated skills.

Kai: This gives us a repeatable way to do systematic microscopic analyses and compare results across different models using this common AI framework.

Mira: It’s about making the connection between complex model output and scientific planning much more transparent for everyone involved in nuclear physics.

More episodes

← Home