Swap one Schelling agent for an LLM and the segregation trajectory holds, but runtime climbs from 2 seconds to 8,917

A class of agent-based simulation models is starting to hand individual agents' local judgments to a large language model instead of writing every rule symbolically. In the Schelling segregation model each agent only counts how many neighbors share its group, then decides whether it is happy and whether to move. Previously such models specified individual decisions through explicit, reproducible symbolic rules, which made them inspectable and repeatable under controlled conditions. The change lands in a single method: ordinary agents still compare symbolic types, while the one LLM-enabled agent receives natural-language descriptions of its neighbors and accumulates similar and different counts through tool calls.

Statistical model checking turns the question of whether replacing that rule changes system behavior into a quantitative estimate with a confidence interval, rather than a look at a few simulation curves. The usual practice runs the simulation many times and compares observables, where the number of runs, stopping criteria and comparison style all affect reliability. MultiVeStA reports intervals of width at most 0.1 at 95% confidence and runs point-wise t-tests on the happy-agent ratio trajectories of the original and LLM-extended models; each analysis needed only 20 simulations, indicating little variance under this parametrization.

The LLM may only call two counting tools and cannot move agents, alter the grid or decide happiness directly, so its errors can be attributed separately from global dynamics. Letting the model take part in scheduling or movement policy would mix semantic errors, missed calls and stuck generation with structural changes to the model. Across Qwen3.5 sizes 0.8b, 2b, 4b and 9b, each message was run three times under both blue and red roles; 0.8b got stuck on the first red-role iteration, 2b got stuck in both roles on the multi-neighbor task, and 4b and 9b passed everything.

The demonstrated scope is a Mesa Schelling segregation model with a 20×20 grid, homophily threshold 0.7, density 0.8 and minority percentage 0.2, with a single LLM-enabled agent and transient analysis over steps 1 to 60. Robustness checks ran on an Apple M4 machine with 24GB of unified memory, and the full simulations on an NVIDIA DGX Sparx with 120GB of unified memory. Planned next steps include richer decision tasks, more LLM-enabled agents, other ABMs, and systematic prompt and tool-interface designs.

The authors frame the experiments as preliminary, and runtimes depend on the local machine and serving infrastructure, so they are not a detailed benchmark. The finding of no statistical difference in dynamics is limited to one LLM agent, 60 transient steps and the stated parameters; behavior with more LLM agents, larger populations or richer decision tasks remains untested.