The paper, posted on arXiv, addresses the growing use of large language model (LLM) agents within agent-based models (ABMs). The stated premise is that richer, context-sensitive human behavior can improve empirical fidelity in simulations. To support this direction, the authors outline a development framework centered on what they call the "model capacity frontier."
According to the abstract, LLM-supported ABMs introduce an issue or set of issues, but the text cuts off at that point. The framework appears intended to help developers navigate the trade-offs between model capacity and simulation goals. Because the source is incomplete, the specific mechanisms or recommendations of the framework are not available in the provided material.
What is clear is that the paper positions the capacity frontier as a key consideration for anyone building or extending LLM-driven social simulations. The framing suggests a need for deliberate choices about how much complexity to embed in agents, rather than assuming more capacity is always better.