Large Language Models (LLMs) have drawn the attention of the agent‑based modelling (ABM) community thanks to their increasingly sophisticated natural‑language understanding and generation capabilities. LLMs can summarize, generate, analyze, classify, transcribe, translate text, answer questions, provide explanations, sustain dialogue, extract information from unstructured sources, and perform logical reasoning and problem‑solving, which gives them strong potential throughout the modelling workflow.
This paper first surveys existing applications of LLMs in ABM, then systematically examines possible entry points across the six stages of the modelling cycle—problem formulation, conceptual model construction, implementation & verification, experiment design, result analysis, and documentation & communication. Specifically, LLMs may be used to:
- Quickly scan literature and produce natural‑language descriptions of research questions during problem formulation;
- Assist in abstracting agent behaviours, automatically generate pseudo‑code or UML diagrams during conceptual modelling;
- Offer code snippets, debugging tips, and parameter explanations during implementation;
- Propose experimental setups, evaluation metrics, and sensitivity‑analysis ideas during experiment design;
- Produce textual summaries, visual explanations, and causal interpretations of large simulation outputs during result analysis;
- Automatically draft reports, create presentation slides, and provide multilingual translations during documentation and communication.
The paper also highlights several challenges: LLM outputs may lack verifiability, inherit biases from training data, behave inconsistently on domain‑specific terminology, and often generate “hallucinated” explanations of model internals. Consequently, practical use should combine human review, transparent prompting strategies, and quantitative assessment of uncertainty.
In summary, LLMs can become valuable assistants in ABM workflows, but their deployment must remain critical, with thorough evaluation of reliability and interpretability before widespread adoption. Review