Abstract for: From Text to Map; Why Not Vice Versa, and Beyond? Investigating the Application of LLMs in Interpreting SD Modes' Output Graphs

Rapid advances in LLMs have opened new possibilities for interpreting SD model outputs. This paper investigates whether LLMs can analyze SD output graphs—without access to underlying model structures—and generate coherent textual interpretations, exploring how LLMs may reveal causal relationships and feedback loops embedded in output graphs, and use it in combination with previous AI-based tools that derived SD diagrams from textual data; streamlining an automated cicular modeling process. Two benchmark SD models—a predator-prey system and a SEIRD model—were constructed in Vensim to produce output graphs. Three state-of-the-art LLMs (o1, Claude, and Gemini) were provided with a standardized prompt instructing them to interpret the graphs. Their responses were then evaluated by SD experts using surveys and analyzed via the TOPSIS method to assess accuracy, clarity, comprehensiveness, and creativity in decoding underlying causal structures. LLMs demonstrated a promising ability to interpret SD output graphs. Expert evaluations revealed that among the models tested, Claude produced the most accurate and clear analyses. The generated interpretations successfully identified key system feedback loops and suggested causal relationships, enabling a reverse-engineered depiction of the original model’s structure. Nonetheless, the outputs also highlighted challenges in fully capturing the complexities inherent in dynamic systems. The findings affirm that LLMs can serve as effective adjuncts for interpreting SD model outputs, offering novel avenues for automated analysis and educational support. However, the approach remains exploratory, with current outputs reflecting both potential and limitations. Future research should refine prompting strategies and evaluation methods and explore integrating LLM-driven analysis directly into SD software to enhance practical analytical workflows. Refine the text and prepare error-free textual content for ESL audiences.