Abstract for: Automating Group Model Building Sessions with LLM-Based Agents Acting as Domain Experts to Generate Causal Loop Diagrams

One technique to develop a Causal Loop Diagram (CLD) is Group Model Building (GMB). GMB requires multiple subject matter experts (SMEs). These SMEs are not always available; they have limited time, are restricted, and are often not ready to commit to long projects. This work explores whether multi-agent large language model systems can assist in generating preliminary CLDs, accelerating early system structure, and supporting modelers before any engagement with stakeholders. The approach emulates Group Model Building sessions using a multi-agent large language model system. Expert agents retrieve and analyze evidence from domain-specific knowledge bases to propose variables and causal relationships. Moderator agents evaluate reasoning and consolidate results into a causal loop diagram. Python, LangChain, LangGraph, and vector databases are used to support the framework. Preliminary results show that the multi-agent system can generate coherent sets of variables and causal relationships related to a focal variable and context. Early tests indicate that the quality and quantity of documents in the expert knowledge bases significantly influence the relevance of the resulting causal structures. Further evaluation against human GMB outputs is ongoing. Preliminary results suggest that multi-agent LLM systems could support modelers in the early stages of System Dynamics projects by accelerating the generation of causal structures. In practice, this reduces the need for lengthy GMB sessions. In reality, GMB sessions are often hard to implement; this work solves a problem many practitioners have. Even though the CLD can be automatically generated, it does not replace engagement with stakeholders.