Enhancing GPT-3.5's Proficiency in Netlogo Through Few-Shot Prompting and Retrieval-Augmented Generation

Recognizing the limited research on Large Language Models (LLMs) capabilities with low-resource languages, this study evaluates and increases the proficiency of the LLM GPT-3.5 in generating interface and procedural code elements for NetLogo, a multi-agent programming language and modeling environment. To achieve this, we employed “few-shot” prompting and Retrieval-Augmented Generation (RAG) methodologies using two manually created datasets, NetLogoEvalCode and NetLogoEvalInterface. The results demonstrate that GPT-3.5 can generate NetLogo elements and code procedures more effectively when provided with additional examples to learn from, highlighting the potential of LLMs in aiding the development of agent-based models (ABMs). On the other hand, the RAG model obtained a poor performance. We listed possible reasons for this result, which were aligned with RAG's common challenges identified by the state-of-the-art. We propose future research directions for leveraging LLMs for simulation development and instructional purposes in the context of ABMs.