Converting natural language descriptions of optimization or satisfiability problems into valid MiniZinc models is challenging because it requires both logical reasoning and constraint programming expertise. This paper introduces Gala, a framework that uses a global agent approach to address this challenge. Gala decomposes the modeling task into multiple specialized large-scale language model (LLM) agents by global constraint type. Each agent is dedicated to detecting a specific type of global constraint and generating code, and a final assembler agent integrates these constraint snippets into a complete MiniZinc model. By decomposing the problem into smaller, well-defined subtasks, each LLM handles a simpler reasoning task, reducing overall complexity. We conducted initial experiments using multiple LLMs, demonstrating better performance than baselines such as one-time prompts and chain-of-thought prompts.
