Generating Robust Portfolios of Optimization Models Using Large Language Models

This paper investigates how large language models (LLMs) can support mathematical optimization modeling by generating portfolios of candidate optimization models from natural language descriptions. The authors propose a framework in which an LLM simultaneously acts as a stochastic model generator and as a reasoning-based evaluator to create diverse, high-quality optimization candidates. The approach provides theoretical guarantees that robust portfolios will contain strong solutions when either the generator or evaluator aligns with human preferences. Empirical evaluations demonstrate strong performance across multiple optimization tasks, supporting human-in-the-loop decision-making for complex planning and resource allocation problems.