LLMs Enhance Operations Research with Uncertainty-Aware Modeling

A new inference framework has been developed to enhance the reliability of Large Language Models (LLMs) when applied to operations research (OR) tasks. The challenge with current LLM approaches is their tendency to make myopic decisions during formulation, which can lead to errors that propagate throughout the optimization model. The proposed method operates without requiring additional training, focusing instead on evaluating intermediate steps in the model generation process. It employs short lookahead simulations to quantify the predictive uncertainty of partial formulations, allowing the system to dynamically select steps that are more likely to lead to a globally consistent and valid optimization model. Evaluations across several OR benchmarks, including NL4OPT, MAMO, and IndustryOR, demonstrate that this uncertainty-aware framework consistently outperforms standard LLM baselines. This establishes a more efficient and reliable paradigm for generating OR mathematical formulations.