LLMs Can Design Near-Optimal OR Algorithms

TL;DR - This paper evaluates whether LLMs can design algorithms for well-specified operations research problems. Its strongest tested model, GPT-5.6 Sol, matched or exceeded leading specialized methods on nearly all evaluated instances, suggesting frontier LLMs are becoming credible baselines for OR algorithm design.

• Covers inventory control, queueing network control, and assortment optimization.
• Tests both instance-specific solutions and general algorithms created before evaluation instances are revealed.
• Uses a single untuned prompt, minimal human input, and a Python sandbox with a fixed compute budget.
• Reports sharp performance gains across models released within an eight-month period.