The authors asked an LLM to solve three well-specified operations research problems: inventory control, queueing network control, and assortment optimization. In one protocol the model received one instance and returned a solution; in another it received only a problem class description and parameter ranges, and returned a fixed algorithm. The strongest model tested matched or outperformed the best existing method on almost all instances, even when the algorithm was fixed before seeing evaluation instances.
The result matters because no prompt tuning, human hints, or specialized knowledge were used; the model only had a problem description and a Python sandbox. Performance also improved sharply across models released less than eight months apart, so this capability appears to be improving quickly. For well-specified OR problems, a cheap one-query LLM baseline may now be a serious competitor to specialist algorithms.