What they did

The study covers inventory control, queueing-network control, and assortment optimization. At level 1, each model received a particular problem instance and returned a solution; at level 2, it received only a problem-class description and broad parameter ranges, then produced an algorithm that had to handle unseen evaluation instances.

Human intervention was limited to a single untuned prompt. Models could use a Python sandbox under a fixed compute budget, and their outputs were compared with the best existing methods used for each task.

Key findings

• The strongest model, gpt-5.6-sol, matched or outperformed the best existing method on almost all evaluated instances.
• Competitive performance persisted at level 2, where the generated algorithm was fixed before the model saw any evaluation instance.
• Performance improved sharply across models released less than eight months apart.
• A single untuned model query was sufficient to produce competitive algorithms in all three studied operations-research domains.

Why it matters

The results suggest that frontier LLMs should be treated as serious empirical baselines for algorithm design when objectives, constraints, and parameter ranges can be specified precisely. Their ability to generate reusable solution procedures—not merely solve supplied instances—could reduce the manual effort required to develop specialized heuristics.

Caveats

The evidence is limited to three well-specified problem families and the tested parameter ranges, compute budget, and model lineup. The reported competitiveness does not establish broad reliability across operations research, theoretical optimality guarantees, or robustness when specifications are ambiguous or substantially different from the evaluation setup.