What's this about?
Can LLMs design effective algorithms for well-specified operations research problems? The paper tests inventory control, queueing network control and assortment optimisation, and finds the designs come out near-optimal.
What is operations research?
The mathematics of running things well: how much stock to hold, how to route deliveries, how to staff a queue. Its problems have two useful properties for this test — they are precisely specifiable, and there is often a known optimal or near-optimal answer to compare against. That makes "did the model design a good algorithm" an objectively checkable question, unlike most claims about model creativity.
Why does this matter to you?
- Designing an algorithm is a step above implementing one, and this is one of the cleaner tests of it.
- Verifiability is the whole point — this is a domain where you cannot bluff.
- The practical read: for well-specified optimisation problems, the design phase may be a reasonable thing to delegate.