Language models as designers of heuristics
Matheuristics and local search are among the fastest-growing keywords in the operations research journals, and recent systems let language models write and evolve heuristic code. Only 2 of the 252 heuristics papers in OR journals carry a language-model keyword. Can generated heuristics beat hand-designed ones on standard routing and scheduling instances?
What prompted this
In the operations research journals, "matheuristic" had 14 papers in July–September 2026 against about 8 expected, and "local search" 15 against 7 (see the operations research dashboard). Since October 2025, 547 papers carry a heuristic, local-search or matheuristic keyword (254 in AI journals, 252 in operations research). Only 13 of them carry a language-model keyword, 2 of those in operations research journals.
Questions
- On standard instances (Solomon, CVRPLIB, Taillard), do heuristics that language models write and refine beat a well-tuned adaptive large neighborhood search?
- Are the generated heuristics new ideas or recombinations of known operators, and can people read them?
- How much computing do they cost compared with what they gain?
Data that could help
Benchmark instances from the optimization category of the dataset ranking, with their published best-known solutions.
Next step
Let a language model evolve destroy and repair operators for an adaptive large neighborhood search on the Solomon instances, and compare the gaps to the best-known solutions.