AI Can Now Invent Its Own Problem-Solving Recipes
Cheaper delivery routes and smarter schedules may be coming — without hiring a math genius.
A new tutorial paper looks at how large language models are increasingly being used as variation operators in metaheuristics — generating or modifying candidate solutions, heuristics, or programs inside iterative search loops. This shift reframes variation as a model call conditioned on different types of information. The authors introduce an operator-level framework with two descriptors: the type of prompt-conditioning information at variation time (Numeric, Symbolic, Linguistic), and artifact persistence — what survives the model call (Transient, Amortized, Transfer). The tutorial shows how to classify, build, and select these operators, through a worked build template, a method survey, an evidence table, and a cost-aware decision guide.
- Planning software used to rely on rules written by human engineers; now AI models can invent those rules on the fly
- The paper sorts these AI helpers into two simple categories: what you tell the AI, and what useful thing comes back
- It includes a cost guide, because every AI call costs money — the trick is knowing when it's actually worth it
Why It Matters
Better problem-solving means cheaper shipping, tighter schedules and less waste — if the AI calls stay affordable.