Researchers show AI can learn to write optimization code by running it, not just studying examples

Instead of training AI models on thousands of hand-labeled examples of how to write optimization code, researchers built a system that generates code, runs it, and learns from whether the output actually works. This removes the expensive bottleneck of needing human experts to annotate training data, and it means the same AI can be retrained to work with different solvers by just swapping out the execution environment rather than rebuilding everything from scratch.

For years, the constraint on automating optimization modeling was the cost and brittleness of training data — you needed expensive process supervision or you got models that only worked for one specific solver and overfitted to its quirks. This approach treats execution as the only signal that matters: did the code run without errors and produce a valid solution? That's radically simpler and cheaper. It means companies could potentially retrain these models themselves on their own solvers without restarting from scratch, and the same model could adapt to multiple optimization backends. The practical question is whether this actually reduces the time and cost of building decision-automation tools in production, or whether the gains flatten out once you hit the real-world messiness of constraint modeling.