An LLM-based tool called OptiMUS-0.3 can now take a plain-language description of a business problem and produce working solver code — no operations research degree required.
The system, detailed in a paper posted to arXiv, targets mixed integer linear programming: the class of math that decides things like factory scheduling, delivery routing, and hospital staffing. OptiMUS-0.3 breaks the problem into modules, writes and debugs solver code, checks its own output, and iterates. That pipeline beat direct-prompting baselines by 43% on easy instances and 18% on hard ones. On real-world case studies, it solved 28.6% of problems where fine-tuned specialist models solved exactly zero.
The finding that should interest AI engineers is buried in the ablation section: architecture beat raw model capability. A weaker model running through the structured, error-correcting pipeline matched a stronger model prompted naively. That cuts against the current industry reflex of reaching for a bigger model whenever benchmark numbers disappoint.
Most optimization problems in manufacturing and logistics are still solved by hand because translating a messy real-world constraint into a solver's input format is genuinely hard. If a tool like this holds up outside benchmark conditions, it could shrink the gap between "we have a problem" and "we have a solution" for teams without a dedicated operations research function — though the paper is careful to note it is aimed at practitioners who already understand the domain, not a black box for novices.