Operations research (OR) problems in the real world rarely arrive as neat mathematical formulations. Clients describe messy, context-rich situations, and experts translate them into solvable models. A new paper on arXiv introduces OR-License, a benchmark designed to test whether large language model (LLM) agents can skip that expert step and work directly from original client materials.
The authors note that existing OR benchmarks typically begin with self-contained problem statements, already stripped of ambiguity and reformulated for solution. OR-License instead evaluates agents on raw client inputs, asking whether they can deliver expert-level solutions without prior expert reformulation. This shifts the challenge from solving a given model to understanding and modeling the problem itself.
The abstract does not yet report results, but the benchmark's design highlights an important gap in current evaluations. If LLM agents can handle the full pipeline from client description to solution, they could broaden access to OR expertise. If not, the benchmark may help identify where human intervention remains necessary.