Democratizing Large-Scale Re-Optimization with LLM-Guided Model Patches

Deployed optimization models must respond to new data, disruptions, and business rules long after implementation. We introduce ReOpt-LLM, which combines a natural-language interface with a re-optimization toolbox. Given a validated, typed model change, an LLM selects suitable techniques from saved solutions, heuristic warm starts, and tuned solver configurations. A deterministic programmer normalizes the edits before the validator configures the selected techniques and calls the solver. On an industrial supply-chain model, toolbox orchestration cuts mean solve time from 194.88 to 97.20 seconds and raises mean shipment fulfillment from 86.57\% to 95.65\% relative to re-solving without the selector. On large exam-scheduling MIPs, it eliminates five no-incumbent cases and reduces the median objective difference from 1,441 to zero. Structured patches separately reach 96.7\% final success on the supply-chain study, compared with 0\% for direct code editing. The selector must choose from an allowed list of executable strategies, while the solver remains responsible for feasibility and solution quality.