Abstract
<title>Abstract</title> <p>Open-pit mine production scheduling (OPMPS) is an NP-hard optimization problem traditionally addressed through Mixed Integer Linear Programming (MILP), whose correct formulation requires mathematical expertise that many mine planning practitioners lack. Large Language Models (LLMs) can automate this procedure; however, their reliability frequently diminishes with increasing problem complexity. This paper proposes a staged context-enrichment framework for LLM-based generation of Gurobi-compatible MILP code. The framework uses four prompt configurations of increasing informational density. Starting from a structured problem description (Basic Prompt), then the description is augmented with a greedy simulation log, inclusion of simulator source code and finally with an inline-annotated version of that code. We evaluate four LLMs, ChatGPT, Gemini, DeepSeek, and Copilot, across all configurations against a verified benchmark MILP. Formulation quality is measured using net present value (NPV) accuracy and constraint recovery across five categories. A central finding is that whole-objective feasibility alone is an unreliable indicator of formulation correctness, since many nominally feasible formulations are silent failures that never enforce one or more constraint groups. Once these cases are identified and excluded, simulator code emerges as the most informative element for whole-objective accuracy, whereas annotated code achieves the highest feasibility rates and the strongest average constraint-level recovery.</p>