Abstract
Recent research has introduced large language model (LLM)-based chatbots to help practitioners interpret solutions of optimization models. Existing chatbots primarily support the interpretation of a solution obtained from a single solve. In rolling horizon optimization, however, practitioners need to understand how and why solutions change across successive solves as the planning window advances and changes occur in model data and/or structure. We present an LLM-based agentic framework for this comparison task, works for time-indexed linear programming (LP) and mixed-integer linear programming (MILP) models represented in Pyomo. Four deterministic analytical modules detect model changes, quantify solution differences, assess whether prior decisions remain feasible under the updated model, and and links differences between successive plans to changes in model parameters and structure. LLM agents interpret user questions, select the corresponding module, and synthesize its numerical results into natural-language explanations. The approach is demonstrated on a multi-period workforce scheduling model with approximately 5,300 variables and 6,400 constraints. The framework correctly identifies the parameter changes and constraint families responsible for rescheduling and completes in under 20 seconds. The framework supports optimization models across domains, while the present evaluation is limited to the workforce scheduling problem.