AI Research

Democratizing Large-Scale Re-Optimization with LLM-... | AI Research

Key Takeaways

Democratizing Large-Scale Re-Optimization with LLM-Guided Model Patches Optimization models are essential tools for industrial decision-making, but they ofte...
Optimization models developed by operations research (OR) experts are often deployed as decision-support systems in industrial settings.
However, real-world environments are dynamic, with evolving business rules, previously overlooked constraints, and unforeseen perturbations.
In such contexts, end users must rapidly re-optimize models to recover feasible and implementable solutions.
This paper introduces an agentic re-optimization framework in which a large language model (LLM) acts as an OR expert, dynamically supporting end users through natural-language interaction.

Paper AbstractExpand

Optimization models developed by operations research (OR) experts are often deployed as decision-support systems in industrial settings. However, real-world environments are dynamic, with evolving business rules, previously overlooked constraints, and unforeseen perturbations. In such contexts, end users must rapidly re-optimize models to recover feasible and implementable solutions. This paper introduces an agentic re-optimization framework in which a large language model (LLM) acts as an OR expert, dynamically supporting end users through natural-language interaction. The LLM translates user prompts into structured updates of the underlying optimization model, selects suitable re-optimization techniques from an optimization toolbox, and solves the resulting instance to return implementable solutions. The toolbox leverages primal information, including historical solutions, valid inequalities, solver configurations, and metaheuristics, to accelerate re-optimization while preserving solution quality. The proposed framework enables interactive and continuous adaptation of deployed optimization models, reducing dependence on OR experts and improving the sustainability of decision-support systems. Extensive experiments on two complementary large-scale real-world case studies demonstrate the effectiveness and scalability of the proposed framework. The first considers online supply chain re-optimization, where solutions must be generated rapidly while remaining close to the deployed plan, whereas the second focuses on offline university exam scheduling, where solution quality is prioritized over runtime. Results show that the toolbox-driven architecture significantly improves computational efficiency through primal-based and solver-aware re-optimization techniques, while the structured patch-based updates improve interpretability and traceability of model modifications.

Democratizing Large-Scale Re-Optimization with LLM-Guided Model Patches
Optimization models are essential tools for industrial decision-making, but they often become outdated as business rules change or unexpected disruptions occur. Traditionally, updating these models requires constant intervention from operations research (OR) experts, which is slow and costly. This paper introduces "ReOpt-LLM," an agentic framework that uses a large language model (LLM) to act as an OR expert. By allowing end users to communicate changes in natural language, the framework automatically translates these requests into structured updates for the underlying mathematical model, selects the best re-optimization techniques, and generates new, implementable solutions.

How the Framework Works

The system functions as an interactive bridge between the user and the optimization model. When a user describes a change—such as a new constraint or a shift in business logic—the LLM interprets the request and creates a "patch" for the model. Instead of treating the model as a black box, the LLM uses a specialized "toolbox" of OR techniques. This toolbox includes historical solution data, solver configurations, and domain-specific heuristics. By combining these tools with the updated model, the framework can efficiently navigate the complex, tightly coupled structure of large-scale problems to find a new, feasible solution without needing a human expert to manually re-program the system.

Bridging Language and Mathematics

A core innovation of this approach is the use of structured, traceable patches. Rather than simply generating code, the LLM performs explicit, auditable modifications to the optimization model. This ensures that every change is transparent and verifiable. The framework is designed to handle the "reasoning" aspect of optimization—understanding how a local change, like adding a single constraint, might ripple through a massive system to affect global feasibility. By offloading this reasoning to an LLM-orchestrated layer, the framework allows non-experts to maintain complex decision-support systems while ensuring the mathematical rigor of the original model remains intact.

Performance and Scalability

The researchers tested the framework on two large-scale, real-world scenarios: online supply chain re-optimization and offline university exam scheduling. In the supply chain case, the framework successfully generated rapid updates that kept the new plan close to the original, meeting strict time constraints. In the scheduling case, it prioritized high-quality solutions over speed. The results demonstrate that the toolbox-driven architecture significantly improves computational efficiency compared to standard re-solving methods. By leveraging primal information and solver-aware techniques, the framework proves that it can handle the scale and complexity of industrial-grade problems while maintaining the flexibility required for dynamic environments.

Comments (0)

No comments yet

Be the first to share your thoughts!