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

Researchers present an LLM-powered framework that enables non-expert end users to re-optimize deployed decision-support systems through natural language interaction, eliminating dependency on operations research specialists. The system combines language models with an optimization toolbox to dynamically adapt models to changing business conditions while maintaining solution quality and interpretability.

This research addresses a critical operational bottleneck in enterprise decision-making systems. When optimization models deployed in supply chains, scheduling, and resource allocation encounter real-world changes—market shifts, regulatory updates, or unexpected constraints—organizations traditionally require specialized operations research experts to manually reconfigure them. This creates friction, delays, and organizational bottlenecks. The proposed framework democratizes this capability by using large language models as intermediaries that translate business requirements into structured model modifications and select appropriate re-optimization techniques automatically.

The approach represents a broader trend of AI systems augmenting specialist expertise rather than replacing it. By enabling business users to interact with complex optimization problems through natural language, organizations can iterate faster and adapt to dynamic environments without waiting for expert intervention. The two case studies—supply chain re-optimization requiring rapid decisions and university exam scheduling prioritizing solution quality—demonstrate the framework's versatility across different optimization objectives.

For enterprises relying on optimization-driven decision support, this technology has immediate practical value. Faster adaptation cycles mean competitive advantages in responsive supply chains and operational efficiency gains in scheduling. The architecture's emphasis on interpretability through "patch-based updates" addresses growing concerns about AI system transparency in critical business functions.

The scalability and effectiveness demonstrated on large real-world problems suggest this framework could become a standard layer in enterprise optimization stacks. Future development likely focuses on integrating additional constraint types, expanding the optimization toolbox, and refining how LLMs reason about tradeoffs between solution quality and computational speed across diverse problem domains.

• →LLM-guided framework enables non-expert business users to re-optimize complex models through natural language without OR specialists
• →Toolbox-driven architecture uses primal information and solver configurations to accelerate re-optimization while preserving solution quality
• →Structured patch-based model updates improve interpretability and traceability of modifications made to deployed systems
• →Demonstrated effectiveness on supply chain and exam scheduling problems shows versatility across different optimization objectives
• →Framework reduces organizational dependency on specialist expertise while improving sustainability of decision-support systems