Large Language Model-Powered Supply Chain Optimization: A Multi-Agent Framework With Reasoning for Modeling and Solver Interpretation

Abstract

This thesis presents a novel framework, GPTOpt, for integrating Large Language Models (LLMs) into supply chain optimization workflows. The framework addresses two major challenges: translating natural language problem descriptions into executable optimization models, and interpreting solver outputs in a decision-support context. To overcome limitations in arithmetic and reasoning accuracy, this research leverages advanced prompting strategy and incorporates Agentic AI principles. The multi-agent reasoning setup is utilized to enhance the reliability of LLM-generated solutions by enabling structured step-by-step problem solving and output verification. We conducted a proof-of-concept demonstration using GPT-4 to guide the model in solving and interpreting three transportation problems of increasing complexity. In addition to these structured examples, GPTOpt was also tested on large-scale, real-world problem instances to validate its scalability and robustness. Across both settings, the framework demonstrated its ability to translate informal natural language descriptions into Gurobi-compatible models and provide interpretable decision support. The results showed that while LLMs perform comparably to traditional solvers in simpler scenarios, they can also explore broader solution spaces and discover more optimized solutions for complex problems. The proposed GPTOpt framework serves as a practical guide for researchers and practitioners in supply chain management (SCM) and operations research (OR), demonstrating how LLMs can automate, enhance, and interpret optimization tasks with greater accessibility and flexibility.

Date of Award	2025
Original language	American English
Awarding Institution	HBKU College of Science and Engineering