RAG pipeline lifts Qwen 3 optimization modeling accuracy to 72%

Optimization and constraint modeling still trips up large language models—they produce structurally inconsistent or incomplete formulations, especially for combinatorial problems in logistics, healthcare, and supply chain. In a new arXiv paper, researchers Prateek Roy and Akash Singirikonda show that a Retrieval-Augmented Generation (RAG) pipeline can fix this without costly fine-tuning. They synthesized 500 optimization problems using seed descriptions from the Text2Zinc dataset, enriched with LLM-generated professional personas specified in JSON, and paired each with validated Python solver scripts. These problem-solution pairs were stored in a Chroma vector database, and for each inference query, a LangChain agent fetched semantically similar examples as context to guide Qwen 3 30B Instruct.

Results across three benchmarks are striking: accuracy jumped from 40% to 72% on NL4OPT, 40% to 56% on MAMO Easy, and 32% to 56% on MAMO Complex. The synthetically validated examples improved both solution accuracy and structural consistency. The authors argue that combining curated synthetic corpora with retrieval augmentation offers a practical alternative to fine-tuning—letting organizations deploy LLM-based optimization tools in real-world decision-support contexts without model retraining. This approach could democratize AI-driven optimization for operations research, enabling faster, cheaper deployment of specialized modeling assistants.