LLM-Driven Framework Accelerates Large-Scale Optimization Model Development

Global: LLM-Driven Framework Accelerates Large-Scale Optimization Model Development

Researchers have introduced a new workflow construction framework called LEAN-LLM-OPT, which automates the formulation of large‑scale optimization models. The approach was detailed in a paper posted to arXiv in January 2026 and aims to reduce the labor‑intensive steps traditionally required for building such models. By accepting a problem description and related datasets, the system orchestrates multiple large language model (LLM) agents to generate a complete optimization formulation.

Framework Architecture

LEAN-LLM-OPT employs a two‑stage upstream process in which separate LLM agents dynamically design a step‑by‑step workflow for modeling problems similar to the query. A downstream LLM agent then follows this workflow to produce the final optimization model, integrating both planning and data‑handling tasks.

Workflow Advantages

The modular workflow standardizes the modeling process into a series of structured sub‑tasks, delegating routine data manipulation to auxiliary tools. This design lessens the planning burden on the LLM, allowing it to focus on interpreting unstructured components of the problem description.

Performance Evaluation

Extensive simulations reported in the paper indicate that LEAN-LLM-OPT, instantiated with GPT‑4.1 and the open‑source gpt‑oss‑20B, delivers strong results on large‑scale optimization modeling benchmarks and remains competitive with existing state‑of‑the‑art methods.

Industry Application

A case study involving Singapore Airlines’ choice‑based revenue management demonstrated the framework’s practical value, achieving leading performance across a variety of operational scenarios.

Benchmark Contributions

The authors also introduced two new benchmark suites, Large‑Scale‑OR and Air‑NRM, which constitute the first comprehensive collections for evaluating large‑scale optimization auto‑formulation techniques.

Open Access Resources

The full codebase and datasets supporting the research are publicly available on GitHub at https://github.com/CoraLiang01/lean-llm-opt.

This report is based on information from arXiv, licensed under Academic Preprint / Open Access. Based on the abstract of the research paper. Full text available via ArXiv.