Title: MIRROR: An Iterative Adaptive Revision and Hierarchical Retrieval Multi-Agent Framework for Optimizing Operations Research Modeling

Abstract:

Traditional Operations Research (OR) modeling is a labor-intensive and delicate process that depends heavily on human expertise, making it poorly equipped to handle emerging or novel scenarios. Although large language models (LLMs) offer the capability to automatically convert natural language descriptions into optimization models, current methodologies present significant drawbacks. These typically involve expensive post-training procedures or utilize multi-agent systems that still suffer from unreliable collaborative error correction and a lack of task-specific retrieval capabilities, frequently resulting in inaccurate outputs.

To address these challenges, we introduce MIRROR, an end-to-end, fine-tuning-free multi-agent framework designed to directly transform natural language optimization problems into both mathematical models and solver code. MIRROR incorporates two fundamental mechanisms: first, execution-driven iterative adaptive revision, which enables automatic error correction; and second, hierarchical retrieval, which accesses relevant modeling and coding examples from a meticulously curated library of exemplars.

Our experimental results demonstrate that MIRROR surpasses existing approaches on standard OR benchmarks, achieving particularly strong performance on complex industrial datasets, including IndustryOR and Mamo-ComplexLP. By merging systematic error correction with precise external knowledge infusion, MIRROR offers a robust and efficient OR modeling solution for non-expert users, effectively mitigating the inherent limitations that general-purpose LLMs exhibit in specialized optimization tasks.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC