COOPA: A Modular LLM Agent Architecture for Operations Research Problems

A modular agent architecture called COOPA has been proposed to improve how large language models handle Operations Research problems, combining self-evaluation, source traceability, and multi-solver routing to address accuracy and transparency gaps in existing systems. The system, detailed in a preprint posted to the arXiv repository on June 25, 2026, is designed to tackle a core challenge in Operations Research (OR): translating complex, real-world decision problems into formal mathematical models that solvers can process [1][2]. The authors note that effective OR modeling demands substantial domain knowledge, mathematical abstraction, and solver expertise, limitations that have hindered earlier LLM-based automation attempts [2]. COOPA, short for COoperative OPerations Agent, introduces three integrated components to address these shortcomings [1][2]. First, an iterative confidence-based modeling module generates multiple candidate formulations for a given problem, evaluates them across several modeling dimensions, and selects the best one using a max-min confidence criterion [2]. Second, the architecture produces element-level provenance and confidence explanations, linking each variable, parameter, constraint, and objective directly to quoted source text to create an audit trail for human verification [1][2]. Third, a multi-solver router dispatches problems to specialized optimizer agents depending on the problem class [2]. The researchers evaluated COOPA across three OR benchmarks, eight LLM backbones, and four baselines under identical conditions [1][2]. The system achieved the highest macro-average accuracy on six of the eight backbones tested and outperformed the strongest baseline by up to 6.7 percentage points [1][2]. A within-system ablation study isolated the specific contribution of the iterative confidence-based modeling component, while additional case studies illustrated the practical value of source traceability and multi-solver dispatch [2]. The preprint appears on arXiv, an open-access repository that has hosted scientific e-prints since 1991 and now receives roughly 24,000 submissions per month [6]. The work has not yet undergone formal peer review, consistent with the platform's role in providing rapid dissemination of research findings before journal publication [6].