COOPA Agent Automates Operations Research with LLMs and Multi-Solver Routing

Operations Research (OR) offers a powerful framework for high-stakes decision-making, but its application often demands deep domain knowledge, mathematical abstraction skills, and expertise in various solvers. While recent LLM-based systems have attempted to automate parts of this process, they frequently struggle with accuracy on complex problems, lack transparency, and support a limited range of solvers. Researchers propose COOPA (COoperative OPerations Agent), a modular LLM-agent architecture aimed at providing interpretable and scalable OR decision support. COOPA integrates three core components: iterative confidence-based modeling, which generates and self-evaluates multiple formulation candidates before selecting the most confident one; element-level provenance and confidence explanations, linking model components to source text for auditability; and multi-solver routing, dispatching problems to specialized optimizer agents. Across three OR benchmarks, COOPA demonstrated superior performance, achieving the best macro-average accuracy on six out of eight LLM backbones and improving over the strongest baseline by up to 6.7 percentage points. Its design emphasizes transparency and adaptability, making it a significant step forward for automating complex OR challenges.