Back to AI Research

AI Research

OptiAgent: End-to-End Optimization Modeling via Mul... | AI Research

Key Takeaways

  • OptiAgent is a multi-agent framework designed to bridge the gap between natural language descriptions of Operations Research problems and the precise, execut...
  • We propose OptiAgent, a multi-agent framework that, given a natural language description of an Operations Research problem, is able to output a solver-ready mathematical formulation as well as executable code.
  • Our architecture prioritizes the mathematical modeling step, where dedicated agents extract structures, such as decision variables and constraints, enabling iterative self-correction.
  • Alongside accuracy, our modular design improves the process of solving optimization problems by improving transparency, as each agent exposes its reasoning and feedback, making the full modeling process auditable.
  • Our framework achieves state-of-the-art performance on 3 out of 4 benchmarks across LP, MILP, and Nonlinear Programming tasks, while remaining highly competitive on the remaining dataset.
Paper AbstractExpand

We propose OptiAgent, a multi-agent framework that, given a natural language description of an Operations Research problem, is able to output a solver-ready mathematical formulation as well as executable code. Our architecture prioritizes the mathematical modeling step, where dedicated agents extract structures, such as decision variables and constraints, enabling iterative self-correction. We introduce a novel multi-loop validation architecture with four specialized feedback mechanisms, each targeting a distinct failure mode such as misinterpretation, structural defects, mathematical inconsistencies, validation failures, and code errors. Alongside accuracy, our modular design improves the process of solving optimization problems by improving transparency, as each agent exposes its reasoning and feedback, making the full modeling process auditable. Our framework achieves state-of-the-art performance on 3 out of 4 benchmarks across LP, MILP, and Nonlinear Programming tasks, while remaining highly competitive on the remaining dataset.

OptiAgent is a multi-agent framework designed to bridge the gap between natural language descriptions of Operations Research problems and the precise, executable code required by mathematical solvers. By moving away from "black-box" AI models, this system uses a specialized team of agents to break down complex optimization tasks into manageable steps—such as identifying variables, defining constraints, and writing code—while maintaining a transparent, auditable process that allows for automated error correction.

A Specialized Multi-Agent Workflow

Unlike systems that treat optimization as a single, end-to-end task, OptiAgent organizes the modeling process into six distinct roles. These include an Interpretation agent to understand the problem, a Formulation agent to build the mathematical model, an Analysis agent to categorize the problem type, a Validation agent to check for errors, a Code Generation agent to write the Python script, and a Solver agent to compute the final result. This modular design ensures that each stage of the process is handled by a dedicated component, mirroring the workflow of human experts.

Automated Error Correction

A key innovation in OptiAgent is its multi-loop validation architecture. If an agent detects a mistake—such as a missing constraint, a mathematical inconsistency, or a coding error—the system automatically triggers a feedback loop. This routes the task back to the appropriate upstream agent with specific instructions on what needs to be fixed. For example, if the validation agent finds a logic error in the constraints, it sends the model back to the formulation agent for revision. To ensure efficiency, these loops are capped at three iterations, preventing the system from getting stuck in infinite cycles while still ensuring high-quality, accurate outputs.

Performance and Transparency

OptiAgent was tested across four major benchmarks covering Linear Programming, Mixed-Integer Linear Programming, and Nonlinear Programming. It achieved state-of-the-art performance on three of these four datasets, showing particularly strong results on complex, modern problems where other models often struggle. Beyond raw accuracy, the framework provides a significant advantage in transparency: because every agent exposes its reasoning and feedback, users can inspect exactly how a model was built, what corrections were made, and why specific decisions were reached. This makes the system a reliable tool for both non-experts needing assistance and experienced professionals requiring an auditable modeling process.

Comments (0)

No comments yet

Be the first to share your thoughts!