COOPA: A Modular LLM Agent Architecture for Operations Research Problems
COOPA is a modular LLM-agent architecture that enhances Operations Research decision support through iterative confidence-based modeling, element-level provenance explanations, and multi-solver routing, achieving state-of-the-art accuracy and interpretability across diverse benchmarks.
Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
Imagine you are trying to build a complex machine, like a custom car, based on a vague description written by a customer. You need to figure out the exact parts, the engine specs, and the safety rules before you can even start building.
Operations Research (OR) is the field of math used to make these kinds of high-stakes decisions (like how to route delivery trucks or schedule factory workers). The problem is that doing this manually requires a PhD-level expert.
Recently, we've tried using Large Language Models (LLMs)—the same kind of AI that writes poems or code—to do this math for us. But the paper argues that current AI attempts are like hiring a very confident but slightly confused intern: they often get the basic math wrong, they don't explain why they made a mistake, and they can only use one specific type of tool.
Enter COOPA (COoperative OPerations Agent). Think of COOPA not as a single worker, but as a specialized construction crew designed to build these math models correctly. Here is how it works, using three main tools:
1. The "Second Opinion" Team (Iterative Confidence-Based Modeling)
Current AI usually tries to solve a problem once, writes the code, and hopes it works. If the math is wrong, the code might still run, but it will give you the wrong answer.
COOPA is different. Imagine you ask three different architects to draw a blueprint for your car.
- Step A: They all draw their own versions.
- Step B: A "Safety Inspector" (the AI itself) reviews each blueprint. It doesn't just say "Good" or "Bad." It gives a confidence score (0–100) for four specific parts: the parts list (parameters), the moving pieces (variables), the goal (objective), and the safety rules (constraints).
- Step C: The system picks the blueprint that has the highest "worst-case" score.
The Analogy: If Architect A has a perfect engine but a terrible safety rule (score: 90, 90, 90, 30), and Architect B has a slightly weaker engine but perfect safety (score: 85, 85, 85, 95), COOPA picks Architect B. It knows that one weak link breaks the whole chain. This process repeats a few times, allowing the AI to fix its own mistakes before it starts building the machine.
2. The "Paper Trail" (Source Traceability)
When current AI gives you an answer, it's often a "black box." You get a number, but you don't know where it came from. If you ask, "Why did you assign Worker A to Station 5?", the AI might just say, "Because I said so."
COOPA acts like a meticulous forensic accountant.
- Every time it writes down a variable or a rule, it attaches a "sticky note" that quotes the exact sentence from the original problem description that justified it.
- The Benefit: If the final answer is wrong, a human expert can instantly look at the "sticky notes," see that the AI misread a sentence, and fix it. It turns a mystery into a transparent audit trail.
3. The "Specialized Tool Shed" (Multi-Solver Dispatch)
Imagine you have a toolbox. If you need to hammer a nail, you use a hammer. If you need to cut wood, you use a saw.
- Current AI systems often try to use a Swiss Army Knife for everything. They force every math problem into one specific format (like a linear equation), even if the problem is actually a complex puzzle that needs a different approach.
- COOPA has a Tool Shed with four different specialized experts:
- The Mathematician: Handles standard algebra and linear equations.
- The Combinatorics Expert: Handles routing and scheduling puzzles (like the Traveling Salesman problem).
- The Metaheuristic Expert: Handles messy, non-linear problems that need evolutionary algorithms (like finding the best shape for a car part).
- The Generalist: Handles custom simulations.
COOPA looks at the problem, identifies what kind of "puzzle" it is, and hands it to the right specialist. This means it doesn't force a square peg into a round hole.
The Results: Does it work?
The authors tested COOPA against other AI systems using three different sets of math problems and eight different "brain" models (from GPT-5 to open-source models).
- The Score: COOPA won the most often, achieving the highest average accuracy on 6 out of the 8 different AI brains tested.
- The Improvement: On the strongest models, it improved accuracy by up to 6.7 percentage points.
- The Secret Sauce: The paper found that the "Second Opinion Team" (the iterative process of generating and selecting the best blueprint) was the main reason for the success. Without it, COOPA performed just as well as the other systems; with it, it became the clear winner.
The Catch
The paper admits that COOPA is slower and more expensive to run than the other methods. Because it asks the AI to think, re-think, and evaluate its own work three times, it uses more computing power. However, the authors argue that for high-stakes decisions (like supply chains or public policy), paying a little extra to get the math right is worth it.
In short: COOPA is a system that makes AI better at math by making it think twice, show its work, and use the right tool for the job.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.