Multi-Agent Framework Auto-Generates QUBO Formulations at 68% Accuracy

QUBO stands for Quadratic Unconstrained Binary Optimization. It expresses combinatorial optimization problems using binary variables and a quadratic objective, with constraints folded into penalty terms. It matters because it maps directly onto quantum, hybrid quantum-classical, and quantum-inspired solvers.

A direct approach asks a model to produce the formulation in one shot. The new framework uses multiple agents plus an iterative self-repair loop that detects and corrects errors, lifting accuracy to 68% on QUBOBench, a 22% gain over the baseline.

The framework and benchmark are open-sourced, so pilots are low-risk for use cases like energy scheduling, supply chain optimization, and fleet routing. But 68% accuracy still requires human review of generated formulations before production use in sensitive applications.

An open benchmark of 100 combinatorial optimization problems spanning 12 application domains, curated from peer-reviewed literature, competitions, and canonical NP-hard problems. It provides a shared yardstick for systems that generate QUBO formulations from natural language.