Multi-agent framework converts natural-language problems into QUBO formulations with 68% accuracy

Researchers propose an end-to-end multi-agent framework that automatically generates Quadratic Unconstrained Binary Optimization (QUBO) formulations from natural-language problem descriptions. The system is designed to identify binary variables, constraints, objective functions, penalty terms, and suitable penalty weights without requiring substantial domain expertise.

To evaluate the framework, the authors introduce QUBOBench, a benchmark containing 100 combinatorial optimization problems spanning 12 application domains. These problems are curated from peer-reviewed literature, competitions, and canonical NP-hard problem sets.

Experimental results indicate the framework achieves 68% accuracy on QUBOBench, outperforming a direct single-call baseline by 22 percentage points. The authors identify iterative self-repair as the most important component contributing to the improved performance.

The data and code for the framework and benchmark are open-sourced at a public repository linked in the paper.