An end-to-end multi-agent framework was proposed to translate natural-language problem descriptions into QUBO formulations. The framework supports structured or unstructured test cases. Evaluation was conducted using QUBOBench, a benchmark consisting of 100 combinatorial optimization problems across 12 application domains. The framework achieved 68% accuracy on the benchmark, surpassing a direct single-call baseline by 22%. Iterative self-repair was identified as the key component contributing to improved performance.
Source: https://arxiv.org/abs/2609.10629