Quantum circuits no longer keep only the winner: fusing five models picked by cognitive diversity beats the best single model on all four datasets

Quantum machine learning pipelines usually keep the single best circuit found by architecture search and discard the rest; this work keeps those candidates as independent scoring systems and fuses them with a diversity-aware selection and combination rule, and the fused model beats the strongest single model in the pool on all four binary classification benchmarks. Prior quantum architecture search, including differentiable quantum architecture search, ultimately returns one chosen circuit and discards every other candidate evaluated during the search; ensemble work in quantum machine learning has mostly relied on simple averaging or ad hoc weighting of output probabilities without a formal diversity criterion. On Heart disease, Ionosphere, Pima diabetes and Spambase, the fused models improve over the best single model from AUC 0.9106 to 0.9444, accuracy 0.9434 to 0.9811, 0.7845 to 0.8448, and 0.9349 to 0.9407, with the 6.03% gain on Pima diabetes the largest of the four.

The approach treats every quantum circuit or classical model as an independent scoring system, first selects a subset of models that differ most from one another, then combines their predictions either in score space or in rank space. Cognitive diversity is a measure from combinatorial fusion analysis that compares the distance between two systems' rank-score functions, that is, how differently two models distribute discrimination across the ranked population after sorting items by score, rather than how correlated their outputs are or how often they disagree. The pool holds 29 scoring systems: 12 circuit configurations from an extended differentiable quantum architecture search space, 8 curated quantum circuits and 9 classical models; selection first removes models with validation AUROC at or below 0.5, sorts pairwise cognitive diversity and applies the Kneedle algorithm to find the knee, then clamps the candidate set to five models, yielding a different quantum-classical subset on each dataset.

The reported scope is offline experiments on four binary classification benchmarks covering medical diagnosis, signal processing and cybersecurity, with 270 to 4601 samples and 8 to 57 features. It is aimed at researchers and engineers working on quantum machine learning ensembles, model fusion and diversity measures. The authors propose extending the current single-layer fusion pipeline into multi-layer combinatorial fusion with an expansion-reduction algorithm, and evaluating the framework against quantum-induced noise.

The authors scope the work to a single-layer combinatorial fusion pipeline and explicitly list multi-layer combinatorial fusion and robustness under quantum noise as future work, so whether the diversity patterns and fusion gains hold in noisy environments is untested. The selected model subsets differ across the four datasets and no single model is picked for all of them, so the best combination depends on the dataset. On Spambase the gain is only 0.58%, which the authors attribute to the models' rank-score functions being very similar to one another on that dataset.