Fancier quantum architectures do not automatically outperform simpler ones — and on standard tabular benchmarks, they often underperform them.
Researchers compared four families of variational quantum circuits (VQCs) — a type of algorithm designed to run on near-term quantum hardware — across five regression and classification tasks. The headline result: basic fully-connected VQCs reached 90–96% of the accuracy of attention-based quantum transformers while using 40–50% fewer parameters. On the Boston Housing benchmark, the simple FC-VQC averaged an R² of 0.829 against 0.753 for a comparable classical neural network. Explicit quantum self-attention, the mechanism that makes transformer architectures expensive, added marginal accuracy gains while meaningfully inflating parameter counts. The study also found that circuit expressibility stops improving at a depth of roughly three layers — meaning the quantum equivalent of "just add more layers" hits a ceiling fast.
The findings matter because quantum machine learning carries significant hype about transformers specifically, given their dominance in classical AI. This paper is a useful corrective: the architectural assumptions that made transformers powerful on language and images do not transfer cleanly to quantum circuits operating on tabular data. It also gives hardware-focused teams a practical signal — shallow, simple circuits are not a compromise; they are often the right call.
The noise results add a wrinkle worth watching: the fully quantum transformer degraded gracefully under simulated hardware noise, while the hybrid quantum-classical version collapsed. That gap could matter more as real quantum hardware becomes the deployment target rather than a simulator.