Researchers tested whether classically simulated quantum computing components could improve performance or reduce parameters in vision-language models, reporting accuracy gains and parameter reductions that depended on where quantum layers were placed.
Using the LLaVA-OneVision model as a backbone, the team evaluated 4- and 8-qubit quantum circuits at multiple locations, either inserting quantum layers alongside classical components or replacing classical components entirely. The quantum layers were classically simulated, not run on quantum hardware.
On the Visual Spatial Reasoning dataset, inserting a quantum layer after the visual projector reached 80.7% development accuracy, 2.28 percentage points above classical full fine-tuning, according to the researchers. Replacing the final decoder feed-forward sublayer retained 79.22% accuracy while reducing trainable parameters by 13.07 million.
The benefits varied significantly by location. Earlier replacements lowered accuracy, and parameter reductions did not reliably shorten training time or reduce computation, the researchers reported. The team also evaluated robustness under six attack settings to assess potential security benefits, though specific results were not detailed in the abstract.
The study maps where quantum layers help or hurt within vision-language models as quantum computing moves from the noisy intermediate-scale quantum era toward early fault-tolerant systems. Whether these gains generalize to other models or datasets remains open, the researchers noted. Code is available at the repository linked in the preprint.
The preprint was submitted October 20, 2026, to arXiv: https://arxiv.org/abs/2610.28964