AI and the World's Hard Problems · Quantum physics & quantum computing · arXiv · 2026-10-22
Can Quantum Layers offer Performance and Security Advantages in Vision-Language Models?
Using LLaVA-OneVision as the backbone, the study evaluates 4- and 8-qubit basic-ring and strongly entangling circuits, either inserted into or replacing classical components at four locations each, and compares performance and computational cost against classical models. On the Visual Spatial Reasoning dataset, inserting a quantum layer after the visual projector reaches 80.7% development accuracy, 2.28 percentage points above classical full fine-tuning, while replacing the final decoder feed-forward sublayer retains 79.22% accuracy with 13.07 million fewer trainable parameters. The benefits depend strongly on placement—earlier replacements lower accuracy and fewer parameters do not reliably shorten training or reduce computation—robustness was also evaluated under six attack settings, and whether these gains generalize remains open.