Total: 1
Quantum Transformer models represent a significant research direction in quantum machine learning (QML), leveraging the parallelism and entanglement properties of quantum computing to overcome the computational complexity and expressive limitations of classical Transformers. Parameterized quantum circuit (PQC)-based Transformer models are the primary focus of current research, employing PQCs to achieve varying degrees of quantumization, including strategies such as QKV-only Quantum mapping, Quantum Pairwise Attention, Quantum Global Attention, and Quantum-Assisted Acceleration. These approaches are well-suited to Noisy Intermediate-Scale Quantum (NISQ) devices, demonstrating potential in small-scale tasks to reduce complexity or enhance performance. The strength of PQC-based methods lies in their compatibility with existing quantum hardware, positioning them as the main pathway toward the practical implementation of quantum Transformers. However, these methods face challenges such as limited scalability, the absence of standardized testing benchmarks, and the "barren plateau" problem during training. As a complementary approach, Quantum Linear Algebra (QLA)-based Transformer models rely on future fault-tolerant quantum computing, utilizing techniques like block-encoding and Quantum Singular Value Transformation (QSVT) to achieve efficient matrix operations and theoretically significant complexity reductions, though they remain in the theoretical exploration stage. Future research should prioritize optimizing PQC-based hybrid architectures and quantum global attention models, establishing unified evaluation frameworks, and addressing training difficulties, while also exploring hybrid PQC-QLA approaches to advance the development of quantum Transformers.