Untrained Filtering with Trained Focusing for Superior Quantum Architecture Search

Quantum architecture search (QAS) represents a fundamental challenge in quantum machine learning. Unlike previous methods that treat it as a static search process, from a perspective on QAS as an item retrieval task in vast search space, we decompose the search process into dynamic alternating phases of coarse and fine-grained knowledge learning. We propose quantum untrained-explored synergistic trained architecture (QUEST-A),a framework through coarse-grained untrained filtering for rapid search space reduction and fine-grained trained focusing for precise space refinement in progressive QAS. QUEST-A develops an evolutionary mechanism with knowledge accumulation and reuse to enhance multi-level knowledge transfer in architecture searching. Experiments demonstrate QUEST-A's superiority over existing methods: enhancing model expressivity in signal representation, maintaining high performance across varying complexities in image classification, and achieving order-of-magnitude precision improvements in variational quantum eigensolver tasks, providing a transferable methodology for QAS.