Zero-shot Quantum Neural Architecture Search

A new framework called MZeQAS uses a zero-shot surrogate model and Monte Carlo Tree Search to discover high-performing quantum circuit architectures without the high computational cost of training every candidate, according to research posted on arXiv [1]. Variational Quantum Algorithms (VQAs) are a leading method for exploiting near-term quantum hardware, but their practical deployment is hindered by the difficulty of designing circuit architectures that balance expressivity, trainability, and hardware constraints [2]. Existing evolutionary-based quantum neural architecture search methods address these challenges but suffer from high computational costs because they require repeated training of candidate circuits [2]. The MZeQAS framework, proposed by Tung Dao, tackles this bottleneck by identifying a setting in which the Gram matrix of the Quantum Neural Tangent Kernel converges [1]. This observation underpins a zero-shot surrogate model that estimates candidate performance without full training, significantly accelerating the search process [1]. MZeQAS integrates this proxy-based performance estimation with Monte Carlo Tree Search (MCTS) exploration to efficiently discover architectures [2]. Experimental results demonstrate that MZeQAS outperforms existing approaches in both search efficiency and solution quality, providing a scalable method for advancing VQA deployment on noisy intermediate-scale quantum devices [2]. The paper was submitted on 12 May 2026 and revised on 7 June 2026 [1]. The work builds on a growing body of research into quantum architecture search (QAS). A 2021 paper introduced the first neural predictor-based QAS, using a neural network to gauge candidate circuit performance from structure alone and integrating it into a search workflow that substantially accelerated the process [5]. That approach demonstrated that predictor-guided QAS could discover state-of-the-art quantum architectures for variational quantum eigensolver and quantum machine learning tasks while using an order of magnitude fewer circuit evaluations than a random-search baseline [5]. More recently, the QNAS framework unified hardware-aware one-shot evaluation, multi-objective optimization, and cutting-overhead awareness for hybrid quantum-classical neural networks, jointly optimizing validation error, a runtime cost proxy, and estimated subcircuit count under a target qubit budget [4]. Neural architecture search itself is a well-established technique in classical deep learning, where networks with multiple hidden layers learn hierarchical representations of data [7]. The field has produced architectures such as convolutional neural networks for vision, recurrent neural networks for sequential data, and transformers that use attention mechanisms to model long-range dependencies [7][8]. MZeQAS adapts the search paradigm to the quantum domain, aiming to deliver the same kind of automated design efficiency for parameterized quantum circuits that classical NAS has provided for deep neural networks [2][6].