Zhang and Colleagues Develop Vibe Calibration for 112-Qubit Superconducting Quantum Processors

Huikai Xu and colleagues from Beijing Academy of Quantum Information Sciences, Hangzhou Institute for Advanced Study, University of Chinese Academy of Sciences, Chinese Academy of Sciences, AI for Science Institute and 1 other institutions, have developed Vibe Calibration. This is a new autonomous system using large language model agents to simplify calibration as superconducting quantum processors surpass one hundred qubits. The system achieves fully autonomous calibration of a 112-qubit processor in 4.7 hours, delivering a 4.5× speedup compared to manual calibration. It provides a reusable, auditable framework for scalable quantum hardware operation.

Autonomous calibration of over one hundred superconducting qubits using large language model agents

Vibe Calibration autonomously calibrated 108 of 112 qubits on a superconducting processor in 4.7 hours, representing a 4.5× speedup compared to previous manual calibration methods. Fully calibrating processors exceeding one hundred qubits was previously limited by the time required for manual operation and the difficulty of scaling expert knowledge. Vibe Calibration utilises large language model agents to translate expert knowledge into reusable “Skills”, packaged as decision trees, enabling autonomous execution and error correction. This approach represents a departure from inflexible scripted routines.

The system also demonstrated transferable calibration workflows, adapting to new processors with minimal interface changes, and establishing a reusable framework for quantum hardware operation. Each Skill combines specific measurement commands, clear criteria for accepting results, and a complete audit trail, allowing independent operation and error correction. Human experts collaborated to capture this knowledge, and a large language model agent, software akin to chatbots, refined its capabilities by training it on validated calibration sequences. This process involved validation of experimental results and summarisation into a library usable by the model, allowing it to learn and refine its calibration approach.

Autonomous quantum processor calibration via expert knowledge distillation and large language models

An autonomous system, Vibe Calibration, utilising large language model agents to calibrate superconducting quantum processors, has developed. The system operates on processors exceeding one hundred qubits, and tested a 112-qubit device featuring frequency-tunable transmons. Distilling expert knowledge into reusable “Skills”, packaged as decision trees, chose to overcome the limitations of inflexible conventional scripts and the scalability issues of manual calibration. Agreement on 14 qubits observed when comparing the system’s performance against expert manual calibration on a smaller 16-qubit processor, validating its accuracy. However, results do not yet demonstrate strong performance across a larger number of qubits, nor address the challenges of maintaining calibration stability over extended periods of operation.

Large language models accelerate quantum computer calibration despite initial human knowledge

The relentless pursuit of more powerful quantum computers demands not just increased qubit counts, but also a solution to the escalating complexity of calibrating these systems, as current methods struggle to keep pace with growing processor sizes. While the system successfully demonstrated autonomous calibration using large language models, it currently relies on a human-in-the-loop process to initially capture expert knowledge and build these “Skills”. Despite this reliance on initial human input to establish the foundational “Skills”, the achievement nonetheless represents a major stride forward. A 112-qubit superconducting processor was fully autonomously calibrated by the Hangzhou Institute for Advanced Study team, a vital step beyond conventional methods hampered by manual intervention and inflexible scripts. By distilling the expertise of quantum computer operators into reusable “Skills”, packaged as decision trees, the system achieved calibration of 108 qubits in under five hours, moving beyond simply automating existing procedures. This advance establishes a reusable framework for quantum hardware operation, potentially accelerating the development of larger and more reliable quantum computers, and raises the question of how to extend this “Skill” abstraction to diverse quantum computing platforms.

The researchers successfully demonstrated an autonomous calibration system for a 112-qubit superconducting processor using large language models. This system distils expert knowledge into reusable “Skills”, enabling it to calibrate 108 qubits in 4.7 hours, a four to five-fold speedup compared to manual calibration. The approach shows agreement with expert calibration on a 16-qubit subset, and importantly, the model’s workflows are transferable between devices. The authors suggest this “Skill” abstraction could be extended to other quantum computing platforms, offering a framework for more efficient hardware operation.