⚛️ quantum physics

QMClaw: A Scalable General-purpose Framework for Quantum Measurement and Control

This paper introduces QMClaw, a scalable, general-purpose framework for quantum measurement and control that combines a robust, rule-based core for low-latency execution with language models for high-level interaction, effectively addressing the complexity and timing constraints of large-scale quantum system calibration.

Original authors: Zhiqiang Fan, Haoran He, Ping Lv, Junchao Wang, Yaqiang Sun, Chenhui Wang, Hanshi Zhao, Geyuyan Ma, Haoran Yang, Pengyu Han, Xiangdong Meng, Lixin Wang, Feng Yue, Weilong Wang, Zheng Shan

Published 2026-09-07

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Zhiqiang Fan, Haoran He, Ping Lv, Junchao Wang, Yaqiang Sun, Chenhui Wang, Hanshi Zhao, Geyuyan Ma, Haoran Yang, Pengyu Han, Xiangdong Meng, Lixin Wang, Feng Yue, Weilong Wang, Zheng Shan

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Quantum computers hold the promise of solving problems that are impossible for today's machines, but they are notoriously fragile. To function, their tiny components, known as qubits, must be kept in a state of perfect balance, free from the slightest interference. This requires a constant, delicate process of tuning and measuring, much like a musician tuning a violin before a concert, but happening millions of times faster and with far less margin for error. In the current era of quantum development, these tuning tasks rely heavily on human experts who manually adjust settings, run tests, and interpret results. As quantum machines grow larger, adding more qubits, this manual approach becomes impossible. The sheer volume of adjustments needed, combined with the need for split-second timing, creates a bottleneck that threatens to stall the progress of the entire field. Scientists need a way to automate these measurements and controls that is not only smart but also incredibly fast and reliable.

A team of researchers has addressed this challenge by building a new system called QMClaw, a framework designed to manage the measurement and control of quantum computers. Rather than relying on a single, all-knowing artificial intelligence to make every decision, the researchers created a hybrid system that separates routine tasks from complex reasoning. At the heart of this system is a rule-based engine, a set of strict, pre-defined instructions that handle the vast majority of daily operations. This engine acts as the fast lane, making decisions in microseconds to keep the quantum hardware running smoothly. It handles the repetitive work of checking frequencies, adjusting pulses, and verifying measurements without hesitation. This approach ensures that the system remains deterministic and predictable, qualities that are essential when dealing with the sensitive physics of quantum devices.

The researchers tested this system by having it perform a complete calibration routine on a single qubit, a fundamental unit of a quantum computer. The process began when a user simply typed a request in plain English, asking the system to start a measurement workflow. The system understood the intent, broke the request down into a series of specific steps, and executed them automatically. It scanned for resonator frequencies, optimized magnetic fields, and measured how long the qubit could hold its state. Throughout the process, the system collected raw data, analyzed it in real time, and produced a detailed report of the results, including specific performance metrics like the qubit's frequency and how long it could maintain its quantum state. The entire sequence was completed without any human intervention, proving that the framework could bridge the gap between a simple human command and the complex, low-level operations required by the hardware.

One of the most significant findings of this work is how the system manages the use of large language models, the powerful artificial intelligence tools often associated with human-like reasoning. The researchers found that relying entirely on these models for control would be too slow and too expensive for large-scale quantum systems. Instead, they designed QMClaw to use these models only as a safety net. The system uses the fast, rule-based engine for everything it knows how to do. It only calls upon the language model when it encounters a situation it has never seen before or when something goes wrong. This strategy drastically reduces the number of times the system needs to consult the language model. In simulations of a million-qubit system, this hybrid approach reduced the number of language model calls by a factor of one hundred compared to a system that relied on the model for every decision. Furthermore, the cost of running the system dropped by a factor of one thousand, while the speed of decision-making remained incredibly fast, with the rule-based engine making decisions in less than a microsecond.

To handle the massive scale of future quantum computers, the researchers also introduced a method of grouping qubits into panels. Instead of asking the system to think about every single qubit individually, which would overwhelm the computer's memory and processing power, the system treats groups of connected qubits as single units. This allows the system to process calibration tasks for large arrays of qubits efficiently, compressing the information it needs to manage. By testing the system against historical data from real quantum devices, the researchers confirmed that their method could accurately predict the next steps in a calibration sequence with high precision. The system learned from past experiments, identifying patterns in how the hardware behaved and using that knowledge to guide future actions. This ability to learn from history while sticking to strict rules for the present moment suggests a viable path forward for managing the complexity of tomorrow's quantum machines.

The work presented here does not claim to have solved every problem in quantum control, nor does it suggest that artificial intelligence should be removed from the equation entirely. Instead, it offers a practical blueprint for how to integrate these tools effectively. The researchers argue that for quantum systems to scale, the core control loop must remain fast and reliable, governed by clear rules, while the flexible, reasoning power of artificial intelligence is reserved for the edges of the process. This design allows the system to handle the routine with speed and the unexpected with intelligence. By validating this approach on real hardware and demonstrating its ability to manage complex workflows without human oversight, the team has provided a concrete step toward making large-scale quantum computing a reality. The framework shows that the path to scaling quantum technology lies not in replacing human expertise with a single, all-powerful brain, but in building a structured system where different tools play to their specific strengths.

Technical Summary: QMClaw – A Scalable General-purpose Framework for Quantum Measurement and Control

1. Problem Statement

As quantum computing transitions into the systems engineering stage, Quantum Measurement and Control (QMC) faces critical bottlenecks in scalability. Current QMC workflows are constrained by:

Complex Calibration: Manual expert intervention is required for configuration, tuning, and scheduling due to hardware decoherence, calibration drift, and device non-uniformity.
Fragmented Automation: Existing script-based automation accelerates isolated tasks but lacks a unified architecture for task management, runtime monitoring, closed-loop adaptation, and traceable execution.
Limitations of AI Agents: While Large Language Models (LLMs) and Vision-Language Models (VLMs) show promise in interpreting goals and planning, they suffer from excessive latency, poor generalization across hardware setups, and unreliable diagnostic reasoning. Unconstrained LLM reasoning introduces overhead that violates the strict real-time and low-latency requirements of large-scale quantum systems.

There is an urgent need for a general, cross-platform framework that coordinates multi-stage calibration under real-world constraints while balancing the flexibility of natural language with the determinism required for hardware control.

2. Methodology: The QMClaw Framework

QMClaw is a general-purpose, workflow-oriented framework designed to organize the full path from task intake to backend execution and structured reporting. It employs a "local-first, tool-governed, robust" architecture that separates high-level semantic understanding from low-latency control logic.

2.1 Six-Layer Execution Stack

The framework is structured into six interconnected layers to transform natural language requests into executable workflows:

Natural-Language Interaction Layer: Receives user intent via NLP, abstracting underlying control complexities.
Task Understanding and Planning Layer: Extracts constraints, maps tasks to predefined workflow templates, and structures goals into actionable formats.
Workflow Construction and Execution Control Layer: Maintains runtime control logic, state monitoring, and dynamic task allocation to ensure robust execution paths.
Tool Scheduling and Backend Invocation Layer: Schedules and invokes software tools, simulators, and hardware drivers, managing computational resources.
Hardware/Simulator Interaction Layer: Bridges the framework with physical quantum devices or simulators, ensuring real-time data collection and synchronization.
Data Analysis and Structured Reporting Layer: Interprets experimental outputs, evaluates validity, and generates structured reports with execution traces.

2.2 Three-Layer Hybrid Decision Architecture

To address latency and determinism, QMClaw avoids placing LLMs on the critical control path. Instead, it utilizes a three-tier decision hierarchy:

Layer 1 (RuleEngine): A minimal, deterministic engine that uses structured context and explicit transition rules to resolve standard calibration steps. It provides a fast, inspectable path for canonical transitions.
Layer 2 (Mined State-Machine): Captures recurrent workflow patterns (e.g., frequency scans, pulse loops) observed in historical data. These patterns are converted into reusable operational knowledge, reducing the need for repeated model calls.
Layer 3 (LLM Fallback): A bounded, high-latency layer activated only when L1 and L2 cannot resolve a context. It handles exception support, diagnosis, and candidate rule acquisition. Validated L3 resolutions are materialized into L2, creating a knowledge accumulation loop.

2.3 Multi-functional Agent Scheduling

The system follows a centralized-scheduling, distributed-execution model. A "Main Agent" handles high-level semantic interpretation and task decomposition, assigning subtasks to specialized subagents (e.g., task-parsing, execution, data-analysis, report-generation). This role-specialization ensures that different constraints and tool interfaces are managed by appropriate modules under a common orchestration mechanism.

2.4 Panel-Based Qubit Grouping

To mitigate the computational overhead of scaling to millions of qubits, physical qubits are partitioned into independent "panels" based on hardware topology and functional correlation.

Each panel acts as a unified processing unit, encapsulating its context into a single batch for LLM interaction.
This strategy compresses redundant context, avoids scattered single-qubit invocations, and drastically reduces total LLM calls and latency.
Closed-loop adaptive adjustments occur at the panel level, triggering strategy revisions based on in-panel calibration accuracy.

3. Key Contributions

The paper presents four primary contributions:

QMClaw Framework: A scalable, workflow-oriented framework unifying natural-language interaction, structured tasking, tool governance, execution monitoring, and result reporting.
RuleEngine-Centered Architecture: A control design enabling lightweight, deterministic, and traceable workflow orchestration, separating routine control from open-ended model inference.
Validated Implementation: A fully validated state-machine implementation for single-qubit tune-up (a critical procedure in superconducting systems) demonstrating robust closed-loop calibration on real hardware.
Scalability Analysis: A systematic study quantifying real-time performance, bottlenecks, and LLM efficiency under large-scale operating assumptions.

4. Experimental Results and Validation

The framework was validated using real quantum device datasets and natural language interfaces.

Natural Language Control: The system successfully executed a complete single-qubit tune-up workflow (including resonator scanning, flux bias optimization, spectroscopy, Rabi oscillation, Ramsey interferometry, and $T_1$ measurement) via a WeChat interface. It autonomously collected data, performed on-the-fly analysis (e.g., fitting decay curves), and returned calibrated parameters ( $f_{10} = 2.980$ GHz, $T_1 = 37.39$ µs, etc.) without manual intervention.
Historical Data Prediction: Using a dataset of 73 qubits and 2,476 experimental steps, a multi-layer Markov prediction strategy was tested. A second-order Markov model achieved 71.6% accuracy in predicting the next experiment type, outperforming first-order Markov models (66.3%) and LLM-based predictors (68%). The overfitting gap was minimal (2.2%), confirming strong generalization.
LLM Overhead and Latency:
- Latency: The RuleEngine achieved an average latency of 0.19 µs per evaluation, approximately 500,000 times faster than typical LLM inference (100–500 ms).
- Scalability: In a simulated million-qubit scenario, the hybrid architecture reduced daily LLM invocations from 43.2 million (pure-LLM baseline) to 432,000.
- Cost: The hybrid approach reduced daily operational costs by a factor of 1,000 compared to a pure-LLM design.

5. Significance and Claims

The authors claim that QMClaw provides a feasible system architecture for building scalable intelligent quantum experiment platforms. Its significance lies in:

Resolving the Latency-Determinism Trade-off: By confining LLMs to exception handling and high-level interpretation, the framework maintains the real-time performance and auditability required for large-scale quantum systems.
Generalizability: Unlike specialized, task-specific frameworks, QMClaw offers a unified, cross-platform architecture applicable to diverse hardware environments.
Empirical Evidence: The work provides quantitative evidence that rule-centered architectures are a promising design choice for scalable quantum-system calibration, demonstrating that AI modules should be embedded within standardized, constrained workflows rather than acting as unrestricted decision executors.

The paper modestly acknowledges limitations, noting that current verification is limited to specific experimental sessions and single-type devices, and that future work is needed to expand to multi-qubit calibration, optimize panel grouping for diverse topologies, and integrate advanced anomaly detection.