IQM uses a chart-reading AI agent to calibrate a 20-qubit chip

IQM Quantum Computers says an AI agent calibrated all 20 qubits of a processor by interpreting measurement plots and choosing the next laboratory action. The demonstration connected NVIDIA Ising vision-language models with IQM’s control software, moving the model from answering questions about charts to acting on measurements from a real device.[1]

A superconducting quantum processor cannot be treated like a conventional chip that simply starts working when power arrives. Its qubits require careful tuning. An engineer runs an experiment, studies the resulting plot, adjusts a control parameter and tries again. The next plot might show the expected resonance—or reveal that the right setting sits beyond the range measured. Repeating those choices across a growing device is a tedious and specialised part of making quantum hardware useful.

The company’s 21 October account describes one useful adjustment: when the relevant setting lay outside an initial measurement range, the agent expanded the range and reran the affected measurements. It therefore used a plot to recognise that the experiment itself needed to change, rather than only assigning a label to the result.[1]

A shared test for chart interpretation

IQM also helped develop QCalEval, an open benchmark intended to make AI calibration comparisons less dependent on private demonstrations. The dataset contains 243 charts from 22 types of experiment. Fifty-seven charts were obtained from real hardware, with the remainder simulated. Each chart is associated with six questions a quantum engineer might ask, including what the graph shows and which step should follow. Most experiments concern superconducting qubits, although other modalities are represented.[1]

IQM reports that specialised NVIDIA Ising Calibration models scored up to 19 points above selected alternatives on QCalEval. That is a company-reported comparison on questions about calibration charts. The dataset includes 57 charts from real hardware and 186 simulated charts, making its composition relevant when interpreting the result.[1]

The benchmark and the chip demonstration answer complementary questions. Shared chart questions make models easier to compare under common conditions. A physical run tests whether interpretation can be connected to an instrument and a sequence of actions. Neither alone measures how reliably the system would maintain a processor over a long operating period.

From one adjustment to a stable processor

The gap between image understanding and laboratory control is crucial. Real hardware drifts. A tuning decision can influence neighbouring qubits, and a parameter that looks correct in one isolated measurement may create a problem when a more complex gate is exercised. Calibration succeeds operationally when the chip meets well-defined performance tolerances and remains stable enough to run the intended computation. Reading a resonance plot is an important part of that chain, not the entire chain.

The announcement does not provide a controlled before-and-after comparison of calibration time or gate fidelity. It therefore supports the specific claim that the workflow operated across this 20-qubit device, rather than a quantified productivity gain or superiority to an expert under equivalent conditions.[1]

IQM also says the agent documented its actions in natural language. For engineers reviewing an unexpected adjustment, that can provide a readable account of the sequence. Its usefulness depends on retaining the corresponding measurements and control calls, so the explanation can be checked against what the instrument actually did.[1]

Further evaluations could run the same workflow on larger processors and unfamiliar failure cases, measuring time, achieved performance and interventions by engineers. These would reveal whether the agent handles routine work consistently and recognises situations that need a specialist.

The demonstration’s distinctive contribution is already clear: a model interpreted a deficient measurement and changed the next experiment accordingly. That feedback loop is the part that could reduce repetitive calibration work if it remains dependable as the device and its operating conditions become more complex.

Report an error or suggest a correction