Quantum computers need their refrigerators fixed fast — and now an LLM panel can help figure out what broke.
Researchers built Onnes, a digital-twin simulator of a dilution refrigerator — the hardware that keeps superconducting quantum computers near absolute zero. The system pairs a physics model with a noise fingerprint learned from real BlueFors logs, then runs a multi-agent LLM layer on top to diagnose faults. In a 1,000-turn evaluation, a zero-shot LLM panel matched a supervised machine-learning classifier on fault detection, though it struggled with classification on closely overlapping fault types. Adding just six labeled examples via contrastive few-shot demonstrations pushed classification accuracy from 0.685 to 0.990, matching the supervised classifier's 0.985 with no retraining.
The result matters because dilution refrigerators are notoriously difficult to diagnose — current systems mostly tell operators that something is wrong, not what or why. A sim-to-real check using real BlueFors telemetry logged 100% recall on injected physics faults and a 6.4% false-alarm rate, suggesting the approach is not purely a lab exercise. Getting LLM agents to this accuracy level with minimal labeled data could lower the barrier for quantum hardware operators who don't have large labeled fault datasets.
The broader implication: few-shot prompting is quietly closing the gap with supervised ML in specialized scientific domains, which should make the "you need thousands of labels" crowd at least mildly uncomfortable.