Agentic AI for Scientific Reasoning in Autonomous Quantum Sensing Experiments
Takuya Isogawa, Ryotaro Okabe, Nutdech Phadetsuwannukun, Mingda Li, Paola Cappellaro
quant-ph, cs.AI
2026-07-28
An LLM agent runs NV-center quantum-sensing experiments unattended for 18.9 h. Deeper reasoning raises false positives for resonance calls; an expected-signal check keeps them low.
Nitrogen-vacancy (NV) centers in diamond are quantum defects used to sense magnetic and electric fields, temperature, and strain. A full experiment means repeatedly picking a center, calibrating the resonance frequency, running pulse sequences, and judging whether a signal is real. The bottleneck is people: what is scarce is not instrument uptime but an experimentalist who can watch the run and call shots in real time. And the people who need NV measurements are usually experts in a target field such as materials or biology, not in NV sensing itself.
This paper puts an LLM agent in charge of such experiments. It does not touch hardware design; given that the apparatus is already computer-controlled, it hands the full measurement workflow to the agent for autonomous decisions.
The architecture hard-separates scientific reasoning from hardware control. The LLM agent reads the project brief, human advice, recent results, and knowledge documents, writes short Python scripts for calculation, analysis, and simulation, and on that basis forms hypotheses and decides what to measure next. All progress goes into persistent project records that track current state, supported conclusions, and unresolved uncertainties.
The hardware side is deterministic control software. Every measurement request passes through its checks; it queues, executes, and records data, and it enforces the safety boundary. The agent never touches the device directly.
A standard autonomous run chains these steps: confocal scan to pick an NV center, track the candidate, use a strong-π pODMR to locate the resonance coarsely (more robust to offset), then a weak-π pODMR to calibrate the frequency precisely, then Ramsey to measure T2, and if Ramsey hints at nearby carbon-13 coupling, autonomously add a CPMG sequence to check.
The main experiment ran 18.9 hours unattended. The agent selected a single NV center, refined the resonance from 3.87646 GHz on the strong-π pODMR (about 14% depth) to 3.87650 GHz on the weak-π pODMR, measured a static field of about 359 G, computed the expected carbon-13 Larmor frequency of about 384.6 kHz, and saw a response with CPMG N=8 at τ ≈ 1/(4fc). T2 landed at the 2 to 3 µs scale. Its verdict on carbon-13 was a "likely weak or moderate signature", a tentative result, not a confirmation. Two more cases (6.3 h with no resolved coupling; 22.2 h of reanalysis that needed a human hint about residual offset) show the agent recovering from failed measurements and revising its plan.
More useful are two offline benchmarks that isolate reasoning. The Ramsey checkpoint benchmark (5 checkpoints, 400 runs per model) tests whether the agent spots a missing residual calibration offset:
| Model (per 100, low to xhigh) | Pass count |
| GPT-5.4 | 7 / 13 / 17 / 20 |
| GPT-5.5 | 11 / 17 / 21 / 34 |
| GPT-5.6 Sol | 16 / 42 / 55 / 49 |
More reasoning effort raised the pass rate (GPT-5.6 Sol drops back to 49 at xhigh), and no model passed cp04.
The second benchmark, judging whether a pODMR measurement contains a resonance (96 measurements, 24 with resonance), gives a counter-intuitive result. With only pulse-sequence information, deeper reasoning raises the false-positive rate: GPT-5.4 climbs from 1.39% to 16.67%, GPT-5.5 to 53.24%, and GPT-5.6 Sol to 45.83%, and newer models false-positive more. Once the agent must compute an expected signal before judging, the false-positive rate drops to between 0% and 3.70% across all three models and every reasoning setting.
The practical takeaway is a division of labor. Use reasoning where it earns its keep: stitching evidence together, forming hypotheses, hunting anomalies. Use deterministic code for the rest: driving hardware, enforcing safety, doing routine data calls. Judging whether a resonance is present is a yes/no task where a more powerful reasoner over-reads noise into a signal that is not there; forcing it to compute the expected response first is the right constraint.
For autonomous experiments as a direction, it demonstrates a safe, repeatable split of an agent that proposes hypotheses backed by deterministic code, rather than letting an LLM command hardware directly.
The authors' stated bounds: the workflow assumes an already computer-controlled apparatus with an explicit safety boundary and does not address building new systems; it is tested only on NV centers, so other platforms and measurements await validation; there are only three end-to-end cases, which is why the benchmarks were added; and newer models false-positive more in some conditions, so prompt engineering must keep up.
The weight of "likely" deserves a harder look. The carbon-13 weak feature is hedged by the agent itself as possible, and the discussion reminds readers not to treat it as confirmed. The easiest trap for autonomous experiments is the agent reporting noise as a confident signal, and the pODMR benchmark, where false positives rise with reasoning, lands exactly on that risk. The paper contains the problem for resonance judgment with a forced expected-signal check, but whether that check generalizes to harder scientific calls is not yet shown.