Collaborative AI agents for autonomous experimental research with solid-state quantum spins

El Mustapha Mansouri; Eikichi Kimura; Masato Koga; Keigo Arai

cs.AIOct 29, 2026

Collaborative AI agents for autonomous experimental research with solid-state quantum spins

Authors: El Mustapha Mansouri, Eikichi Kimura, Masato Koga, Keigo Arai

Organizations: School of Engineering, Institute of Science Tokyo, Yokohama, Kanagawa 226-8501, Japan

Abstract

An autonomous research laboratory could conduct an entire study, from formulating questions to preparing manuscripts. Artificial intelligence (AI) offers one route towards this goal. Recent AI systems take on several research roles, but may draw conclusions that the evidence does not support. Here, we present an autonomous research laboratory in which agents based on large language models assess one another's work and can stop or redirect investigations. They can generate and reject research questions, require separate theoretical checks and conduct experiments designed to challenge their claims. We demonstrate these capabilities by investigating ensembles of quantum spins in diamond. The agents derived the minimum number of calibration settings needed to predict responses to a defined family of microwave pulse sequences. They checked the result computationally and designed experiments to distinguish competing predictions. An initial experiment agreed with the theoretical predictions, but repeated reference measurements showed that the signal was drifting during data collection. The agents therefore treated the result as inconclusive and revised the measurement protocol without relaxing the acceptance criteria. The revised experiment met those criteria and supported the theoretical predictions. This study demonstrates that collaborating AI agents can select research questions, derive theoretical results and test their predictions experimentally.

Figures & tables

Figure 1: Scientific workflow and coordination of the collaborative AI laboratory. (a) Five-stage research cycle with checkpoints overseen by the principal investigator (PI). Dotted lines indicate PI involvement; dashed arrows indicate rejection and revision. (b) Specialized AI agents and their exchanges of research outputs, with safety-agent approval required before experimentation. (c) Document flow during one pass through the API-orchestrated workflow, from left to right (not to scale). The orchestrator routes documents to the persistent PI conversation (blue). Dots mark document deliveries, circled ticks and crosses indicate PI decisions, and dashed links trace rejected work. Upper branches represent separate model calls; lower blocks represent coding-agent jobs.

Figure 2: How the accepted proposal emerged. (a) Every proposal at its delivery time, marked by the PI’s decision. The dashed line marks when the repeatedly rejected idea families were listed in the proposers’ instructions. The ringed proposal was rejected but named the best seed, and its revision was accepted. (b) Reasons for rejecting 75 proposals, in four categories: logically invalid (false or invalid claims, overclaims), already known, not significant (no core theorem, trivial, renamed mathematics, too narrow) and no feasible test. Dark bars count the decisive reason and light bars every reason cited.

Figure 3: Discovery and verification of a minimal ensemble-calibration law. (a) The revised proposal supplied an upper bound of

8L^{2}+6L+1

calibrations, which the principal-investigator agent refined to the exact necessary-and-sufficient count

L^{2}+1

under persistent pulse errors. (b) Calibration requirements compared with the number of possible sequences for five phase choices per pulse, with the first phase fixed. (c) Certified prediction-error and calibration-ambiguity bounds under hidden phase rotations, relative to photon-counting noise.

Figure 4: Experimental design to distinguish the calibration law from interpolation. (a) Rejected design using a complete grid of three-pulse calibrations, from which the withheld targets can already be determined by Fourier interpolation. (b) Revised design combining five three-pulse calibrations with five shorter sequences to predict four withheld targets (

T_{1}

–

T_{4}

). Filled symbols indicate calibration sequences and open symbols indicate withheld targets. In both maps, rows give the phase of pulse 2 and columns the phase of pulse 3. Both designs also include the same five shorter calibrations of zero to two pulses, which are not drawn. (c) Measured signals of the final run along each row of (b), relative to the mean of the five three-pulse calibrations. Dots are measured calibrations, whose standard errors are smaller than the dots. The calibration law (blue) predicts each row from all ten calibrations, whereas the two interpolation-based alternatives (grey, solid and dashed) use only the five three-pulse calibrations. Rings mark the withheld targets at the law’s predictions, fixed before the target measurements were analysed; the measured targets are shown in Fig. 5 c.

Figure 5: Critical rejection and experimental support for the AI-derived calibration predictions. (a) Signal of the first run in each 10-min sub-run, in weighted counts per shot, for the two acquisition programs; the dashed line marks 11:48, when the room light came on. (b) The global-phase and reference-pulse checks of each acquisition, as a ratio to their preset limits; checks in the shaded region fail. (c) The four withheld targets, relative to the mean of the five three-pulse calibrations: locked predictions of the calibration law (blue) and of the two interpolation-based rival rules (grey), and the final measurements (orange), whose standard errors are smaller than the symbols. (d) The target contrast

S

formed from the four targets in

10^{-3}

weighted counts per shot; the shaded band marks the predefined tolerance of

\pm 20\%

around the law’s locked prediction.

AI Agents for Scientific Discovery

AI agents for scientific discovery use language models, tools, and experiments to generate, test, and refine scientific knowledge.

See all 293 papers

Agentic AI for Scientific Reasoning in Autonomous Quantum Sensing Experiments
Jul 27, 2026Takuya Isogawa, Ryotaro Okabe, Nutdech Phadetsuwannukun +2AI Agent Benchmarks Laboratory Automation
Department of Nuclear Science and Engineering, Massachusetts Institute of Technology, Cambridge, MA 02139, USA · Department of Chemistry, Massachusetts Institute of Technology, Cambridge, MA 02139, USA · Department of Physics, Massachusetts Institute of Technology, Cambridge, MA 02139, USA
HERA: Assessing an Autonomous Agent for Quantum Device Measurement
Oct 29, 2026Zachary W. Gomez, Nathan Bittner, Jordan T. McCourt +7Autonomous Experimentation Autonomous Research Agents
Stanford Institute for Materials and Energy Sciences, SLAC National Accelerator Laboratory, Menlo Park, CA 94025 · Department of Physics, Stanford University, Stanford, CA 94305 · Research Center for Electronic and Optical Materials, National Institute for Materials Science, 1-1 Namiki, Tsukuba 305-0044, Japan +1
Claw AI Lab: An Autonomous Multi-Agent Research Team
May 21, 2026Fan Wu, Cheng Chen, Zhenshan Tan +12Multi-Agent LLM Systems LLM Agent Orchestration
1NTU · 2A*STAR · 4NUIST +3

Jul 27, 2026J/K move · Enter open · S save

Agentic AI for Scientific Reasoning in Autonomous Quantum Sensing Experiments

Takuya Isogawa, Ryotaro Okabe, Nutdech Phadetsuwannukun, Mingda Li+1

Department of Nuclear Science and Engineering, Massachusetts Institute of Technology, Cambridge, MA 02139, USA · Department of Chemistry, Massachusetts Institute of Technology, Cambridge, MA 02139, USA · Department of Physics, Massachusetts Institute of Technology, Cambridge, MA 02139, USA

AI Agent Benchmarks Laboratory Automation Agentic Reasoning Autonomous Experimentation Autonomous Research Agents

Open paper AskPDF arXiv

We implement an agentic AI workflow built around a large language model (LLM) agent for autonomous experiments with nitrogen-vacancy (NV) centers in diamond. NV centers are a widely used platform for quantum sensing, and the ability to control many measurements from a computer makes NV experiments a natural setting for autonomous workflows. We make two main contributions. First, we demonstrate an autonomous NV experiment workflow that combines persistent project records, quantitative calculation and data analysis tools, and deterministic experiment control. In one autonomous experiment, the agent selected a single NV center, calibrated its resonant frequency, measured

T_2^\ast

with Ramsey measurements, and added a Carr--Purcell--Meiboom--Gill (CPMG) measurement to check a weak feature that could be related to nearby

^{13}\mathrm{C}

. Second, we introduce two offline benchmarks that evaluate the agent's reasoning separately from laboratory execution. We evaluated both benchmarks with GPT-5.4, GPT-5.5, and GPT-5.6 Sol. In the Ramsey checkpoint benchmark, greater reasoning effort generally improved recognition of a residual resonance calibration offset. By contrast, in the pulsed optically detected magnetic resonance (pODMR) data evaluation benchmark, pulse sequence information alone produced more false positive resonance judgments at higher reasoning effort. Requiring an expected signal calculation kept false positive rates low across all three models and reasoning settings. The results suggest a clear division of labor for autonomous experiments. The agent forms scientific hypotheses and uses quantitative tools to evaluate data, while deterministic code controls the hardware and enforces safety constraints.

Abstract

AI Agents for Scientific Discovery

Explore similar work

Agentic AI for Scientific Reasoning in Autonomous Quantum Sensing Experiments

HERA: Assessing an Autonomous Agent for Quantum Device Measurement

Claw AI Lab: An Autonomous Multi-Agent Research Team