Pith. sign in

REVIEW 4 major objections 5 minor 46 references

An LLM-driven agentic workflow can carry a quantum protocol from a published paper or patent through emulation to overnight execution on cloud neutral-atom hardware, while the human files the scientific choices—and the paper shows the human

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-01 01:20 UTC pith:WDAHV4LQ

load-bearing objection Credible vendor demo of agent-driven neutral-atom QPU workflows with real data and honest failure reporting, but the corpus' nearly-half-implementable claim is inflated by a decision tree that ignores the quantitative device limits the paper itself shows are decisive. the 4 major comments →

arxiv 2607.25834 v1 pith:WDAHV4LQ submitted 2026-07-28 quant-ph cond-mat.quant-gascs.AI

Lowering the implementation barrier of neutral-atom quantum computing with agentic workflows

Constantin Dalyac , Alexandre Dauphin , Lo\"ic Henriet , Christophe Jurczak This is my paper
classification quant-ph cond-mat.quant-gascs.AI
keywords neutral-atom quantum computingagentic workflowlarge language modelsRydberg arraysquantum cloud executionexperiment automationprotocol extractionquantum simulation
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

This paper argues that the bottleneck between theoretical quantum proposals and real hardware experiments has shifted from hardware availability to a coordination problem, and that an LLM-driven agentic workflow can automate the entire chain from protocol extraction to cloud QPU execution. In three case studies, the agent turned a published paper or patent into an overnight QPU campaign, producing data that matched the original experiments, a clear 1/3-filling plateau, and chromatic-number estimates within two of the classical optimum. However, the agent twice produced plausible but wrong science—an inadequate observable and a false hardware diagnosis—that only domain-expert review caught. On the corpus side, a second agent classified 526 Rydberg-array theory papers and found 49% implementable on present-day neutral-atom QPUs, with XY interactions and local addressing as the main missing capabilities. If correct, the practical takeaway is that a researcher can delegate the engineering to an agent and focus on scientific judgment, but the evidence also shows the human's role becomes more decisive, not less.

Core claim

The central claim is that an agentic workflow built from six modular LLM skills—protocol extraction, pulse-sequence generation, validation emulation with and without device noise, QPU submission, and data harvesting—operating on a single structured experiment specification, can translate a quantum protocol from the literature into a real, cloud-executed QPU experiment with minimal human interaction. The paper demonstrates this on three tasks of increasing complexity: reproducing a known many-body correlator, preparing a frustrated magnetic phase from a theory paper, and executing a patent's graph-coloring algorithm. In each case the workflow produced publishable data, while the human supplie

What carries the argument

The load-bearing object is the agentic workflow itself: six composable skills—extracting a protocol from a source, converting it to a pulse sequence, validating it with a noiseless and a noise-aware emulator, submitting it to a cloud QPU, and processing the returned data—that share a single machine-readable experiment specification (a structured artifact encoding the register geometry, pulse parameters, observables, and constraints). The specification is the human–agent interface and the propagation channel: any modification flows to all downstream stages. For the corpus analysis, the machinery is a five-rule decision tree (N≤100, global drive only, Ising or blockade interactions, no mid-cir

Load-bearing premise

The headline statistic rests on the assumption that the second agent's automated classification of 526 papers—applied by the same kind of LLM that chose a wrong observable and invented a false hardware story earlier in this paper—is accurate enough that 49% is a reliable number; the paper gives no human-validated ground-truth rate for that classifier.

What would settle it

Hand-label a random sample of, say, 100 of the 526 classified papers using a neutral-atom experimentalist's judgment and compare against the agent's labels; if agreement is materially below 90% (say), the 49% claim lacks support. A sharper falsifier: re-run the three case studies with the human-validation steps deliberately disabled and check whether the agent catches its own wrong observable or false diagnosis; based on the paper's narrative, it would not, implying that the claimed automation is conditional on human expertise that the workflow does not itself supply.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • If a researcher can go from a paper or patent to a QPU campaign overnight, experimental throughput on neutral-atom platforms rises by orders of magnitude, and researchers without hardware expertise can test theoretical proposals directly.
  • If the 49% corpus figure is robust, roughly half of the current Rydberg-array theory literature constitutes immediate experimental targets, while the other half is gated primarily by two specific hardware upgrades: XY-type interactions and local addressing.
  • The demonstrated failure modes imply that agent output must be treated as a scientifically unvalidated draft; the paper's own prescription is a researcher-in-the-loop refinement loop, not full autonomy.
  • Executable specifications could turn publications into reproducible artifacts: a protocol stored as a structured specification can be re-instantiated on new hardware and revalidated with minimal manual effort.
  • The corpus analysis supplies a demand-driven hardware roadmap: adding XY interactions and local addressing would unlock the largest fraction of currently inaccessible literature.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The 49% statistic is inherently platform-relative: it was computed against the capabilities of one specific neutral-atom architecture (no site-selective driving, Ising-only interactions, a 100-atom cap). A different architecture with different constraints could yield a substantially different fraction, and the paper does not quantify how the fraction would shift under plausible alternative capabil
  • A testable extension is to run the same corpus classifier against a deterministic rule base without an LLM and compare fractions; if the LLM's 49% is robust, the workflow would serve as a reusable benchmark for future literature-feasibility tools.
  • The 'wrong observable' failure suggests a general failure class: LLM agents optimize for tractability over scientific validity. A practical mitigation, not explored in the paper, is to require the agent to defend its observable choice against a formal criterion—e.g., whether the observable distinguishes the target phase from all other states at the same density—before QPU submission.
  • If the 49% figure is even approximately right, the agentic workflow could be re-run after each hardware upgrade, giving a quantitative, literature-driven measure of progress in neutral-atom quantum computing—an implicit benchmark the paper leaves unconsummated.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

4 major / 5 minor

Summary. The paper reports an agentic workflow that takes a natural-language objective derived from a published paper or patent and carries it through protocol extraction, pulse-sequence generation, noiseless and noise-aware emulation, cloud submission to two Pasqal QPUs (FC1 and SA1), and data reduction. Three end-to-end demonstrations are presented: reproduction of the Z2 Kibble-Zurek correlator of Keesling et al., preparation of the 1/3-filling ordered phase in a triangular Rydberg array studied by Guo et al., and a graph-coloring algorithm taken from a Pasqal patent. The authors candidly document two scientific failures of the agent: it chose an inadequate observable in Example B and produced a plausible but incorrect hardware diagnosis in Example C. A separate corpus agent is then used to classify 526 Rydberg-array theory papers, of which 258 (49.0%) are labeled implementable on present-day Pasqal QPUs, with XY interactions and local addressing identified as the main missing capabilities.

Significance. If the central claims are correct, the paper would be a useful demonstration that modern LLM agents can reduce the engineering overhead between a theoretical protocol and cloud quantum hardware, and that a substantial fraction of the Rydberg-array theory literature is, at least in a coarse sense, within reach of current neutral-atom devices. The paper has real strengths: it anchors two of its three demonstrations against external results (Keesling et al. and Guo et al.), it runs real QPU campaigns rather than only simulations, and it is unusually transparent about failure modes, explicitly attributing two important errors to the agent. The corpus analysis is also framed in an auditable way with a stated decision tree and query strings. However, the headline 49% claim is not yet supported, because the decision tree omits the very quantitative device limits that the case studies show to be decisive, and the classifier is not validated against a human-labeled ground truth. The lack of raw data or workflow artifacts also prevents independent verification of the claimed pipeline and results.

major comments (4)
  1. [§IV, Appendix B, Table I] The five-rule decision tree in Appendix B does not contain any quantitative device limits (maximum detuning, Rabi-frequency cap, coherence-time bound, minimum spacing). Yet §III.B shows these limits are decisive: the 2/3-filling phase of Guo et al. is declared infeasible on FC1 solely because |δ|/2π ≤ 10 MHz forces Ω/2π ≲ 0.6 MHz and a ramp that exceeds the coherence time. That same protocol satisfies all five Appendix B rules and would be marked 'implementable'. The 49.0% figure in Table I is therefore an upper bound on rule compatibility, not a practical measure of implementability, and the overestimate is systematic rather than random. Either augment the decision tree with device-specific quantitative consistency checks (detuning range, Rabi ceiling, ramp duration vs. coherence, spacing floor) or relabel the category as 'rule-compatible' and revise the abstract's 'nearly half' claim a
  2. [§IV and Appendix B] The corpus classification is performed by the same kind of LLM agent that, within this manuscript, selected a scientifically inadequate observable (§III.B) and fabricated a plausible but incorrect hardware diagnosis (§III.C). No human-validated ground-truth labels are provided for any of the 526 classified papers, and 107 of 633 (17%) are dropped as unreadable or insufficiently detailed. The abstract claims 'nearly half' of the literature, but a classifier error rate comparable to the demonstrated scientific-validity failure rate could shift this fraction substantially. The authors should provide a confusion matrix or human-audited random sample, with confidence intervals, and should state how the 107 unclassified papers affect the denominator. Additionally, the 'Pulser/Pasqal/FRESNEL' query accounts for 1,188 of 1,966 raw hits and may bias the corpus toward Pasqal-compatible papers; thi
  3. [§III and Appendix A] The central claim is that the agent can go 'from published paper or patent to a QPU campaign run overnight', yet no artifacts are made available: no experiment_spec.json files, no prompt transcripts, no code repository, no raw bitstring data, and no job metadata. Figures 2–4 report fitted correlation lengths, emulation bands, and QPU-shot statistics without the underlying data or uncertainty estimates for the fitted values. The paper itself argues in §V that publications should evolve into executable scientific artifacts; this contribution would be substantially strengthened, and independently verifiable, if the workflow and at least one full specification/data set were released. As it stands, the demonstrations cannot be reproduced beyond the narrative description.
  4. [Title, §III.A, §III.B] The title and introduction promise a 'lowering' of the implementation barrier, but no baseline comparison is provided. The examples are anecdotal: we are told Example A required three prompts and Example B required 43 exchanges, but there is no measurement of the wall-clock time, expert effort, or success rates of a human-only or non-agent workflow on the same tasks. The claim is plausible, and the transparency about failures is welcome, but a quantitative or even semi-quantitative comparison is needed before the headline claim is fully supported.
minor comments (5)
  1. [Fig. 2] The caption reports fitted correlation lengths ξ = 3.9, 4.2, and 4.2 sites, but no uncertainties or fit ranges are given; please provide them or a reference to a data table.
  2. [§III.C] The text says 75 instances are used across N ∈ {15,20,30,40,50,60,70,80}, but only the N=20 count (10 instances) is stated. Please report the number of instances per size and the exact random-graph generation parameters (e.g., density definition, placement model) so that the benchmark is reproducible.
  3. [Appendix B] The 'pulser_pasqal_sp' query is described as targeting 'Pulser, Pasqal, or FRESNEL'; the acronym FRESNEL is not expanded. Please define it and confirm whether the query is intended to capture neutral-atom hardware papers or software-related papers.
  4. [§I and Appendix A] The agent models are named ('Opus 4.8', 'Fable 5'), but the specific prompts, skill definitions, and tool-call loops are only sketched. A short pseudocode or skill interface specification would clarify the architecture and aid reproducibility.
  5. [§III.B] The term 'marginal' in the verdict table (1/2-filling phase) is not defined. Please state the quantitative criterion used to distinguish feasible, marginal, and infeasible.

Circularity Check

0 steps flagged

No load-bearing circularity; the case studies are externally anchored and the corpus statistic is an acknowledged coarse screen rather than a circular derivation.

full rationale

The paper's central demonstrations are independent of its own conclusions: Example A reproduces a published Keesling et al. result (Fig. 2) and compares against values taken from the external paper; Example B benchmarks QPU data against quantum Monte Carlo from Guo et al. (Fig. 3); Example C evaluates the graph-coloring pipeline against a classical DSATUR baseline. These provide external anchors, and the paper candidly reports two scientific-validity failures (wrong observable in Example B; misdiagnosed blockade leakage in Example C), which is inconsistent with a self-confirming design. The PASQAL-authored stack (Pulser, emu-mps, cloud SDK) and QPUs are self-cited but are operational software artifacts, not unverified claims; the patent used in Example C is a PASQAL patent, which limits the generality of the 'patents as sources' claim but is not circular because the agent's output is externally benchmarked. The corpus-scale '49% implementable' result is an operationalization: Table I labels are the output of the five-rule decision tree in Appendix B. This is a screening definition, not a derivation, and the paper qualifies it as 'back-of-the-envelope' and notes that real signals still require adaptation. The absence of quantitative device limits (e.g., detuning/Rabi caps shown decisive in §III.B) and of ground-truth labels raises validity and reliability concerns about the 49% figure, but these are not circularity: the classifier's criteria are not fitted to the outcome and the count is reported transparently. Score 2 reflects the minor self-referential elements (self-authored patent, PASQAL-centric corpus query) without treating them as load-bearing circular steps.

Axiom & Free-Parameter Ledger

3 free parameters · 5 axioms · 0 invented entities

The central demonstrations rely on the standard neutral-atom Ising/blockade model and on the fidelity of Pasqal's own emulator/cloud stack; the headline corpus statistic depends on hand-selected classification thresholds applied by an unvalidated LLM agent. No new physical entity is introduced, but the paper's conclusions are sensitive to self-defined criteria and vendor-supplied software.

free parameters (3)
  • Beam-search width w and shots per round n_s = w=3, n_s=200
    Hand-picked for Example C without sensitivity analysis; these hyperparameters directly determine the reported coloring-gap statistics.
  • Calibration-offset band on noise-emulation envelope = unspecified
    The acceptance region against which QPU data are judged is "widened by a calibration-offset band" (Appendix A); the width is not reported, so agreement between emulation and hardware is partially constructed.
  • Implementability classification thresholds = N≤100; global drive only; Ising/blockade interactions; no mid-circuit measurement; no basis rotations
    These hand-chosen rules define "implementable today" and hence the 49% headline; no sensitivity analysis or alternate-threshold check is given.
axioms (5)
  • domain assumption Neutral-atom registers implement the Rydberg Ising/blockade Hamiltonian with register, Ω(t) and δ(t) as the only controls.
    Used for all protocol extraction/adaptation; carried from cited platform literature [1,2,28], not re-derived.
  • domain assumption The noisy MPS emulator with live cloud-SDK noise parameters plus calibration-offset band is a faithful predictor of QPU data.
    The GO/NO-GO decision to submit and the later claim that emulation "faithfully envelopes" QPU data both depend on this model being unbiased.
  • ad hoc to paper The reported agent dialogues are representative and not cherry-picked.
    No full transcripts, seeds, or model-version logs are provided; the "3 messages" and "43 exchanges" counts are the only evidence of autonomy.
  • ad hoc to paper The five-rule decision tree is a sufficient proxy for implementability of a Rydberg protocol.
    The corpus claim relies on these coarse criteria; there is no human ground-truth validation and 107/633 papers were excluded.
  • domain assumption Digitized reference data from Keesling et al. Fig. 1d and Guo et al. QMC are accurate.
    The reproduction claims in Examples A and B compare against these published points; any digitization error propagates into the claimed agreement.

pith-pipeline@v1.3.0-alltime-deepseek · 11884 in / 17068 out tokens · 153484 ms · 2026-08-01T01:20:32.957551+00:00 · methodology

0 comments
read the original abstract

Quantum computers are moving from research laboratories to industrial machines accessible via the cloud and integrated into high-performance computing facilities. However, translating theoretical quantum protocols into hardware experiments remains a major bottleneck, requiring expertise across protocol design, compilation, simulation, and cloud execution. Here, we introduce an agentic workflow that automates this pipeline on neutral-atom quantum processors (here two Pasqal QPUs available on the cloud) while keeping the researcher in the loop for critical validation. In three case studies from many-body physics and optimization, the agent went from published paper or patent to a QPU campaign run overnight. In particular, human intervention was crucial to ensure scientific validity: the agent selected an inadequate observable in one experiment and constructed a plausible but incorrect hardware diagnosis in another, with both failures detected only through domain-expert review. Finally, we use a second agent to classify a corpus of 633 Rydberg-array arXiv papers and show that nearly half are implementable on present-day QPUs while identifying specific hardware upgrades needed for the rest. Together, these results demonstrate that agentic workflows provide a practical bridge between theoretical ideas and physical hardware, opening quantum experimentation to a much broader scientific community.

Figures

Figures reproduced from arXiv: 2607.25834 by Alexandre Dauphin, Christophe Jurczak, Constantin Dalyac, Lo\"ic Henriet.

Figure 1
Figure 1. Figure 1: FIG. 1 [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: FIG. 2. Density-density correlator [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: FIG. 3. Normalised order parameter [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: FIG. 4. Example C, QPU beam-search graph coloring on FC1 (unit-disk graphs at density 0 [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

46 extracted references · 8 linked inside Pith

  1. [1]

    Context and task Our objective is to reproduce known results that have already been run on hardware. We select a landmark paper by Keeslinget al.[32] from 2019 where the authors study the Kibble-Zurek mechanism when driving a 1D array of atoms through a quantum phase transition from the paramagnetic to several ordered states, namelyZ2,Z 3 andZ 4. From a d...

    2019

  2. [2]

    Workflow and dialogue The agent extracts the protocol directly from the pa- per: a 51-atom 1D chain, the blockade ratioR b/adefin- ing the ordered lobe (R b/a≈2.16–2.43 forZ 3, withR b the blockade radius andathe atomic spacing), and the pulse schedule together with the sweep rates tabulated 4 for each interaction range. It then fetches FC1’s live de- vic...

    2000

  3. [3]

    Comments This first example already demonstrates excellent agentic capability. In addition to independently con- structing the noiseless-vs-ideal comparison, it also re- covered from infrastructure failures (a paused QPU, er- rors in jobs) without derailing the broader campaign. Agents have the healthy habit of consistently adding checkpoints, error logs ...

  4. [4]

    order-by-disorder

    Context and task Guoet al.[33] explore the quantum magnetic phases of a frustrated triangular Rydberg-atom array. Using quan- tum Monte Carlo, the authors map out a phase diagram containing three ordered lobes as a function of the re- duced detuningδ/Ω: a 1/3-filling plateau, a 1/2-filling “order-by-disorder” phase known to be experimentally fragile, and ...

  5. [5]

    Its verdict table again demonstrated a firm grasp of the hardware constraints

    Workflow and dialogue Invoking thepaper-to-specskill, the agent correctly parsed the paper’s core results, identifying the 1/3 and 2/3 plateaus alongside the delicate 1/2-filling order-by- disorder phase. Its verdict table again demonstrated a firm grasp of the hardware constraints. The decisive lim- itation is FC1’s maximum detuning,|δ|/2π≤10 MHz: since ...

  6. [6]

    Comments The agent demonstrated a significant ability to ex- tract, translate, and execute a hardware experiment di- rectly from a theory paper, eliminating tedious boiler- plate script generation and hardware parameter mapping to sharply reduce the time-to-QPU. 6 However, we observed a key failure mode: the agent consistently optimized for the nearest tr...

  7. [7]

    Context and task Graph coloring (assigning colors to the vertices of G= (V, E) such that adjacent vertices receive differ- ent colors) is equivalent to a sequence of independent- set (IS) problems: peel one IS per round until all ver- tices are covered, with the round count upper-bounding the chromatic numberχ(G). Aproject-and-expelalgo- rithm maps this d...

  8. [8]

    This eliminates re-queuing latency between rounds while holding a single QPU queue slot

    Workflow and dialogue On present hardware, physical atom expulsion is re- placed by a classical measurement-feedback loop: the agent compiles the fullN-atom layout once into a MappableRegisterbatch and submits each round as a separate job within that open batch, activating only the subset of trap sites corresponding to the current subgraph. This eliminate...

  9. [9]

    It also 7 FIG

    Comments This example extends the workflow beyond a single- shot protocol to amulti-round, state-dependentpipeline: the agent compiles the register once, routes ˆχsequen- tial jobs within one open batch, manages beam-search branches across rounds, and aggregates the final coloring, an orchestration pattern that would require substantial bespoke engineerin...

  10. [10]

    Browaeys and T

    A. Browaeys and T. Lahaye, Nature Physics16, 132 (2020)

    2020

  11. [11]

    Henriet, L

    L. Henriet, L. Beguin, A. Signoles, T. Lahaye, A. Browaeys, G.-O. Reymond, and C. Jurczak, Quantum 4, 327 (2020)

    2020

  12. [12]

    A. J. Menssen, M. Gullans, T. Manovitz, J. M. Tay- lor, J. Cong, J. Sinclair, Z. Aqua, D. J. Blumen- thal, J. Ataides, J. Borregaard,et al., arXiv preprint arXiv:2607.21554 (2026)

    Pith/arXiv arXiv 2026

  13. [13]

    Ebadi, A

    S. Ebadi, A. Keesling, M. Cain, T. T. Wang, H. Levine, D. Bluvstein, G. Semeghini, A. Omran, J.-G. Liu, R. Samajdar,et al., Science376, 1209 (2022)

    2022

  14. [14]

    Dalyac, L

    C. Dalyac, L. Leclerc, L. Vignoli, M. Djellabi, W. d. S. Coelho, B. Ximenez, A. Dareau, D. Dreon, V. E. Elfving, A. Signoles,et al., The European Physical Journal A60, 177 (2024)

    2024

  15. [15]

    D. A. Boiko, R. MacKnight, B. Kline, and G. Gomes, Nature624, 570 (2023)

    2023

  16. [16]

    A. M. Bran, S. Cox, O. Schilter, C. Baldassari, A. D. White, and P. Schwaller, Nature Machine Intelligence6, 525 (2024)

    2024

  17. [17]

    C. Lu, C. Lu, R. T. Lange, J. Foerster, J. Clune, and D. Ha, arXiv preprint (2024), arXiv:2408.06292 [cs.AI]

    Pith/arXiv arXiv 2024

  18. [18]

    Zhenget al., arXiv preprint (2025), arXiv:2505.13259 [cs.AI]

    T. Zhenget al., arXiv preprint (2025), arXiv:2505.13259 [cs.AI]

    arXiv 2025

  19. [19]

    Ghafarollahi and M

    A. Ghafarollahi and M. J. Buehler, Proceedings of the National Academy of Sciences122, e2414074122 (2025)

    2025

  20. [20]

    Ghafarollahi and M

    A. Ghafarollahi and M. J. Buehler, Digital Discovery3, 1389 (2024)

    2024

  21. [21]

    Tomet al., Chemical Reviews124, 9633 (2024)

    G. Tomet al., Chemical Reviews124, 9633 (2024). 9

    2024

  22. [22]

    Simet al., Matter7, 2959 (2024)

    M. Simet al., Matter7, 2959 (2024)

    2024

  23. [23]

    Mandalet al., Nature Communications16, 9104 (2025)

    I. Mandalet al., Nature Communications16, 9104 (2025)

    2025

  24. [24]

    Reueret al., Nature Communications14, 7138 (2023)

    K. Reueret al., Nature Communications14, 7138 (2023)

    2023

  25. [25]

    Sivaket al., arXiv preprint (2025), arXiv:2511.08493 [quant-ph]

    V. Sivaket al., arXiv preprint (2025), arXiv:2511.08493 [quant-ph]

    Pith/arXiv arXiv 2025

  26. [26]

    Jernet al., inIEEE International Conference on Quantum Computing and Engineering (QCE)(2025)

    L. Jernet al., inIEEE International Conference on Quantum Computing and Engineering (QCE)(2025)

    2025

  27. [27]

    Asifet al., arXiv preprint (2025), arXiv:2503.02497 [quant-ph]

    H. Asifet al., arXiv preprint (2025), arXiv:2503.02497 [quant-ph]

    Pith/arXiv arXiv 2025

  28. [28]

    Mikuriyaet al., arXiv preprint (2025), arXiv:2510.26101 [cs.CL]

    T. Mikuriyaet al., arXiv preprint (2025), arXiv:2510.26101 [cs.CL]

    arXiv 2025

  29. [29]

    M. Tao, Y. Li, T. Yue, M. Zhang, and A. Arrieta Marcos, arXiv preprint (2026), arXiv:2607.00939 [cs.SE]

    Pith/arXiv arXiv 2026

  30. [30]

    Shiraishiet al., inInternational Conference on Quantum Communications, Networking, and Computing (QCNC)(2026)

    M. Shiraishiet al., inInternational Conference on Quantum Communications, Networking, and Computing (QCNC)(2026)

    2026

  31. [31]

    S. Arlt, X. Gu, and M. Krenn, Towards autonomous quantum physics research using LLM agents with access to intelligent tools (2025), arXiv:2511.11752 [cs.AI]

    arXiv 2025

  32. [32]

    Kumar, S

    S. Kumar, S. Datta, V. Singh, D. Datta, S. K. Singh, and R. Sharma, IEEE Access12, 75735 (2024)

    2024

  33. [33]

    X. Tang, Q. Jin, K. Zhu, T. Yuan, Y. Zhang, W. Zhou, M. Qu, Y. Zhao, J. Tang, Z. Zhang, A. Cohan, Z. Lu, and M. Gerstein, Nature Communications16, 8317 (2025)

    2025

  34. [34]

    Messeri and M

    L. Messeri and M. J. Crockett, Nature627, 49 (2024)

    2024

  35. [35]

    Huang, X

    J. Huang, X. Chen, S. Mishra, H. S. Zheng, A. W. Yu, X. Song, and D. Zhou, inInternational Conference on Learning Representations (ICLR)(2024)

    2024

  36. [36]

    Malinowski, Can AI answer open questions in physics?,https://m-malinowski.github.io/2026/06/ 05/ai-open-physics-problems.html(2026), accessed: June 11, 2026

    M. Malinowski, Can AI answer open questions in physics?,https://m-malinowski.github.io/2026/06/ 05/ai-open-physics-problems.html(2026), accessed: June 11, 2026

    2026

  37. [37]

    Silv´ erio, S

    H. Silv´ erio, S. Grijalva, C. Dalyac, L. Leclerc, P. J. Kar- alekas, N. Shammah, M. Beji, L.-P. Henry, and L. Hen- riet, Quantum6, 629 (2022)

    2022

  38. [38]

    Bidzhiev, S

    K. Bidzhiev, S. Grava, P. le Henaff, M. Mendizabal Pico, E. Merhej, and A. Quelle, arXiv preprint (2025), arXiv:2510.09813 [quant-ph]

    arXiv 2025

  39. [39]

    Pasqal, pasqal-cloud: Python SDKs to interact with Pasqal Cloud Services,https://github.com/ pasqal-io/pasqal-cloud(2026), accessed: 2026-06-10

    2026

  40. [40]

    Wennersteen, K

    A. Wennersteen, K. Bidzhiev, M. D’Arcangelo, M. Moreau, A. Quelle, A. Dauphin, and M. Beji, inQuantum Engineering Sciences and Technologies for Industry and Services (QUEST-IS 2025), Communica- tions in Computer and Information Science, Vol. 2744, edited by F. Barbaresco and F. Gerin (Springer, Cham,

    2025

  41. [41]

    Keesling, A

    A. Keesling, A. Omran, H. Levine, H. Bernien, H. Pich- ler, S. Choi, R. Samajdar, S. Schwartz, P. Silvi, S. Sachdev, P. Zoller, M. Endres, M. Greiner, V. Vuleti´ c, and M. D. Lukin, Nature568, 207 (2019)

    2019

  42. [42]

    Guoet al., Order-by-disorder and emergent Kosterlitz– Thouless phase in a triangular Rydberg array (2023), arXiv:2302.08963 [cond-mat.quant-gas]

    S. Guoet al., Order-by-disorder and emergent Kosterlitz– Thouless phase in a triangular Rydberg array (2023), arXiv:2302.08963 [cond-mat.quant-gas]

    Pith/arXiv arXiv 2023

  43. [43]

    Scholl, M

    P. Scholl, M. Schuler, H. J. Williams, A. A. Eberharter, D. Barredo, K.-N. Schymik, V. Lienhard, L.-P. Henry, T. C. Lang, T. Lahaye, A. M. L¨ auchli, and A. Browaeys, Nature595, 233 (2021)

    2021

  44. [44]

    Henry and L

    L.-P. Henry and L. Henriet, A method for estimating the chromatic number of a graph, PCT Patent Application No. PCT/EP2024/081690 (2024), assignee: PASQAL; priority EP 23306936.8, filed 8 November 2024

    2024

  45. [45]

    Vercellino, G

    C. Vercellino, G. Vitali, P. Viviani, E. Giusto, A. Scionti, A. Scarabosio, O. Terzo, and B. Montrucchio, in2023 IEEE International Conference on Quantum Comput- ing and Engineering (QCE), Vol. 2 (2023) pp. 141–147, arXiv:2605.03524

    Pith/arXiv arXiv 2023

  46. [46]

    rydberg atom array

    C. Jurczak, QOuLiPo: What a quantum computer sees when it reads a book (2026), arXiv:2605.14188 [cs.CL]. ACKNOWLEDGMENTS The hardware results reported here were obtained on the Pasqal QPUs FC1 and SA1 through standard Pasqal Cloud accounts available to external users. FC1 is lo- cated in Sherbrooke, in the DistriQ quantum-innovation zone (Qu´ ebec, Canada...

    Pith/arXiv arXiv 2026