Pith. sign in

REVIEW 3 major objections 3 minor 65 references

Quantum circuits as a game: A reinforcement learning agent for quantum compilation and its application to reconfigurable neutral atom arrays

T0 review · 3 major / 3 minor · reviewed 2026-08-07 · deepseek-v4-flash

Pith's one-line read A reinforcement-learning agent learns to reconfigure atom arrays during circuit execution, reducing logarithmic infidelity across benchmarks up to 100 qubits.

desk verdict A genuine RL-based move synthesis method for neutral atom arrays with a real transferability test, but the headline 20% reduction is measured in the paper's own unvalidated proxy, and a factor-of-10 inconsistency in a key parameter needs fixing. read the letter →

arxiv 2506.05536 v1 pith:DHQAFVYG submitted 2025-06-05 quant-ph

classification quant-ph PACS 03.67.Lx
keywords reinforcementlearningquantumcompilationneutralatomarraysreconfigurablemovesynthesislayoutreconfigurationtransformerpolicycircuitfidelity
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper tries to establish that compiling a quantum circuit for reconfigurable neutral atom arrays can be framed as a game and solved by a reinforcement learning agent. In this game, each move is a choice of where to place atoms in the storage region before each round of parallel two-qubit gates, and the reward is a proxy for logarithmic infidelity. The agent, called QC-Daemon, is a transformer policy trained with proximal policy optimization, and the paper reports cost reductions of up to about 20% across benchmarks up to 100 qubits. It also reports that the policy transfers to unseen circuits. A sympathetic reader would care because atom-array hardware can dynamically change qubit positions, and a learned policy could automate a difficult combinatorial planning step.

What carries the argument

The QC-Daemon policy is an autoregressive transformer that decides, for each playable atom in turn, which grid trap it should occupy next. Static features come from an MLP-Mixer over the current layout; dynamic features come from a Gate Transformer, which embeds the gate chunk containing the current atom and nearby future chunks, and a Move Transformer, which embeds already-planned moves; attention masks force information to flow from atoms to grid positions. The reward is built from a conflict graph that encodes the crossed-AOD constraints (many-to-one and ordering), with the number of moves estimated by a divide-and-conquer heuristic whose depth is logarithmic in the graph size. That estimate feeds the cost $J(D,M)=\alpha DN + \beta M$, where $D$ counts move duration, $M$ counts atom touches, $\alpha$ is the inverse coherence time, and $\beta$ is the per-touch loss.

What would settle it

Run a benchmark circuit (for example, the 100-qubit QFT instance) on a zoned reconfigurable atom array twice—once with the QC-Daemon's reconfigurations and once without—and compare measured process infidelities; the central claim fails if the agent's layout does not reduce the measured infidelity.

Watch

Extended reading notes

Core claim

The paper's central claim is that the Atom Game can be solved by a transformer-based reinforcement learning agent: given a circuit broken into parallel two-qubit gate chunks, the agent plans the next storage layout for each atom, and this planning lowers the logarithmic-infidelity proxy $J(D,M)=\alpha DN + \beta M$ by up to about 20% relative to no reconfiguration on QFT, QNN, QAOA, VQE, random, and Fermi-Hubbard circuits with up to 100 qubits. The paper also claims that the learned policy generalizes: a model trained on 30 random Hamiltonian-evolution circuits achieves positive cost reductions on unseen circuits of 40–50 and 80–100 qubits.

Load-bearing premise

The headline numbers are computed from the paper's cost model $J(D,M)=\alpha DN + \beta M$, not from a real device; if that model does not track actual move and gate errors, the reported reductions will not translate to real fidelity gains.

Editorial extensions

If this is right

  • If the transferability claim holds, a single pretrained policy can generate layout moves for new circuits without per-circuit training, which removes a major computational bottleneck of RL-based compilation.
  • If the proxy reduction is real, neutral-atom compilers can treat reconfiguration as an optimizable planning stage rather than a fixed constraint and combine it with existing gate scheduling.
  • The modular separation of the gate cost $G$ and layout cost $L$ means the same agent can be adapted to different hardware by swapping in new cost estimators.
  • The extension to logical qubits suggests the same game formulation could plan lattice-surgery or braiding moves in fault-tolerant topological codes.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Editorial inference: Because all reported gains are measured in the proxy $J$, the claim that real circuits would run better is conditional on $J$'s fidelity; a device-level experiment is the missing test.
  • Editorial inference: The transfer experiments sample test circuits from the same random-Hamiltonian family used in training, so the demonstrated generalization is within-distribution; cross-family transfer (for example, training on Hamiltonians and testing on QFT or QAOA) is not yet shown.
  • Editorial inference: The cost estimator assumes an effectively infinite swap region and that a greedy MaxCut partitioning always succeeds; a finite swap region or adversarial layouts could make the true move count larger than the estimated $n_m$, shrinking the real gains.
  • Editorial inference: A stronger validation would compare the QC-Daemon against classical scheduling heuristics on the same cost function to quantify how much of the reduction comes from RL rather than from the look-ahead formulation.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 3 minor

Summary. The paper introduces QC-Daemon, a reinforcement learning agent for the Atom Game, a move-synthesis problem on reconfigurable neutral atom arrays with a zoned storage/gate architecture. The authors formalize the compilation problem as an MDP (the QC-Game), define a cost proxy J(D,M) = αDN + βM for logarithmic infidelity (Eq. 8), and train two transformer-based policies with PPO. Experiments on benchmarks up to 100 qubits report cost reductions of up to about 20% relative to a no-reconfiguration baseline, and transfer experiments on 40–50 and 80–100 qubit circuits show positive reductions of roughly 6–10% on unseen circuits.

Significance. If the proxy accurately reflects execution infidelity, the paper would be a meaningful step toward learned compilation for reconfigurable neutral atom arrays. The formal MDP framing, the physically motivated transformer architecture, the multi-benchmark evaluation with multiple seeds, and the transferability experiments are genuine strengths. The public reconfiguration-cost estimator code is also a useful artifact. However, the central quantitative claim is currently established only in an unvalidated proxy, with an internal parameter inconsistency, and without comparison to existing reconfiguration compilers; these gaps substantially limit the significance of the headline 20% reduction.

major comments (3)
  1. [§V A and Appendix A, Table IV] The physical parameter β is inconsistent: Section V A states that Eq. (8) is evaluated with α = 0.02 and β = 0.002, while Appendix A, Table IV lists β = 0.02. Since every reported reduction is a function of J(D,M) in Eq. (8), a factor-of-10 change in the per-touch move-error term changes the trade-off between idling error and move error and can alter the learned policy. The authors must state which value was actually used and, ideally, report sensitivity of the headline reductions to β.
  2. [Abstract, §V, and §VI] The abstract claims a 'reduction of the logarithmic infidelity,' but the measured quantity is the proxy J(D,M) = αDN + βM, with D and M themselves estimated from the log-depth conflict-graph heuristic (Sec. III C). Section VI explicitly states that the reward is only a rough estimate and that no full simulation of scheduling or device execution was performed. The central claim should be reworded to refer to the proxy cost, and ideally supported by a pulse-level scheduler simulation or a hardware comparison, before it is presented as an infidelity reduction.
  3. [§V, Figs. 5 and Tables I–II] The evaluation compares only against a no-reconfiguration baseline and an untrained model; it does not compare against existing reconfiguration compilers and schedulers, such as Refs. [20,31–37], which solve closely related problems. Without such a comparison, the practical advantage of the RL approach over current state-of-the-art move synthesis is not established, even if the proxy is accepted.
minor comments (3)
  1. [§III C 2] The claim that the iteration exponent δ is 'irrelevant to the action of the QC-Daemon' is not strictly correct: because G(st, Ct) in Eq. (11) includes the transfer time TG, which is not multiplied by δ, the relative weight of the move cost versus the inter-zone transfer cost does depend on δ. The text should qualify this statement.
  2. [Fig. 5] The vertical-axis tick labels appear to omit minus signs for the negative reduction values (e.g., '-60', '-40', '-20'), which makes the early-training behavior harder to read.
  3. [§V B] The term 'n-shot' is defined in the text, but Tables I–II would be clearer if the caption or table header explicitly repeated that the reported value is the best over n independent runs of the trained model.

Circularity Check

1 steps flagged · score 4.0 of 10

The reported log-infidelity reduction is measured in the same proxy J that defines the RL reward; the paper itself disclaims device-level validation.

  1. self definitional [Sec. III B Eq. (7); Sec. III C Eq. (8); Sec. V A (Fig. 5); Abstract]
    "For each chunk Ct, we can associate a cost—or, equivalently, a reward—that reflects the log infidelity of executing the underlying AOD moves: J (D, M) = αDN + βM. ... With these cost functions, we define the reward at time t as R(st, st+1, Ct) := −L(st, st+1) − G(st+1, Ct) + G(s0, Ct). ... We see that ... we can achieve a reduction in cost of up to about 20% through layout changes."

    By Eqs. (9)-(10), both L and G in the reward are evaluations of the same J defined in Eq. (8), which the paper equates with the log infidelity of the AOD moves. The 'cost reduction' reported in Fig. 5 and Tables I-II is exactly the cumulative reward R that PPO maximizes, relative to the no-reconfiguration baseline G(s0,Ct). Therefore the abstract's claim of 'a reduction of the logarithmic infidelity' is, by the paper's own construction, a statement about the optimized proxy objective rather than an independently measured fidelity.

full rationale

The paper's central quantitative claim does not fully reduce to a fit, because the RL agent genuinely learns policies that improve the proxy relative to no reconfiguration, and the transfer experiment shows generalization to unseen circuits. The step above is flagged because the headline 'reduction of logarithmic infidelity' is definitionally tied to J(D,M)=αDN+βM, which is also the reward being optimized; the claim is thus partially self-definitional. This is a validation weakness rather than a logical contradiction, and the paper is transparent about it in Section VI. The internal β discrepancy (Sec. V A sets β=0.002 while Appendix A Table IV lists β=0.02) is a correctness risk, not a circularity issue. The heuristic from Ref. [38], cited with a co-author overlap, is an externally published method with stated assumptions and provided code, so it does not constitute load-bearing self-citation in the circularity sense. The paper is benchmarked on MQT Bench circuits, but the reduction numbers are all in the proxy, not in an external device simulation. Overall, partial circularity: score 4.

Assumptions & free parameters 8 free parameters · 6 assumptions · 0 invented entities

The reward model and the conflict-graph scheduler carry several free parameters and unverified assumptions. Most importantly, the reward is both the optimization objective and the evaluation metric, and the paper does not validate it against a real device. No new physical entities are introduced.

free parameters (8)
  • α (inverse coherence time) = 0.02
    Weights the idling error term in the reward J = αDN + βM. Chosen by hand in Section V A; no sensitivity analysis.
  • β (atom move loss) = 0.002 in Section V A, 0.02 in Appendix A (inconsistent)
    Weights the move error term in the reward. The paper gives two different values, which affects the landscape the agent optimizes.
  • ϵ (touches per move) = 0.5
    Assumes half the atoms are moved per step when estimating total touches ML = ϵ n_m |A|. Stated in Section III C.
  • δ (log-depth reconfiguration exponent) = ~1
    Assumed order one scaling factor for the move count estimate n_m = δ log |V(Gc)|. From Section III C 2.
  • γ (acceleration constant) = 1
    Appears in characteristic move time τ = sqrt(max distance / γ); set to 1 in the dimensionless simulation. Appendix A.
  • TG (inter-zone transfer time) = 10
    Added to gate duration DG = DL + TG; chosen in Appendix A without justification.
  • Grid size per benchmark = 4x10 to 10x20 depending on benchmark
    Grid dimensions are chosen per circuit and affect the action space and attainable cost reduction. Listed in Appendix A.
  • Window size W and horizon length K = W=2, K=5
    Model hyperparameters that determine how much look-ahead and which atoms are playable; chosen without ablation. Appendix A.
assumptions (6)
  • domain assumption The circuit can be decomposed into chunks of parallel two-qubit gates; one-qubit gates can be ignored for move synthesis.
    Section II A states the pre-processing step and says one-qubit gates are ignored because they are local and higher fidelity.
  • domain assumption Gates within each chunk commute and act on independent qubits.
    Section II A defines chunks this way, which is required for one-time-step parallel execution.
  • domain assumption The transition function P is deterministic.
    Section II A explicitly says 'we let P be a deterministic function in our simulations'.
  • domain assumption The conflict graph captures all relevant crossed AOD constraints (many-to-one and ordering).
    Section III C 1 defines the two constraint classes and uses them to build the move scheduler model.
  • domain assumption The swap region is large enough that the landscape is convex and greedy MaxCut always succeeds.
    Section III C 2 explicitly says 'Here we simply assume that the swap region is large enough so that the landscape is convex and a greedy MaxCut heuristic always succeeds.' This supports the logarithmic move count estimate.
  • ad hoc to paper The reward proxy J = αDN + βM represents logarithmic infidelity.
    The cost model is introduced in Section III C and used as both training reward and evaluation metric; it is not derived from a device calibration or validated against experiment.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Quantum circuits as a game: A reinforcement learning agent for quantum compilation and its application to reconfigurable neutral atom arrays." pith.science (2026). https://pith.science/paper/DHQAFVYG

@misc{pith2026250605536,
  author       = {Pith},
  title        = {Pith review of: Quantum circuits as a game: A reinforcement learning agent for quantum compilation and its application to reconfigurable neutral atom arrays},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/DHQAFVYG}},
  note         = {Machine review of arXiv:2506.05536}
}
read the original abstract

We introduce the "quantum circuit daemon" (QC-Daemon), a reinforcement learning agent for compiling quantum device operations aimed at efficient quantum hardware execution. We apply QC-Daemon to the move synthesis problem called the Atom Game, which involves orchestrating parallel circuits on reconfigurable neutral atom arrays. In our numerical simulation, the QC-Daemon is implemented by two different types of transformers with a physically motivated architecture and trained by a reinforcement learning algorithm. We observe a reduction of the logarithmic infidelity for various benchmark problems up to 100 qubits by intelligently changing the layout of atoms. Additionally, we demonstrate the transferability of our approach: a Transformer-based QC-Daemon trained on a diverse set of circuits successfully generalizes its learned strategy to previously unseen circuits.

Figures

Figures reproduced from arXiv: 2506.05536 by the authors.

Figure 1
Figure 1. FIG. 1. Training and inference with the QC-Game. During training, QC-Daemon accumulates experiences from playing the [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. FIG. 2. The structure of the atom game on a zoned reconfig [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. FIG. 3 [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: FIG. 4. The implementation of the atom-by-atom policy (QC-Daemon) [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]
Figure 5
Figure 5. Figure 5: FIG. 5. Cost reduction (%) for each benchmark across training iterations, as achieved by our QC-Daemon in the experiment [PITH_FULL_IMAGE:figures/full_fig_p011_5.png]
Figure 6
Figure 6. Figure 6: FIG. 6. The number of reconfiguration actions for each benchmark at each iteration in the same experiment as Fig. [PITH_FULL_IMAGE:figures/full_fig_p011_6.png]
Figure 7
Figure 7. Figure 7: FIG. 7. The change in the mean reward for each iteration during the training of the transferable models with 40–50 qubit [PITH_FULL_IMAGE:figures/full_fig_p012_7.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

65 extracted references · 55 canonical work pages

  1. [1]

    From that subset, iden- tify all “active” columns and rows Ac = [ q {c(t) q , c(t+1) q }, A r = [ q {r(t) q , r(t+1) q } in that rearrangement

    Accumulate the active participants Given a layout change st 7→ st+1, identify all atoms q for which v(t) q ̸= v(t+1) q . From that subset, iden- tify all “active” columns and rows Ac = [ q {c(t) q , c(t+1) q }, A r = [ q {r(t) q , r(t+1) q } in that rearrangement. The active participants A are the atoms that are in one of the active rows or columns and ar...

  2. [2]

    6 For more details on estimating n(t) m , see Sec

    Estimate the number of moves Given the active participants A, the number of moves n(t) m is computed based on a log-depth recon- figuration heuristic [38] with infinite swap space. 6 For more details on estimating n(t) m , see Sec. III C 2

  3. [3]

    Here, we assume that half the atoms are moved per step ϵ = 0 .5

    Compute the total reconfiguration cost The total number of touches is equal to the number of moves times the number of active participants ML = ϵn(t) m |A| scaled by some constant ϵ ∼ 1 that estimates how many atoms are touched per move. Here, we assume that half the atoms are moved per step ϵ = 0 .5. The total duration is equal to the number of moves tim...

  4. [4]

    The conflict graph The Conflict Graph Gc is a tool to schedule moves un- der the crossed AOD constraints and is defined as follows. Given a layout change st 7→ st+1, each vertex is an atom participating in the change, with an edge between ver- tices if the two moves cannot be done in parallel due to violating the crossed AOD constraints (cf Fig. 3). There...

  5. [5]

    carved out

    Estimating the number of moves n(t) m A simple parallel move scheduler may be implemented using a vertex coloring of the conflict graph Gc. A vertex colouring partitions k subsets of vertices such that every partition is an independent set with no two vertices shar- ing an edge. If a subset of vertices on the conflict is an independent set, it can be done...

  6. [6]

    They are then concatenated into a vector e: e = et ⊕ ea ⊕ el (15) Finally, a multilayer perceptron is applied to e to obtain d := {dj}ngrid j=1

    Static feature extraction The input of the static feature extraction is converted into a vector embedding: t → et, atom id → ea, and (v′ q)N q=1 → el. They are then concatenated into a vector e: e = et ⊕ ea ⊕ el (15) Finally, a multilayer perceptron is applied to e to obtain d := {dj}ngrid j=1 . For building el, we use the MLP-Mixer [50]. The MLP-Mixer is...

  7. [7]

    The output from those components is combined and processed in the readout layer

    Dynamic feature extraction Two types of Transformers, the Gate Transformer and the Move Transformer, are used to extract dynamic fea- tures. The output from those components is combined and processed in the readout layer. a. Gates Transformer In each chunk, we consider the gate containing the current playable atom (PA-gate) along with other gates (non-PA-...

  8. [8]

    We also describe how the value function is constructed for the actor-critic type RL algorithm

    The QC-Daemon ’s policy and its value function F and d are used to calculate the policy (QC-Daemon). We also describe how the value function is constructed for the actor-critic type RL algorithm. a. Policy (QC-Daemon) For each position, MLP is applied to fj for each j, and converted to one-dimensional output oj. The logit is defined by wj := oj +dj. The g...

Show all 65 references
  1. [9]

    Parallel implementation of high-fidelity multiqubit gates with neutral atoms,

    H. Levine, A. Keesling, G. Semeghini, A. Omran, T. T. Wang, S. Ebadi, H. Bernien, M. Greiner, V. Vuleti´ c, H. Pichler, and M. D. Lukin, “Parallel implementation of high-fidelity multiqubit gates with neutral atoms,” Phys. Rev. Lett., vol. 123, p. 170503, Oct 2019

  2. [10]

    Ghallab, D

    M. Ghallab, D. Nau, and P. Traverso, Automated Plan- ning: theory and practice . Elsevier, 2004

  3. [11]

    G. C. Pflug and A. Pichler, Multistage stochastic opti- mization, vol. 1104. Springer, 2014

  4. [12]

    Pddl— the planning domain def- inition language,

    C. Aeronautiques, A. Howe, C. Knoblock, I. D. McDer- mott, A. Ram, M. Veloso, D. Weld, D. W. Sri, A. Barrett, D. Christianson, et al., “Pddl— the planning domain def- inition language,” Technical Report, Tech. Rep., 1998

  5. [13]

    The formal semantics of processes in pddl,

    D. McDermott, “The formal semantics of processes in pddl,” in Proc. ICAPS Workshop on PDDL , pp. 101– 155, Citeseer, 2003

  6. [14]

    Reinforcement learning: An introduc- tion,

    R. S. Sutton, “Reinforcement learning: An introduc- tion,” A Bradford Book, 2018

  7. [15]

    An introduction to deep rein- forcement learning,

    V. Fran¸ cois-Lavet, P. Henderson, R. Islam, M. G. Belle- mare, J. Pineau, et al. , “An introduction to deep rein- forcement learning,” Foundations and Trends ® in Ma- chine Learning, vol. 11, no. 3-4, pp. 219–354, 2018

  8. [16]

    Mastering the game of go with deep neural networks and tree search,

    D. Silver, A. Huang, C. J. Maddison, A. Guez, L. Sifre, G. Van Den Driessche, J. Schrittwieser, I. Antonoglou, V. Panneershelvam, M. Lanctot, et al. , “Mastering the game of go with deep neural networks and tree search,” nature, vol. 529, no. 7587, pp. 484–489, 2016

  9. [17]

    Human-level control through deep reinforcement learning,

    V. Mnih, K. Kavukcuoglu, D. Silver, A. A. Rusu, J. Ve- ness, M. G. Bellemare, A. Graves, M. Riedmiller, A. K. Fidjeland, G. Ostrovski, et al. , “Human-level control through deep reinforcement learning,” nature, vol. 518, no. 7540, pp. 529–533, 2015

  10. [18]

    Artificial intelligence for quantum com- puting,

    Y. Alexeev, M. H. Farag, T. L. Patti, M. E. Wolf, N. Ares, A. Aspuru-Guzik, S. C. Benjamin, Z. Cai, Z. Chandani, F. Fedele, et al., “Artificial intelligence for quantum com- puting,” arXiv preprint arXiv:2411.09131 , 2024

  11. [19]

    Logical quantum processor based on reconfigurable atom arrays,

    D. Bluvstein, S. J. Evered, A. A. Geim, S. H. Li, H. Zhou, T. Manovitz, S. Ebadi, M. Cain, M. Kali- nowski, D. Hangleiter, et al., “Logical quantum processor based on reconfigurable atom arrays,” Nature, vol. 626, no. 7997, pp. 58–65, 2024

  12. [20]

    High-fidelity parallel entan- gling gates on a neutral-atom quantum computer,

    S. J. Evered, D. Bluvstein, M. Kalinowski, S. Ebadi, T. Manovitz, H. Zhou, S. H. Li, A. A. Geim, T. T. Wang, N. Maskara, et al. , “High-fidelity parallel entan- gling gates on a neutral-atom quantum computer,” Na- ture, vol. 622, no. 7982, pp. 268–272, 2023

  13. [21]

    Quantum error correction below the surface code threshold,

    G. A. team, “Quantum error correction below the surface code threshold,” Nature, vol. 638, p. 920–926, Dec. 2024

  14. [22]

    Experiments with the four-dimensional surface code on a quantum charge-coupled device quantum computer,

    N. Berthusen, J. Dreiling, C. Foltz, J. P. Gaebler, T. M. Gatterman, D. Gresh, N. Hewitt, M. Mills, S. A. Moses, B. Neyenhuis, P. Siegfried, and D. Hayes, “Experiments with the four-dimensional surface code on a quantum charge-coupled device quantum computer,” Physical Re- vie...

  15. [23]

    Scaling and logic in the color code on a superconducting quantum processor,

    N. Lacroix, A. Bourassa, F. J. Heras, L. M. Zhang, J. Bausch, A. W. Senior, T. Edlich, N. Shutty, V. Sivak, A. Bengtsson, et al. , “Scaling and logic in the color code on a superconducting quantum processor,” arXiv preprint arXiv:2412.14256, 2024

  16. [24]

    Scaling and network- 14 ing a modular photonic quantum computer,

    H. Aghaee Rad, T. Ainsworth, R. Alexander, B. Al- tieri, M. Askarani, R. Baby, L. Banchi, B. Baragiola, J. Bourassa, R. Chadwick, et al. , “Scaling and network- 14 ing a modular photonic quantum computer,” Nature, pp. 1–8, 2025

  17. [25]

    Optimality study of existing quan- tum computing layout synthesis tools,

    B. Tan and J. Cong, “Optimality study of existing quan- tum computing layout synthesis tools,” IEEE Transac- tions on Computers , vol. 70, no. 9, pp. 1363–1373, 2021

  18. [26]

    Using reinforcement learning to perform qubit routing in quantum compilers,

    M. G. Pozzi, S. J. Herbert, A. Sengupta, and R. D. Mullins, “Using reinforcement learning to perform qubit routing in quantum compilers,” ACM Transactions on Quantum Computing , vol. 3, p. 1–25, May 2022

  19. [27]

    Optimality study of existing quan- tum computing layout synthesis tools,

    B. Tan and J. Cong, “Optimality study of existing quan- tum computing layout synthesis tools,” IEEE Transac- tions on Computers , vol. 70, no. 9, pp. 1363–1373, 2020

  20. [28]

    Optimized compilation of aggregated instructions for realistic quantum comput- ers,

    Y. Shi, N. Leung, P. Gokhale, Z. Rossi, D. I. Schuster, H. Hoffmann, and F. T. Chong, “Optimized compilation of aggregated instructions for realistic quantum comput- ers,” in Proceedings of the Twenty-Fourth International Conference on Architectural Support for Programming Lan...

  21. [29]

    Compilation for dy- namically field-programmable qubit arrays with efficient and provably near-optimal scheduling,

    D. B. Tan, W.-H. Lin, and J. Cong, “Compilation for dy- namically field-programmable qubit arrays with efficient and provably near-optimal scheduling,” arXiv preprint arXiv:2405.15095, 2024

  22. [30]

    Elementary gates for quantum compu- tation,

    A. Barenco, C. H. Bennett, R. Cleve, D. P. DiVincenzo, N. Margolus, P. Shor, T. Sleator, J. A. Smolin, and H. Weinfurter, “Elementary gates for quantum compu- tation,” Physical review A , vol. 52, no. 5, p. 3457, 1995

  23. [31]

    Efficient discrete approximations of quantum gates,

    A. W. Harrow, B. Recht, and I. L. Chuang, “Efficient discrete approximations of quantum gates,” Journal of Mathematical Physics , vol. 43, no. 9, pp. 4445–4451, 2002

  24. [32]

    The solovay-kitaev algorithm,

    C. M. Dawson and M. A. Nielsen, “The solovay-kitaev algorithm,” arXiv preprint quant-ph/0505030 , 2005

  25. [33]

    Quan- tum compiling by deep reinforcement learning,

    L. Moro, M. G. Paris, M. Restelli, and E. Prati, “Quan- tum compiling by deep reinforcement learning,” Commu- nications Physics, vol. 4, no. 1, p. 178, 2021

  26. [34]

    Quantum circuit optimization with deep reinforcement learning,

    T. F¨ osel, M. Y. Niu, F. Marquardt, and L. Li, “Quantum circuit optimization with deep reinforcement learning,” arXiv preprint arXiv:2103.07585 , 2021

  27. [35]

    Practical and efficient quantum circuit synthesis and transpiling with reinforcement learning,

    D. Kremer, V. Villar, H. Paik, I. Duran, I. Faro, and J. Cruz-Benito, “Practical and efficient quantum circuit synthesis and transpiling with reinforcement learning,” arXiv preprint arXiv:2405.13196 , 2024

  28. [36]

    An abstract model and efficient routing for logical entan- gling gates on zoned neutral atom architectures,

    Y. Stade, L. Schmid, L. Burgholzer, and R. Wille, “An abstract model and efficient routing for logical entan- gling gates on zoned neutral atom architectures,” in 2024 IEEE International Conference on Quantum Computing and Engineering (QCE) , p. 784–795, IEEE, Sept. 2024

  29. [37]

    A quantum processor based on coherent transport of entangled atom arrays,

    D. Bluvstein, H. Levine, G. Semeghini, T. T. Wang, S. Ebadi, M. Kalinowski, A. Keesling, N. Maskara, H. Pichler, M. Greiner, et al. , “A quantum processor based on coherent transport of entangled atom arrays,” Nature, vol. 604, no. 7906, pp. 451–456, 2022

  30. [38]

    Dispersive optical systems for scalable raman driving of hyperfine qubits,

    H. Levine, D. Bluvstein, A. Keesling, T. T. Wang, S. Ebadi, G. Semeghini, A. Omran, M. Greiner, V. Vuleti´ c, and M. D. Lukin, “Dispersive optical systems for scalable raman driving of hyperfine qubits,” Phys. Rev. A, vol. 105, p. 032618, Mar 2022

  31. [39]

    Nmr tech- niques for quantum control and computation,

    L. M. K. Vandersypen and I. L. Chuang, “Nmr tech- niques for quantum control and computation,”Rev. Mod. Phys., vol. 76, pp. 1037–1069, Jan 2005

  32. [40]

    Optimal layout synthesis for quan- tum computing,

    B. Tan and J. Cong, “Optimal layout synthesis for quan- tum computing,” in Proceedings of the 39th International Conference on Computer-Aided Design , pp. 1–9, 2020

  33. [41]

    Qubit mapping for reconfigurable atom arrays,

    B. Tan, D. Bluvstein, M. D. Lukin, and J. Cong, “Qubit mapping for reconfigurable atom arrays,” in Proceed- ings of the 41st IEEE/ACM International Conference on Computer-Aided Design , ICCAD ’22, (New York, NY, USA), Association for Computing Machinery, 2022

  34. [42]

    Powermove: Optimizing compilation for neu- tral atom quantum computers with zoned architecture,

    J. Ruan, X. Fang, H. Zhang, A. Li, T. Humble, and Y. Ding, “Powermove: Optimizing compilation for neu- tral atom quantum computers with zoned architecture,” arXiv preprint arXiv:2411.12263 , 2024

  35. [43]

    Parallax: A compiler for neutral atom quantum computers under hardware constraints,

    J. Ludmir and T. Patel, “Parallax: A compiler for neutral atom quantum computers under hardware constraints,” in SC24: International Conference for High Performance Computing, Networking, Storage and Analysis , pp. 1–17, IEEE, 2024

  36. [44]

    Reuse-aware compi- lation for zoned quantum architectures based on neutral atoms,

    W.-H. Lin, D. B. Tan, and J. Cong, “Reuse-aware compi- lation for zoned quantum architectures based on neutral atoms,” in 2025 IEEE International Symposium on High Performance Computer Architecture (HPCA) , pp. 127– 142, IEEE, 2025

  37. [45]

    Surface code quantum computing by lattice surgery,

    D. Horsman, A. G. Fowler, S. Devitt, and R. Van Meter, “Surface code quantum computing by lattice surgery,” New Journal of Physics , vol. 14, no. 12, p. 123011, 2012

  38. [46]

    Routing- aware placement for zoned neutral atom-based quantum computing,

    Y. Stade, W.-H. Lin, J. Cong, and R. Wille, “Routing- aware placement for zoned neutral atom-based quantum computing,” arXiv preprint arXiv:2505.22715 , 2025

  39. [47]

    Constant-overhead fault-tolerant quantum computation with reconfigurable atom arrays,

    Q. Xu, J. P. Bonilla Ataides, C. A. Pattison, N. Raveen- dran, D. Bluvstein, J. Wurtz, B. Vasi´ c, M. D. Lukin, L. Jiang, and H. Zhou, “Constant-overhead fault-tolerant quantum computation with reconfigurable atom arrays,” Nature Physics, vol. 20, p. 1084–1090, Apr. 2024

  40. [48]

    https://github.com/QuEraComputing/ reconfiguration_cost_estimator

  41. [49]

    On colouring the nodes of a network,

    R. L. Brooks, “On colouring the nodes of a network,” Mathematical Proceedings of the Cambridge Philosophical Society, vol. 37, p. 194–197, Apr. 1941

  42. [50]

    Experimental demonstra- tion of logical magic state distillation,

    P. S. Rodriguez, J. M. Robinson, P. N. Jepsen, Z. He, C. Duckering, C. Zhao, K.-H. Wu, J. Campo, K. Bagnall, M. Kwon, et al. , “Experimental demonstra- tion of logical magic state distillation,” arXiv preprint arXiv:2412.15165, 2024

  43. [51]

    Fault-tolerant quantum computation by anyons,

    A. Y. Kitaev, “Fault-tolerant quantum computation by anyons,” Annals of physics , vol. 303, no. 1, pp. 2–30, 2003

  44. [52]

    Non-abelian anyons and topological quantum computation,

    C. Nayak, S. H. Simon, A. Stern, M. Freedman, and S. Das Sarma, “Non-abelian anyons and topological quantum computation,” Reviews of Modern Physics , vol. 80, no. 3, pp. 1083–1159, 2008

  45. [53]

    Fault-tolerant quan- tum computation with high threshold in two dimen- sions,

    R. Raussendorf and J. Harrington, “Fault-tolerant quan- tum computation with high threshold in two dimen- sions,” Physical review letters, vol. 98, no. 19, p. 190504, 2007

  46. [54]

    The other hyper-parameters are listed in Appendix A

    as the RL algorithm. The other hyper-parameters are listed in Appendix A. The training is performed on a system equipped with eight NVIDIA H100 Tensor Core GPU. In Fig. 5, we show the cost reduction ratio at each iteration for each benchmark problem. The horizontal axis repres...

  47. [55]

    A game of surface codes: Large-scale quan- tum computing with lattice surgery,

    D. Litinski, “A game of surface codes: Large-scale quan- tum computing with lattice surgery,” Quantum, vol. 3, p. 128, 2019

  48. [56]

    Opensurgery for topological assemblies,

    A. Paler and A. G. Fowler, “Opensurgery for topological assemblies,” 2020

  49. [57]

    Surface code compilation via edge-disjoint paths,

    M. Beverland, V. Kliuchnikov, and E. Schoute, “Surface code compilation via edge-disjoint paths,” PRX Quan- tum, vol. 3, no. 2, p. 020342, 2022. 15

  50. [58]

    A high performance compiler for very large scale surface code computations,

    G. Watkins, H. M. Nguyen, K. Watkins, S. Pearce, H.- K. Lau, and A. Paler, “A high performance compiler for very large scale surface code computations,” Quantum, vol. 8, p. 1354, 2024

  51. [59]

    Mlp-mixer: An all-mlp architecture for vision,

    I. O. Tolstikhin, N. Houlsby, A. Kolesnikov, L. Beyer, X. Zhai, T. Unterthiner, J. Yung, A. Steiner, D. Keysers, J. Uszkoreit, et al., “Mlp-mixer: An all-mlp architecture for vision,” Advances in neural information processing systems, vol. 34, pp. 24261–24272, 2021

  52. [60]

    Attention is all you need,

    A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, L. Kaiser, and I. Polosukhin, “Attention is all you need,” Advances in neural informa- tion processing systems, vol. 30, 2017

  53. [61]

    MQT Bench: Benchmarking software and design automa- tion tools for quantum computing,

    N. Quetschlich, L. Burgholzer, and R. Wille, “MQT Bench: Benchmarking software and design automa- tion tools for quantum computing,” Quantum, 2023. MQT Bench is available at https://www.cda.cit.tum. de/mqtbench/

  54. [62]

    Observation of separated dynamics of charge and spin in the fermi-hubbard model,

    F. Arute, K. Arya, R. Babbush, D. Bacon, J. C. Bardin, R. Barends, A. Bengtsson, S. Boixo, M. Broughton, B. B. Buckley, et al. , “Observation of separated dynamics of charge and spin in the fermi-hubbard model,” arXiv preprint arXiv:2010.07965, 2020

  55. [63]

    Proximal policy optimization algorithms,

    J. Schulman, F. Wolski, P. Dhariwal, A. Radford, and O. Klimov, “Proximal policy optimization algorithms,” arXiv preprint arXiv:1707.06347 , 2017. Appendix A: Hyper-parameter of the numerical experiment

  56. [64]

    Model architecture, training, and experimental parameters

    Machine learning model parameters TABLE IV. Model architecture, training, and experimental parameters. The device parameters are used to compute a dimen- sionless quantity, the infidelity. Therefore, while each parameter may originally have a physical unit, the units are omitt...

  57. [65]

    The list of the size of the grids in each benchmark experiment

    Size of the grids TABLE V. The list of the size of the grids in each benchmark experiment. The benchmark ’transfer’ represents the experiment in Section V B. Benchmark Qubits Grid size (row × col) qaoa 14 4 × 10 groundstate 14 4 × 10 random 30 7 × 10 qnn 50 5 × 20 qft 100 10 ×...

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.