acceptodds
Under review as a conference paper at ICLR 2027

Circuit-Frontier Representations for Quantum Architecture Search

Abstract

Hardware-aware quantum architecture search must connect logical operation dependencies with the physical resources used to execute a circuit. We introduce a circuit-frontier representation that maintains a global circuit summary and an ordered embedding for each logical wire. An autoregressive builder uses these embeddings to select operations. We study two placement strategies within the same construction framework: a learned hardware-attention pointer that reuses the embeddings, and a compiled-cost heuristic that maps the constructed circuit directly. This separation lets us test learned placement without assuming that both decisions must be neural policies. A balanced study covers six H/LiH/HO Hamiltonians, single- and ten-calibration training, three paired training blocks, and density-matrix simulation with native-gate and readout noise from IBM Fez calibrations. The learned-placement model has favorable aggregate policy-only point estimates against the learned controls under single-snapshot training, but several structural directions reverse under ten-snapshot training and no registered aggregate advantage survives correction. Retaining acquired circuits reveals a different comparison: the heuristic-placement learner's archive improves eight of twelve task–exposure means relative to its own frozen policy and ties the other four at reported precision, with reductions of 10.455 and 8.876 mHa on the two water tasks under ten-snapshot training. Source-matched remapping shows that archive usefulness depends on deployment placement, while a fixed policy–archive mixture does not improve the full model's normalized aggregate. The results identify circuit acquisition, placement, and candidate retention as distinct requirements for evaluating a learned circuit representation.

Primary area:
reinforcement learning
Keywords:
quantum architecture search, reinforcement learning, graph neural networks, qubit mapping, variational quantum eigensolver

Reviews

Among:
Shape:
1234this paper, Bottom 25%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.75 over 4 reviews, above 25% of the others.
Reviewer_45rfOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper proposes a "circuit-frontier representation" for quantum architecture search: it constructs a DAG of operations based on logical-line dependencies, maintains a "frontier embedding" for each logical line, and forms a global summary. Circuit construction strategies utilize these representations to iteratively select operations.

Critical strengths

Table 2 separately discusses the comprehensive strategy, distinct profiles under shared mapping rules, and mapping interventions for fixed profiles, thereby avoiding the conflation of these different issues.

The authors did not misrepresent task-level mean improvements or pointwise confidence intervals as a generalized, calibrated advantage; they also retained the observed direction reversals and performance degradation associated with mixed deployment under S10.

Critical weaknesses

Existing experiments do not separately isolate the removal of line-by-line readout, the removal of port labels, or the integration of existing circuit DAG encoders into the same training pipeline. Comparisons involving standard GRUs, hardware MLPs, independent encoders, and gradient blocking are insufficient to isolate the effects of these specific design additions.

Table 9 shows that, under a shared heuristic mapping, the overall archive mean for all seven configurations outperforms their respective final policies. Since the CF-H variant still relies on the PPO constructor, the paper does not address whether a simple search combined with archive retention, within the same computational budget would have sufficed.

Questions for authors

Could you include a comparison of encoders, specifically "global-only," "port-label-free," and approaches similar to He et al., and report the results from the archive under both the strategy and the joint heuristic mapping?

Could you add random or local search baselines using the same 1,024 acquisitions and the same upper limit on parameter optimization, while retaining eight archive circuits? At the very least, for H₂, could you provide a reference based on enumeration or the reachable optimum?

Is the direction reversal observed in S1/S10 robust? Could you add independently trained blocks to directly estimate the paired difference-in-differences and their associated uncertainties?

Does the improvement in energy correspond to a better quantum state? Could you provide additional data on the "gate-only expectation" of the returned circuits, the ideal noise-free energy, and diagnostics for particle number and spin sectors? An objective function that incorporates readout distortion cannot, on its own, validate state accuracy.

Reviewer_FesYOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents a paradigm for hardware-aware quantum architecture search. In detail, the authors introduce an autogressive framework to construct the circuit with: [1] CF-L, which uses a pre-trained hardware-attention pointer, and [2] CF-H, which uses a compile-cost heuristic. The constructed circuit is optimized via the proximal policy optimization (PPO) and the learning signal is set as molecular energy estimations. The work performs experiments on six molecular Hamiltonians involving H₂, LiH, and H₂O. The results suggest that the learnt construction can lead to good performance. In summary, this work presents a structured framework for constructing quantum circuits alongwith the empirical evidence of different roles of circuit representation.

Critical strengths

I am familiar with architecture search but not familiar with reinforcement learning. From what I can see, the strengths are as follows:

I think the targeted problem, i.e. circuit construction while preserving logical-wire dependencies, is quite clear and important for the NISQ era quantum computers. The proposed method captures [1] global circuit structures and [2] operation histories of individual logical wires. The adopted appraoch of using ordered, wire-specific emebeddnings, along with a global representation is quite natual thought between circuit construction and hardware assignment. The design makes sense and generally well-aligned with circuit structure. The experimental framework is good. This work includes evaluating several neural models for structure controls. This means different choices are available within a single framework which is good.

Critical weaknesses
  1. Pages in the main manuscript are not fully used and many details and results are deferred to the appendix. This means that the page budget of the main paper can have been used more effectively. Most important supporting evidence should be put in the main text rather than the appendix. This dirctly affects readbility of the paper.

  2. Then, after reading the related litertaure, it is not very clear to me how the proposed work is different from existing methods. There exist works that explored dependency-aware circuit DAG encodings and graph-based circuit generation. It seems to me that the proposed method integrates these components. Additionally, the performed experiments do not provide convincing evidence that the additional architectural complexity gives significantly improved search performance. The reported advantages over simpler encoders vary across settings. Therefore, given the experimental set-up, I have not been convinced that the paper gives a substantial advance towards hardware-aware quantum architecture search.

Questions for authors

See weakness.

Reviewer_W64nOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper investigates the interaction between circuit representations, physical qubit placement, and candidate retention in hardware-aware Quantum Architecture Search (QAS) by introducing a circuit-frontier representation that models circuits as port-labeled dependency DAGs. This framework extracts both a global circuit summary and explicit, ordered wire embeddings to guide an autoregressive builder, enabling a direct comparison between a learned GATv2 hardware-attention pointer (CF-L) and a deterministic compiled-cost heuristic (CF-H). Evaluated across six molecular Hamiltonians and 13 historical IBM Fez calibration snapshots under single-calibration (S1) and ten-calibration (S10) training regimes, the study demonstrates that CF-L’s policy-only point advantage under S1 disappears under broader calibration exposure, where structural trends reverse and no aggregate advantage survives Holm-Bonferroni correction. Crucially, candidate retention captures significant value that final policies overlook: CF-H’s archive improves upon or ties its own frozen policy across all twelve tasks (yielding energy reductions of up to 10.455 mHa on water). Meanwhile, the learned pointer struggles to generalize when deploying its own archive, but remapping those exact same archived circuits using the compiled heuristic restores performance, establishing that circuit acquisition, placement, and candidate retention represent distinct and decoupled requirements in quantum architecture search.

Critical strengths
  • The experimental design uses three independent paired training blocks (df=2df=2), reports 95% confidence intervals, and subjects all primary hypotheses to Holm-Bonferroni family-wise error rate control.

  • The paper provides a clear, valuable negative result: an end-to-end learned neural pointer fails on archived circuits, whereas an additive compiled-cost heuristic reliably redeploys those same candidates with lower energies. Invariance

  • Proposition 1 proves that the circuit-frontier representation is invariant under topological reindexing and independent event ordering, overcoming a key failure mode of sequence models like GRUs. Policy vs. Retention Separation: Demonstrating that candidate archives consistently outperform final frozen policies provides a necessary reality check for RL-QAS evaluation protocols.

Critical weaknesses
  • The construction horizon is restricted to at most 4 actions, and placement is chosen from an enumerated roster of 16 feasible subgraphs. This limits the generality of the conclusions for deep circuits and unconstrained hardware routing.

  • Relying on three paired blocks provides only 2 degrees of freedom. Combined with conservative Holm corrections, no aggregate policy contrast achieves significance (padj≥0.559p_{\text{adj}} \ge 0.559), leaving many architectural conclusions descriptive.

  • Density-matrix simulations model gate and readout errors but omit thermal relaxation (T1/T2T_1/T_2) and finite measurement shot noise, leaving the pointer's resilience to stochastic reward noise unverified.

  • The heuristic baseline is an additive sum of gate and readout errors; comparing against industrial compilation passes (such as noise-adaptive SABRE) would provide a stronger benchmark

Questions for authors
  • Does the performance gap between learned (CF-L) and heuristic (CF-H) placement widen or narrow when extending the search horizon beyond 4 actions?

  • How does the compiled-cost heuristic compare against standard Qiskit noise-adaptive routing passes on these subgraphs?

  • How stable are the terminal frontier readouts and GATv2 pointer scores when expectation values include finite sampling variance (e.g., 10310^3–10410^4 shots)?

Reviewer_nognOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes a graph representation to guide quantum circuit construction and qubit placement. It compares learned and heuristic placement and evaluates circuits retained during training. Results support retaining and remapping circuits but do not establish a consistent advantage for the proposed representation.

Critical strengths

The evaluation separates placement effects from the benefit of retaining circuits found during training.

Critical weaknesses
  • Several results reverse across training conditions, and no aggregate comparison remains significant after correction for multiple testing.
  • Results from two test dates had already been inspected during development, limiting the independence of the final evaluation.
  • The experiments lack direct comparisons with the closest existing architecture-search methods.
  • The ablations do not isolate the benefit of the separate representation for each qubit.
Questions for authors

See weaknesses above

Comments on the paper

1 comment

Hello Reviewers,

As the discussion period is approaching its end, please take a moment to review the authors’ responses. If you have any further comments please share either to authors or privately to other reviewers and me. And if you feel your judgment is refreshed, please also update your assessment if necessary.

Thanks for your thoughtful reviews and discussion!

Bests, AC

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.