acceptodds
Under review as a conference paper at ICLR 2027

MARSA: Joint Learning of Qubit Mapping and Routing with Verifiable Self-Improvement

Abstract

Executing a quantum circuit on a processor with limited connectivity requires choosing an initial qubit mapping and a sequence of routing operations. These decisions are often exposed separately, but they are fundamentally coupled: a mapping is useful only through the routing trajectories it enables, while the best route depends on the mapping from which it starts. We study this coupling directly with **MARSA** (**M**apping **A**nd **R**outing with a **S**emantic **A**utomaton), a single autoregressive policy that generates an initial mapping followed by the complete physical routing trace. This representation creates a structured-generation problem: the legality of every new token depends on the current logical-to-physical mapping, the hardware graph, and the unexecuted circuit. We therefore introduce a *route-state automaton* that maintains these semantics progressively and masks invalid actions during both inference and reinforcement learning rollouts. To learn this long-horizon structured policy efficiently, we initialize from a strong heuristic expert using a two-axis curriculum over circuit width and length, then move beyond imitation through *verifier-guided self-improvement*: MARSA generates its own complete solutions, exact verification evaluates their correctness and SWAP cost, and GRPO reinforces the better generations. On Q20 random circuits, MARSA strictly improves on the training expert for 81.12% of circuits, matches or improves it for 97.15%, and reduces aggregate SWAP count by 16.40%; it also transfers from synthetic training to held-out real quantum algorithm circuits, reducing aggregate SWAP count by 71.35% relative to Qiskit LightSABRE. A separately trained model on a second 2D grid coupling graph also achieves strong performance. Together, these results show that joint mapping and routing can be cast as structured generation, enabling a learned compiler to optimize both and continue improving through its own verified solutions.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Quantum circuit compilation, qubit mapping and routing, AI for quantum computing, reinforcement learning with verifiable rewards, curriculum learning

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_fdZoOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates joint initial qubit mapping and routing for quantum circuit compilation and proposes MARSA. Specifically, an autoregressive policy generates an initial mapping and a complete routing trace, with a route-state automaton enforcing admissible actions. Curriculum learning and verifier-guided reinforcement learning improve the policy beyond expert demonstrations. Theoretical analysis establishes feasibility of accepted traces under the stated routing model. Experiments on two hardware topologies and held-out algorithm circuits report reduced SWAP counts.

Critical strengths
  • Proposition 1 and separate route replay connect generated traces to dependency-consistent, hardware-adjacent circuit execution. The appendices distinguish accepted-route soundness from rollout completion, and justify eager execution and removal of redundant SWAP cycles within the stated objective.
  • Greedy decoding reports a 7.01% aggregate Q20 SWAP reduction over paired valid cases relative to QMSI; gains therefore do not arise solely from best-of-16 selection. Improvements on two topologies and held-out algorithm circuits broaden the evidence. Exact-match, contiguous-subcircuit, and cross-split duplicate exclusions also address direct overlap concerns.
Critical weaknesses
  • The central novelty is insufficiently differentiated and experimentally unisolated. Previous works [1, 2] have already explored learned quantum routing and masked allocation sequences, respectively. The baseline panel omits direct comparisons with AlphaRouter, QAP-Router, CO-MAP, or a matched separately trained mapper/router. Substituting QMSI mappings into the existing learned router changes its conditioning distribution. The degradation shows a benefit from MARSA-generated mappings with this router, without isolating the superiority of joint training. Moreover, the already cited AlphaRouter [1] explicitly proposed using initial virtual SWAPs to learn mapping and routing with the same model. The missing reference [2] already generates complete allocation sequences with dynamic feasibility masks, albeit for modular architectures. Neither source implements MARSA's exact mapping-prefix/full-operation trace; that narrower residual contribution needs attribution and a controlled demonstration of its benefit.
  • The principal teacher-relative gains are not fully auditable. Appendix D.1 omits QMSI's search budget, restarts, and stopping rule. The previous work [3] defers important expert details to an unpublished manuscript; it cannot substitute for an exact teacher specification. Aggregate reductions use paired valid cases without reporting the corresponding completion counts and paired-valid denominators, although total test-set sizes are given. Failures do count as losses in win/tie/loss statistics, so inflation is not established. Unspecified expert effort and potentially differing subsets leave uncertainty about the comparison. Report settings, coverage, and common-subset or fallback-inclusive comparisons. Runtime would additionally establish practical cost.

References

[1] AlphaRouter: Quantum Circuit Routing with Reinforcement Learning and Tree Search

[2] Attention-Based Deep Reinforcement Learning for Qubit Allocation in Modular Quantum Architectures

[3] Leveraging Machine Learning for Quantum Compilation Optimization

Questions for authors
  • Clarification: What QMSI settings, generation limits, candidate-completion rates, and paired-valid circuit counts underlie each comparison?
  • Requested analysis: If any circuits fail to complete, do reductions persist on one common subset, with failures reported separately?
  • Requested control: Does the advantage remain against a router trained on the supplied mappings and comparable learned systems under matched evaluation conditions?
  • Scope clarification: Which workload statistics defined the synthetic training mixture, and when were they frozen relative to evaluation selection?
  • Related-work clarification: Please situate route-state legality against the previous work [1], which already uses evolving qubit locations, gate dependencies and feasible-action masks for distributed compilation. Its different architecture/objective and future-only initial-mapping extension do not anticipate MARSA's full joint trace.

References

[1] Compiler for Distributed Quantum Computing: a Reinforcement Learning Approach

3 comments
AuthorsOpenReview ↗

Response to Reviewer fdZo (Part 1)

We thank the reviewer for the questions on novelty and comparisons. We clarify them as follows.

1. Novelty relative to previous work (W1 & Q5)

  • AlphaRouter (Tang et al., 2024): The original paper states: “AlphaRouter is designed to solve only the circuit routing problem for a given initial mapping.” It explicitly leaves jointly generating mapping and routing to future work, suggesting an initial sequence of virtual SWAPs only as a possible but unvalidated idea. MARSA is the first work to solve this previously unaddressed gap through joint representation and well-suited training recipe.
  • Russo et al. (2024; reviewer weakness reference [2]): (1) This paper addresses qubit-to-core allocation in modular architectures, a completely different problem from the SWAP-based routing discussed in our work. Its model sequentially allocates qubits to cores to minimize the inter-core transfer cost; it does not generate a physical SWAP/CNOT trace nor minimize SWAP count. (2) Its action masks enforce available core capacity, co-location of interacting qubits, and sufficient remaining capacity. These allocation-feasibility rules do not enforce the evolving gate-execution semantics used in MARSA's explicit routed circuit.
  • Promponas et al. (2024; reviewer question reference [1]): (1) The states and actions differ fundamentally. Their RL agent selects routing and communication actions from a given initial mapping, while executable CNOTs are handled by the environment. In contrast, MARSA jointly learns the initial mapping and explicitly generates the complete CNOT/SWAP routing trace. (2) Invalid action masking itself is a well-established idea, but the concrete constraint modeling differs significantly: Promponas et al. perform action-level masking, whereas MARSA performs token-level constrained generation, which enforces not only the evolving semantics of routing but also the grammar of the joint representation. (3) MARSA integrates cost-preserving pruning into the route-state automaton, such as elimination of redundant SWAP reversals, with formal guarantees of soundness for complete accepted traces and preservation of SWAP-optimal solutions. As the reviewer acknowledges, Promponas et al. adopt a different architecture and discuss optimization of initial mapping only as future work. They do not address joint generation of mapping and routing, which is precisely MARSA's central contribution.

2. Comparison with AlphaRouter and other baselines (W1)

We examined AlphaRouter, QAP-Router, and CO-MAP, but could not find publicly released repos for faithful reproduction. The papers identify benchmark sources or families, but do not provide the exact instances for their reported metrics. We therefore cannot fairly compare MARSA's results against aggregate numbers obtained on potentially different circuit subsets.

They also address different learned decisions: AlphaRouter and QAP-Router focus on routing with initial mapping given, whereas CO-MAP focuses on initial mapping only. None demonstrates MARSA's learned full mapping and routing generation. Our reported experiments instead compare on identical circuits against reproducible QMSI, LightSABRE, ML-SABRE, and AIRouting configurations.

3. Comparison against a router trained on the supplied mappings (W1 & Q3)

To further strengthen the evidence for joint training of mapping and routing, we trained a new routing-only policy from scratch on QMSI-supplied mappings. We refer the details to our Response 1 to Reviewer PMFf (Table 1). The routing-only model achieves only 6.61% SWAP reduction over QMSI, compared with 16.40% for joint MARSA on Q20. It uses 11.71% more SWAPs than MARSA. This controlled comparison addresses the distribution shift concern and provides stronger evidence for the value of jointly learning the initial mapping and routing.

AuthorsOpenReview ↗

Response to Reviewer fdZo (Part 2)

4. QMSI settings, completion rates, and common-subset analysis (W2 & Q1–Q2)

QMSI settings. We clarify the teacher configurations:

  • Q20: Depth-3 SWAP search, initial mapping search widths (8, 4, 2, 1, 1, ...).
  • Grid: Depth-2 SWAP search, initial mapping search widths (16, 4, 2, 1, 1, ...). The depth is smaller than Q20 because the topology is significantly larger so depth-3 search times out.

Both topologies use bidirectional mapping refinement and a 120-second per-call limit with one failure-only retry. There's no multiple search restart mechanism similar to LightSABRE in QMSI.

Completion rates and paired-valid comparisons. QMSI successfully generates valid references for all 1,822 Q20 and 2,738 grid evaluation circuits. MARSA generates candidates with a 4,096-token limit without fallback.

Table 6: QMSI and MARSA completion rates and win/tie/loss/failure counts.

TopologyTotal circuitsQMSI completedMARSA completedW / T / L / FMARSA failure rate
Q201,8221,8221,8221,478 / 292 / 52 / 00.000%
Grid2,7382,7382,7172,159 / 320 / 238 / 210.767%

MARSA achieves a very low circuit-level failure rate: 0% on Q20 and only 0.77% on the grid. All 21 grid failures are due to the generation length limit, not invalid completed routes. As the reviewer acknowledges, failures are counted as losses in the win/tie/loss statistics in the paper; there is no inflation of win rates.

Crucially, since QMSI completes every circuit, the reported SWAP reductions already use the exact common-valid subset. The paired results and denominators are:

Table 7: Aggregate SWAP counts and reductions on paired-valid evaluation circuits.

TopologyCommon-valid circuitsQMSI SWAPsMARSA SWAPsSWAP reduction
Q201,82283,86270,10916.40%
Grid2,717407,270364,21310.57%

The 21 incomplete grid cases are excluded from both SWAP sums and counted as losses. Thus, the reported gains are not inflated by differing QMSI and MARSA completion subsets. We have added the clarifications in the corresponding sections of the paper and presented this table as Table 4 in the revision. We refer runtime experiments including compilation-time scaling and matched-budget comparisons in our Response 1 to Reviewer LNzk.

5. Synthetic training distribution (Q4)

Our synthetic generator uses aggregate statistics including: (1) CNOT count band (2) qubit count band (3) pair repeat gap (4) interaction graph density (5) general circuit patterns such as star, path, brick-like structures, etc. The benchmark was fixed before generator development, and the synthetic training mixture and generator were frozen before model training and evaluation. Therefore model performance is not used to modify the workload statistics. Exact benchmark matches, 64-CNOT contiguous overlaps, and cross-split duplicates were excluded.

Dear Reviewer fdZo,

Thank you again for your valuable feedback and suggestions. We have provided additional experiments and clarifications in our rebuttal, and we hope these additions help address your concerns.

We would be grateful for any further feedback and would be happy to clarify any remaining questions or concerns.

Thank you very much for your time and consideration!

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

The paper presents a learned method for jointly placing logical qubits and routing their interactions on quantum processors with limited connectivity, aiming to minimise the number of SWAP operations required. And indeed results are presented that show the method achieves reductions in SWAP counts across two hardware topologies.

Critical strengths
  • Presents an apparently novel strategy for generating initial qubit placement and the complete routing trace with a single learned policy.

  • Constraints ensure that completed, accepted routes respect hardware connectivity and circuit dependencies.

  • Convincing experimental results demonstrate reduced SWAP counts across two hardware topologies, including on held-out circuits.

Critical weaknesses
  • Practical efficiency is unclear, especially since compilation time and its scaling with circuit size are not reported.

  • A separate model is trained for each hardware topology. The cost of adaptation and generalisation to other topologies unclear.

  • The experiments seem to show that learning the initial placement helps, but not that learning placement and routing together is better than learning them separately.

  • The relationship to established graph-based placement methods, including graph embedding and partitioning, is not clear.

Questions for authors
  • How does compilation time scale with circuit size, and how does MARSA compare with the baselines under similar time budgets?

  • On the basis of the currently presented experiments, how can we be sure that learning placement and routing jointly is better than learning them separately?

  • What is needed to adapt MARSA to a new topology? Complete re-training or just fine tuning?

  • Please clarify exact novelty over Li et al. (2019), Ren et al. (2024) and Tang et al. (2024).

4 comments
AuthorsOpenReview ↗

Response to Reviewer LNzk (Part 1)

We thank the reviewer for recognizing the novelty of generating mapping and routing with a single learned policy and the strong results across two hardware topologies. We address the questions on compilation time, joint learning, topology adaptation, and prior work below.

1. Compilation time and matched-budget comparisons

We measured MARSA's compilation time scaling on 500 held-out random circuits per topology, with 100 circuits in each CNOT-count interval. Experiments ran on a dual AMD EPYC 9655 host; each MARSA inference job used one NVIDIA RTX 5090 GPU and eight pinned physical CPU cores. Table 4 reports the median compilation time in seconds.

Table 4: MARSA compilation time scaling with circuit size.

Topology / CNOT count51–100101–150151–200201–250251–300
Q201.9373.1384.1115.4566.893
Grid25.57526.70527.77128.73129.669

Compilation time grows approximately linearly with circuit length over the evaluated range, demonstrating reasonable scaling. Moreover, our current implementation has not been extensively optimized for inference speed, leaving room for further engineering improvements.

For a more controlled comparison, we assigned each baseline a compilation-time budget matched to MARSA's measured runtime for each circuit. Table 5 reports both the original results and the new time-matched results.

Table 5: Aggregate SWAP counts under matched compilation-time budgets.

Topology / evaluationMARSAQMSILightSABREML-SABREAIRouting
Q20 — Original paper3,2103,65311,33411,27211,142
Q20 — Time-matched3,2103,4479,58010,29010,657
Grid — Original paper20,11321,06322,58822,36923,655
Grid — Time-matched20,11320,57222,10122,26123,290

Increasing the baselines' compilation-time budgets improves their performance, but MARSA retains its advantage over every baseline on both topologies. On Q20, MARSA reduces aggregate SWAP count by 66.49%, 68.80%, and 69.88% relative to LightSABRE, ML-SABRE, and AIRouting, respectively, even under matched compilation-time budgets. These results show that MARSA's advantage is not simply due to a larger compilation-time budget. We have added these experiments in Table 6 and 7 in the revised paper.

2. Joint learning of mapping and routing

We have evaluated the effectiveness of jointly learning mapping and routing in Section 5.2. To further strengthen this evidence, we conducted an additional ablation, detailed in our Response 1 to Reviewer PMFf (Table 1). A newly trained routing-only model conditioned on QMSI mappings achieves only 6.61% SWAP reduction, compared with 16.40% for joint MARSA on Q20.

More fundamentally, learning initial mapping and routing separately is difficult to formulate well because the final quality depends on each other. A separately trained mapping model cannot optimize the true SWAP objective without evaluating its decisions through a routing procedure, and vice versa. Therefore, previous methods typically use a fixed classical mapper or router for the learned model, making the model dependent on that mapper/router's behavior. If we directly combine such two modules together, they suffer from distribution shift. Conversely, if two separate modules are instead trained with a shared downstream reward, they are already being jointly optimized, even if implemented as different networks.

MARSA captures this coupling directly: a single policy generates the initial mapping and complete routing trace, and the final verified SWAP cost reinforces both portions of the same trajectory. This is the central motivation and contribution of our formulation, rather than merely learning a better standalone initial mapping.

AuthorsOpenReview ↗

Response to Reviewer LNzk (Part 2)

3. Adaptation to other hardware topologies

MARSA currently requires retraining for each new topology. We note that topology-specific training is standard practice in learned quantum compilation, including Pozzi et al. (2022) [1], QRoute (Sinha et al., 2022) [2], Ren et al. (2024) [3], and AlphaRouter (Tang et al., 2024) [4].

MARSA follows this established setting and demonstrates the same framework on Q20 and a substantially larger 60-qubit grid. Training is a one-time offline cost: our Q20 model required 27h 28min on eight RTX 5090 GPUs. Once trained, the policy can be reused across circuits on the target hardware. Since topology remains fixed throughout a processor's deployment, potentially for years, this modest cost can be amortized over many compilation workloads.

Furthermore, the MARSA formulation also naturally admits a topology-conditioned extension: prefix the input with a hardware specification, e.g., <topo> <edge> u v <edge> ..., and derive automaton legality masks from that graph. This would allow training a single policy across multiple topologies, potentially reducing the need for complete retraining on a new topology. We leave this extension to future work.

4. Relationship to previous work

These works solve related but distinct problems. SABRE/LightSABRE (Li et al., 2019; Zou et al., 2024) are classical heuristic compilers, using hand-designed distance and lookahead objectives; they do not train a neural network. Both Ren et al. (2024) and Tang et al. (2024) are neural policies trained only for routing, assuming initial mapping is provided.

To our knowledge, MARSA is the first neural policy to generate both the initial mapping and the complete routing trace. This requires a new joint representation, a route-state automaton enforcing evolving legality constraints, and trajectory-level self-improvement that directly optimizes both decisions. This joint learned optimization of mapping and routing, rather than neural routing alone, is the key distinction of MARSA.

[1] Pozzi, Matteo G., et al. "Using reinforcement learning to perform qubit routing in quantum compilers." ACM Transactions on Quantum Computing 3.2 (2022): 1-25.

[2] Sinha, Animesh, Utkarsh Azad, and Harjinder Singh. "Qubit routing using graph neural network aided Monte Carlo tree search." Proceedings of the AAAI conference on artificial intelligence. Vol. 36. No. 9. 2022

[3] Ren, Xiangyu, et al. "Leveraging machine learning for quantum compilation optimization." Proceedings of the 61st ACM/IEEE Design Automation Conference. 2024.

[4] Tang, Wei, et al. "Alpharouter: Quantum circuit routing with reinforcement learning and tree search." 2024 IEEE International Conference on Quantum Computing and Engineering (QCE). Vol. 1. IEEE, 2024.

Thank you for the additional experiments and clarifications. The compilation-time measurements and time-matched comparisons substantially address my main concern about practical efficiency. It is encouraging that MARSA retains an advantage when the baselines receive comparable time budgets.

Please make clear which circuits are used in Table 5, how their time budgets were determined, and what computational resources each baseline receives. Matching elapsed time is useful, but does not necessarily match computational cost.

The new routing-only baseline strengthens the evidence for learning placement. I also accept your distinction between separate networks and separate optimisation. The remaining question is what benefit comes specifically from a single policy, compared with separate placement and routing networks trained together; this can reasonably be acknowledged as an open comparison.

The retraining requirement and narrower novelty claim are now clearer. Overall, the response strengthens the confidence I have in my positive-leaning assessment.

We thank the reviewer for the immediate follow-up and for acknowledging that our additional experiments and clarifications substantially address the concerns about practical efficiency and joint learning. We clarify the experimental setup of Table 5 below.

  • Circuits: The same real algorithm circuits used in our main experiments in the paper ("Strong synthetic-to-real transfer" paragraph, Figure 4), evaluated on both Q20 and the grid.
  • Time budgets: We timed MARSA three times per circuit and used the median runtime as that circuit's wall-clock time budget for each baseline. Baseline search settings and restart caps were calibrated using timing measurements, with three independent runs per circuit. For every baseline, the sum of per-circuit runtimes was within 3% of MARSA's corresponding total.
  • Computational resources: Each MARSA job used one RTX 5090 GPU and eight pinned physical CPU cores. Each baseline (QMSI, LightSABRE, ML-SABRE, AIRouting) used eight pinned physical CPU cores without a GPU, as these implementations do not require GPU usage.
Reviewer_PMFfOpenReview ↗
Rating
4(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper frames qubit mapping and routing as a unified structured-generation task: an encoder–decoder Transformer outputs the initial logical-to-physical permutation, followed by the entire CNOT/SWAP trace, as a token sequence. A route-state automaton masks invalid tokens at each step, ensuring validity by construction (Prop. 1) alongside cost-preserving pruning. The policy is first initialized through imitation of the QMSI search expert (Ren et al., 2024) using a two-axis width×length curriculum, and then post-trained via sequence-level GRPO (with mean-centered advantages, a GSPO-style length-normalized ratio, and a squared reference regularizer) on its own verified rollouts, receiving rewards based on normalized SWAP reduction compared to the expert.

Critical strengths
  • Clear motivation and formulation. The argument that a mapping is only valuable through the routes it enables, and that a joint sequence lets a single route-level objective touch both, is accurate and well motivated. The claim of being the first learned policy to output mapping and the full route as a single trajectory is consistent with the related work the paper surveys.

  • The paper is clearly written and it flows really well.

  • The transfer experiments are conducted carefully. The synthetic training distribution is constructed solely from aggregate statistics of the benchmark circuits, the authors reject any generated circuit that exactly matches a benchmark or shares a 64-CNOT segment with one, and they deduplicate across splits. This is more rigorous than most learned-compiler papers I have reviewed, making the synthetic-to-real results believable.

Critical weaknesses
  • The mapping ablation fails to isolate what it claims to. In Section 5.2, the authors substitute MARSA’s generated mapping with QMSI’s mapping while keeping the learned router, concluding from the drop in improvement that MARSA’s mappings are superior. However, the router was trained solely on mapping prefixes it generated itself, meaning that feeding it a QMSI mapping places it in an unseen situation. Part of the drop might simply be the router performing worse on unfamiliar prefixes. A cleaner control would be a routing-only model trained on QMSI mappings, compared against the joint model.

  • The paper exclusively counts SWAPs. A SWAP equals three CNOTs, and SWAPs on disjoint edges can execute in parallel. However, the objective (Eq. 4), the reward (Eq. 20), and the output format (Eq. 7) treat each SWAP as a single serial token without any notion of time. Circuit depth, two-qubit gate count after decomposition, and overall execution time are never reported. The authors even disable gate cancellation in AIRouting to keep the SWAP count exact, meaning the reported numbers are further from the actual physical gate count than necessary. This creates a fairness issue when evaluating the baselines. LightSABRE is run using the decay heuristic, which is SABRE’s depth-aware mode that knowingly inserts a few extra SWAPs to align them into parallel layers. ML-SABRE also optimizes for depth. Therefore, the comparison scores depth-optimizing methods on a metric they are not actively optimizing, which plausibly explains part of the 71% reduction over LightSABRE on Q20. Two steps would resolve this: report depth (ideally, two-qubit depth post-decomposition) for all methods, and re-run LightSABRE using its basic or lookahead heuristic, which directly targets SWAP count.

Questions for authors
  • What are the routed depths, or two-qubit depths after SWAP decomposition, for MARSA and each baseline on both topologies? Is the number of swaps the best metric to compare the compilers?

  • Can you add the routing-only control for the mapping ablation, that is, a router trained from QMSI mappings? How much of the drop survives when the router has seen QMSI-style prefixes during training?

4 comments

We thank the reviewer for recognizing the motivation, formulation, experimental rigor, and clarity of our work. We agree that the mapping ablation and evaluation metrics need further evidence, and have conducted additional experiments to address these concerns.

1. A newly trained routing-only control

The reviewer insightfully notes that replacing MARSA's generated mapping with a QMSI mapping might introduce a distribution shift of unfamiliar prefixes. We therefore trained a new routing-only model from scratch on QMSI mappings. It generates only routing operations and EOS, with both supervised learning and verifier-guided self-improvement applied to the routing continuation.

Table 1: Q20 mapping ablation with routing-only models trained on QMSI mappings.

PolicyMappingKCompletedSWAP reduction vs QMSI
Joint MARSA, supervisedLearned161,822/1,8227.52%
Joint MARSA, + self-improvementLearned161,822/1,82216.40%
Routing-only, supervisedFixed QMSI161,822/1,8222.32%
Routing-only, + self-improvementFixed QMSI161,822/1,8226.61%

As shown in the table, the routing-only model achieves 6.61% swap reduction v.s. QMSI, compared with 16.40% by joint MARSA. It uses 78,320 SWAPs versus MARSA's 70,109, i.e. 11.71% more SWAPs, while both models complete all 1,822 circuits. Thus, even when explicitly trained on QMSI mappings, the routing-only model falls short of joint MARSA substantially. This controlled comparison addresses the distribution shift concern and provides stronger evidence for learning the initial mapping and complete routing trace jointly. We have added this experiment in the main text.

2. Two-qubit depth after SWAP decomposition

We additionally evaluated MARSA and all four baselines on the real circuits for both topologies. We decomposed each SWAP into three CNOTs and measured the resulting two-qubit circuit depth.

Table 2: Aggregate SWAP counts and decomposed two-qubit depths.

MethodQ20 SWAPsQ20 depthgrid SWAPsgrid depth
MARSA3,21042,48120,11389,079
QMSI3,65342,67921,06390,769
LightSABRE11,33464,60422,58891,939
ML-SABRE11,27263,98422,36991,518
AIRouting11,14264,40223,65594,120

MARSA achieves the lowest aggregate SWAP count and two-qubit depth on both topologies. On Q20, it reduces two-qubit depth by 34.24%, 33.61%, and 34.04% relative to LightSABRE, ML-SABRE, and AIRouting, respectively, while also outperforming QMSI. These results demonstrate that MARSA delivers substantial depth improvements despite optimizing SWAP count rather than depth directly. This table was also added as Table 5 in the revised version of the paper.

We choose SWAP count as our primary metric following the convention of previous neural methods on qubit routing. Still, the MARSA framework can also support depth-aware optimization by evaluating circuit depth with the verifier and incorporating it into the self-improvement reward.

3. LightSABRE: decay versus lookahead

We additionally evaluated LightSABRE with the lookahead heuristic, alongside the original decay results.

Table 3: LightSABRE heuristic comparison.

TopologyHeuristicSWAPsTwo-qubit depth
Q20Decay (original)11,33464,604
Q20Lookahead12,57167,429
GridDecay (original)22,58891,939
GridLookahead22,97292,473

The lookahead heuristic does not outperform the original decay configuration on either topology. MARSA outperforms both configurations with similar reduction rate, confirming that its advantage is not attributed to the choice of LightSABRE heuristic.

Dear Reviewer PMFf,

Thank you again for your valuable feedback and suggestions. We have provided additional experiments and clarifications in our rebuttal, and we hope these additions help address your concerns.

We would be grateful for any further feedback and would be happy to clarify any remaining questions or concerns.

Thank you very much for your time and consideration!

Thank you for the detailed response. The routing-only control and the depth measurements address both of my major concerns, and I am raising my score. I ask for the following updates in the camera-ready:

  1. The depth results change the overall picture and should not be left solely in a table. Against the SABRE baselines the depth reduction is roughly 34% rather than the 71% SWAP reduction quoted in the abstract and introduction, and against the QMSI teacher the depth improvement is under 1% on Q20 and under 2% on the grid. Please report both metrics wherever the headline numbers appear, and state clearly that the beyond-teacher gain is currently a SWAP-count gain that does not yet translate into shallower circuits.
  2. The claim of being the first to jointly handle mapping and routing should be qualified. SABRE’s bidirectional layout already lets the router pick the initial mapping by routing forward, backward and forward again, and the same idea has since been proposed for learned routers (see also the reference raised in reviewer fdZo’s thread). This constitutes a joint optimization/consideration of the problems since the one is optimized over the other. The paper mentions this only briefly. The accurate claim is that MARSA is the first to generate both in a single learned trajectory under a single objective. Relatedly and importantly, the new routing-only control uses a fixed QMSI mapping; the more natural comparison is the routing-only model with a mapping obtained from its own forward–reverse–forward passes, since that is the simplest alternative to learning a mapping prefix. Including that row would make the ablation complete and correct.

Thank you for your positive assessment and constructive feedback. We appreciate your suggestions and will incorporate the requested updates in the camera-ready version.

Comments on the paper

1 comment
Area_Chair_VBwaOpenReview ↗

Please respond to the authors

Dear reviewers,

Thank you for your service.

The authors have provided their response to your initial reviews. As we approach the end of the author-reviewer discussion period, could you please respond to the authors’ rebuttal (if you haven’t done so already) and communicate the extent to which your concerns have been addressed, and whether you have any further questions and concerns?

Thanks, Your AC

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.