acceptodds
Under review as a conference paper at ICLR 2027

Learning Gaussian Boson Sampling for Integrated Quantum Photonics

12 views

Abstract

Gaussian boson sampling (GBS) is a leading platform for probing the boundary between quantum and classical computation. Given its impact on practical quantum advantage, emulating GBS is crucial, but exact calculation of its output probabilities is computationally expensive and #P-hard in general. Practical GBS workflows repeatedly query related photonic circuits and detection patterns, creating an opportunity to accelerate this computation through machine learning. We formulate GBS probability evaluation as a supervised learning task and introduce PhOracle, a physics-informed recurrent model that adapts its matrix computations to the requested detection pattern. We also introduce GBSBench, a circuit-disjoint benchmark that evaluates generalization to unseen circuits and unseen detection patterns. PhOracle achieves a median absolute percentage error of 4.56% on held-out circuit-pattern queries, while matched runtime tests show that PhOracle evaluates probabilities up to 79.8 faster than the reference solver. We further apply PhOracle to closed-loop quantum circuit optimization. Without task-specific retraining, PhOracle accelerates single-pair and multipair Bell-state circuit discovery by 17.3 and 36.2, respectively, reaching fidelities close to those achieved by reference-solver search. Together, this work provides an empirical basis for future studies of learned probability models across broader quantum-photonic systems and applications.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Gaussian boson sampling, scientific machine learning, surrogate modeling, physics-informed learning, quantum photonics, datasets and benchmarks

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_uFyVOpenReview ↗
Rating
4
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates learning-based acceleration of Gaussian boson sampling (GBS) probability evaluation, a computationally expensive task in quantum photonics. The authors propose PhOracle, a physics-informed recurrent neural architecture that uses query-conditioned matrix compositions to predict detection-pattern log-probabilities. They also introduce GBSBench, a benchmark designed to evaluate generalization across unseen photonic circuits and detection patterns.

Experiments demonstrate that PhOracle substantially outperforms existing neural baselines, achieving a median absolute percentage error of 4.56% and up to 79.8× faster probability evaluation than the reference solver. Ablation studies validate the effectiveness of the recurrent architecture and query-conditioning mechanism. Furthermore, the authors apply PhOracle to Bell-state circuit discovery without task-specific retraining, achieving 17.3× and 36.2× speedups for single-pair and multipair tasks, respectively, while maintaining competitive final fidelities after reference-solver refinement.

Critical strengths

1.Novel and well-motivated approach. The paper introduces a compelling learning-based formulation of Gaussian boson sampling probability evaluation. PhOracle incorporates physical structure through query-conditioned recurrent matrix compositions, providing a meaningful architectural contribution beyond generic neural surrogates.

2.Strong empirical results and convincing ablations. PhOracle substantially outperforms the evaluated baselines, achieving a median absolute percentage error of 4.56% and up to 79.8× faster probability evaluation than the reference solver. The ablation studies further demonstrate the importance of recurrent computation and query conditioning.

Critical weaknesses
  1. Although the paper reports both median and mean prediction errors, the relatively large gap between these metrics suggests that a more detailed analysis of the error distribution would be valuable. Reporting the 90th and 95th percentile errors on the existing test set would help clarify the model's robustness, particularly for challenging detection patterns.
Questions for authors

Could the authors report the 90th and 95th percentile absolute percentage errors on the existing test set? This would help clarify whether the gap between the median and mean APE is primarily driven by a small number of challenging predictions and provide further evidence of the model's reliability.

2 comments

We thank the reviewer for their strong support and for recognizing the novelty of our learning formulation, the query-conditioned architecture, and the strong experimental results and ablations.

Q1. What are the 90th- and 95th-percentile APEs, and which queries drive the error tail?

PhOracle's 90th- and 95th-percentile APEs are 27.00% and 51.39%, using the same checkpoint and 95,995 valid test queries as Table 2.

The error tail is concentrated in high-photon-number queries, particularly those spread across more output rails. Ten- and twelve-photon events comprise 40.0% of test queries but 76.5% of the highest-APE 5%.

Following the reviewer's suggestion, we added P90/P95 columns to Table 2, highlighted in orange, to make the tail visible alongside mean and median APE. As our work serves as a foundational study, identifying these challenging patterns will provide useful guidance for subsequent research. We thank the reviewer for giving us the opportunity to strengthen the paper through this additional analysis.

Dear Reviewer uFyV,

Thank you again for recommending our paper for acceptance and for your constructive feedback. As the discussion period draws to a close, we would appreciate it if you could review our rebuttal and let us know if you have any further questions or comments.

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

The paper studies supervised prediction of individual coarse-grained Gaussian boson sampling probabilities. PhOracle combines query-conditioned recurrent matrix products with invariant rail features and a Rail Transformer. GBSBench uses circuit-disjoint splits within a fixed nine-rail family with identical photon-pair sources, a five-port interferometer, four retained spectral modes per rail, and two to twelve total detected photons. Reported test Log-MSE is 0.01734, with median and mean APE of 4.56% and 21.30%. Forward evaluation is up to 79.8 times faster than the specified reference solver. A frozen model also guides two Bell-state circuit searches, reaching reference-verified fidelities of 0.9919 and 0.9943 after refinement. Reported accumulated online search-and-refinement speedups are 17.3 and 36.2 times.

Critical strengths
  1. The physical inductive bias is concrete and supported by substantial comparisons. Query conditioning controls successive matrix compositions, while invariant contractions preserve the stated internal-coordinate symmetries. Table 2 reduces Log-MSE from 0.1323 for the similarly sized block Transformer to 0.01734. Table 6 supports recurrence and conditioning inside the updates, with approximately 5.20 and 7.25 times larger errors for the deeper single-step and query-independent controls.

  2. The evaluation has meaningful safeguards. Circuits are separated across splits, and bootstrap intervals resample circuits rather than correlated query rows. The downstream surrogate score is explicitly distinguished from physical fidelity. Selected circuits are rescored and refined using reference probabilities, with positivity, measurement-consistency and analyzer-invariance checks. These support the reported examples without relying on the surrogate to verify itself.

Critical weaknesses
  1. Numerical validation does not yet cover the full label range in the units used for the headline metrics. Appendix A.6 reports excellent agreement with The Walrus over 76,800 queries, but the maximum relative discrepancy is reported only for probabilities above 1e-12. Log-MSE and APE cover all valid positive labels, including low-energy/high-photon-number queries; Appendix C.4 also excludes 26 nonpositive labels without their stratified distribution. A global maximum absolute discrepancy cannot certify arbitrary-small probabilities in relative or log units. This is a validation/reporting gap, not evidence that those labels or the model ranking are wrong. A probability-stratified summary of the existing cross-checks, followed by high-precision checks only where needed, would directly resolve the concern.

  2. The uncertainty and efficiency conclusions should stay conditional on the actual evaluation protocol. The bootstrap holds trained checkpoints fixed; independent training repetitions and tuning budgets are not specified, while PhOracle and baselines use different learning-rate and spacing-loss recipes (Appendix D.2). The circuit-discovery result is one 25-energy by three-start sweep per task, with one selected candidate refined. These 75 candidates are not 75 independent repetitions of the complete selection pipeline. The reported endpoint fidelities and accumulated online time ratios are useful demonstrations, but do not establish general time-to-quality or selection-success guarantees. Existing seed records and paired endpoint or trajectory summaries would substantially strengthen the interpretation.

Questions for authors
  1. Can the authors report label counts and cross-implementation relative or absolute log10 discrepancies by probability decade and photon number, including probabilities at or below 1e-12 and the 26 excluded pairs? If tail discrepancies occur, do limited higher-precision checks change the main metrics or method ranking?

  2. How many independent training runs and what validation-tuning budgets underlie the reported models? For circuit discovery, can existing paired results show endpoint quality and cost by energy/start, or success at a common fidelity threshold? Please distinguish checkpoint-conditional circuit intervals from training and full-pipeline variability.

  3. Can Table 7 include the selected source energies and mean herald-plus-ancilla probabilities? This would clarify the physical use of the conditional-fidelity solutions without changing the stated optimization task. Please also keep learning errors explicitly relative to the four-mode reference and give absolute error values at representative large delays.

2 comments

We thank the reviewer for supporting acceptance and for questions that helped us clarify the reliability and reproducibility of our evaluation.

Q1 & W1. Are rare-event labels reliable?

After high-precision correction, median APE remains 4.465%, Log-MSE changes 0.01393 → 0.01399, and MAPE 16.20% → 16.39% on the original validation mask. PhOracle retains the lowest validation Log-MSE.

Walrus cross-check before correction. Log and relative differences use positive outputs from both solvers.

Stored probability rangeQueriesP95 absolute log₁₀ differenceMaximum absolute log₁₀ differenceMaximum relative difference (%)
P > 10⁻¹²69,9987.84 × 10⁻¹³1.12 × 10⁻⁶0.000258
10⁻²⁰ < P ≤ 10⁻¹²4,9073.25 × 10⁻¹⁰1.41596.15
10⁻³⁰ < P ≤ 10⁻²⁰1,7506.08 × 10⁻¹¹0.0028320.650
0 < P ≤ 10⁻³⁰1424.82 × 10⁻⁷0.038698.523
Stored numerical zeros3———

We clarified the probability-stratified cross-check and the recheck of 16 discrepant queries using exact coefficient extraction with high-precision vacuum normalization in Appendix A.6 (Table 3). We also detailed the 26 excluded labels in Appendix C.4. These additions are highlighted in teal.


Q2(a) & W2. How many training runs and tuning trials were used?

All models use the same eight-run tuning budget, 75,000 updates per run, and validation-based selection. Each selected configuration has three independent runs. We clarified these conditions and separated training variability from circuit-bootstrap uncertainty in Table 2 and Appendices D.2–D.3, highlighted in teal.


Q2(b) & W2. How do circuit-discovery results vary across starts?

Each task used one complete discovery workflow with 25 energies × three paired starts. Reference-evaluated endpoint fidelities before refinement are:

TaskPhOracle fidelity, median [IQR]Reference-search fidelity, median [IQR]Endpoints with F ≥ 0.95, PhOracle vs. reference
Single-pair0.9111 [0.6551, 0.9450]0.9919 [0.9865, 0.9937]15/75 vs. 75/75
Multipair0.9343 [0.6298, 0.9508]0.9928 [0.9879, 0.9944]22/75 vs. 75/75

The 17.3×/36.2× accumulated-time ratios include all searches and one selected PhOracle refinement. We clarified the timing in Section 7 and added endpoint distributions in Appendix G.1, highlighted in teal.


Q3(a). Selected energies and event probabilities

Event probabilities are the mean herald-plus-ancilla probabilities over 32 analyzers.

TaskPhOracle energy/source (pJ)Event probability, before → after refinementReference-search energy/source (pJ)Reference-search event probability
Single-pair81.687 × 10⁻⁴ → 1.851 × 10⁻⁴12.497 × 10⁻⁷
Multipair0.31.938 × 10⁻¹¹ → 2.149 × 10⁻¹¹17.339 × 10⁻⁹

We added source energies to Section 7 and the full results to Appendix F.4 (Table 9), highlighted in teal.


Q3(b). Reference target and large-delay errors

All reported learning errors are computed against the four-mode reference.

We set the delay range based on the temporal-overlap scale at 10 pJ, which is approximately 309 fs. The maximum pairwise delay is 900 fs, approximately three times this scale.

We added median APE and median absolute probability error across the full delay range to Appendix E.3 (Table 7), highlighted in teal.

Dear Reviewer SewE,

Thank you again for supporting acceptance and for your careful questions, which helped us strengthen the rigor and clarity of the paper. We have supplemented the numerical validation and completed matched hyperparameter searches for the baselines, as detailed in our rebuttal and revised manuscript. As the discussion period draws to a close, we would appreciate your feedback on whether these additions address your concerns, and whether any questions remain.

Reviewer_KzU2OpenReview ↗
Rating
2
Confidence
1
Informativeness
3(+1)
Presentation
2(−1)
Soundness
2
Read review
Paper summary

The paper introduces GBSBench and PhOracle, a recurrent model for predicting coarse-grained GBS probabilities in a fixed nine-rail circuit family. PhOracle achieves 4.56% median APE and supports Bell-state circuit optimization without retraining. With reference-solver refinement, the reported post-training search speedups are 17.3× and 36.2×.

Critical strengths
  1. The matrix-composition design has a clear physical motivation. The ablations support the roles of recurrence and query conditioning, including comparison with a deeper single-step Transformer.

  2. The downstream results go beyond prediction accuracy. The frozen model guides circuit search, and the final fidelities are verified using reference probabilities and physical-consistency checks.

Critical weaknesses
  1. The practical speedup needs stronger physical baselines. Tab. 4 shows nearly single-mode sources, but reduced-mode approximations are not evaluated. The identical-source, frequency-independent setup at zero delay may also permit an exact factorization across a common Schmidt basis. It is unclear whether the reference implementation exploits this structure. These comparisons could change the accuracy–runtime advantage.

  2. Generalization remains narrow. Pattern-OOD holds out allocations within known collision signatures, while the source design and circuit size stay fixed. Under temporal mismatch, pooled median APE reaches about 65% at the largest delay. The current results leave uncertainty about performance beyond the controlled training family.

Questions for authors
  1. How do one-/two-mode approximations and, if applicable, a Schmidt-factorized exact solver compare with PhOracle in accuracy and runtime, including setup costs?

  2. How does performance change when entire collision signatures or source configurations are held out?

  3. Please report the selected source energies, heralding probabilities, and reference-evaluated fidelity distributions across the 75 searches.

  4. How much reuse is needed to offset dataset-generation and training costs?

3 comments
AuthorsOpenReview ↗

Response to Reviewer 1 (1/2)

We appreciate the reviewer's thoughtful questions and suggestions. We address each detail below.

Q1 & W1. How does PhOracle compare with physical approximations?

We carried out calculations under the conditions suggested by the reviewer to make our speedup claim more concrete. Please note that we have interpreted “mode” in the reviewer’s question as referring to the “spectral mode” used in Figure 6(b). This is not the same concept as the “Schmidt mode” used in the context of Schmidt factorization. While we appreciate the reviewer’s insightful suggestion of Schmidt factorization as a promising approach in this regime, assessing how broadly it applies to our spectrally resolved formulation would require a separate analysis beyond the scope of the present study.

We compared PhOracle with K=1 and K=2 truncations under the same conditions on GBSBench queries.

MethodMedian APE (%)MAPE (%)Setup (s)Total time, including setup (s)
K=1 truncation0.2711.1030.732192.7
K=2 truncation0.1130.4130.757245.8
PhOracle4.46516.200.50623.9

These results show that physical methods achieve high accuracy within the present benchmark, while PhOracle offers substantially lower evaluation latency than the physical baselines.

We would also like to explain the fairness of the comparison to support our speedup claim. Gaussian boson sampling itself has diverse versions depending on certain assumptions (Oh, 2025). The central takeaway of our manuscript is the comparison between a given exact physics model and its ML model counterpart. For each of these versions, there are various approximation methods that take advantage of its specific physical conditions. We therefore view these approximations less as baselines for our study and more as tools for accelerating costly label generation under those conditions.

Among many options, we have chosen a “frequency-resolved” approach, which is a low-level physics model compared with others (Kim et al., 2024). Such an approach can model physical effects such as the time delay studied in the manuscript and many others. Meanwhile, this level of detail comes at the cost of calculation speed compared with other frequency-agnostic versions. One can always speed up GBS calculations by removing frequency resolution. However, the benefits of the low-level physics model will be lost.

Further study of these approximation methods, including the Schmidt-factorization approach suggested by the reviewer, could help reduce the cost of generating training labels and offers a promising direction for future research.

AuthorsOpenReview ↗

Response to Reviewer 1 (2/2)

Q2(a). Can PhOracle generalize to unseen collision signatures?

We ran a new experiment to test generalization to unseen collision signatures.

Training signaturesLog-MSEMedian APE (%)Mean APE (%)
All 1710.023645.0326.99
137 retained0.0388510.1852.97

Even after withholding one fifth of the signature types entirely, median APE remains about 10%, showing generalization to collision structures absent from training. Following the reviewer's suggestion, we added the result to Section 6.2 and the detailed comparison to Appendix E.5 (Figure 11), with the new text and captions highlighted in blue.


Q2(b). Can PhOracle generalize to unseen source conditions?

We ran an additional source-energy OOD experiment.

Source conditionLog-MSEMedian APE (%)Mean APE (%)
Training-grid energies0.01484.4317.40
Source OOD0.01584.4217.48

PhOracle maintains comparable overall accuracy without retraining. We added this evaluation to Table 1, Section 5.2, Figure 4(f), and Appendix E.4, highlighted in blue. We thank the reviewer for suggesting these experiments, which helped clarify our model’s generalization capabilities.


W2. Why is error high at large delays?

The 67.1% median APE corresponds to a severe stress test, with a maximum pairwise delay approximately three times the source’s temporal-overlap scale. Even under this substantial shift from zero-delay training conditions, updating the adjacency reduces median APE from 75.1% to 67.1% without retraining. This response to interference changes encoded in the adjacency supports our learning formulation’s ability to generalize to unseen physical conditions.

We clarified the temporal-mismatch evaluation and results in Sections 5.2 and 6.2 and Appendix E.3, highlighted in teal.


Q3. What are the selected source energies, heralding probabilities, and fidelity distributions?

Selected PhOracle circuits after refinement:

TaskEnergy per sourceFidelityMean herald-plus-ancilla probability
Single-pair8 pJ0.99191.851 × 10⁻⁴
Multipair0.3 pJ0.99432.149 × 10⁻¹¹

We also added operating points for all comparison endpoints to Appendix F.4, Table 9, and fidelity distributions across 75 starts to Appendix G.1, highlighted in teal.


Q4. When does PhOracle break even?

Estimated break-even is about ten single-pair or four multipair campaigns, each with 75 starts. Preparation and one training run cost approximately 125 GPU-hours, while each campaign saves 13.3/33.2 hours, respectively, including refinement.

We would also like to emphasize that our contribution here is to enable low-latency probability evaluation for iterative design, rather than to minimize the total computational cost of a single optimization task. Faster evaluations allow designers to explore candidates and revise designs with less waiting, giving the surrogate practical value beyond total computational speedup.

Dear Reviewer KzU2,

As the discussion period draws to a close, we would appreciate your feedback on our responses and the additional physical-approximation comparisons and OOD experiments conducted in response to your suggestions. If these additions address your main concerns, we would be grateful if you could consider raising your score. Please let us know if any questions or concerns remain so that we can address them before the discussion closes.

Thank you again for your time and thoughtful feedback.

Comments on the paper

3 comments
AuthorsOpenReview ↗

Code availability

As stated in the manuscript, we are releasing our code during the review period to support reproducibility. We hope this release clarifies implementation details and helps reviewers better understand our method.

Anonymous code repository

AuthorsOpenReview ↗

Summary of reviewer feedback and additional results

Dear Area Chair and Reviewers,

As the discussion period draws to a close, we summarize the reviewers’ assessments and the improvements made in response.

Reviewer uFyV rated the paper 4 (Clear acceptance), highlighting the novel learning formulation, the architectural contribution beyond generic neural surrogates, and the strong empirical results and convincing ablations.

Reviewer SewE rated the paper 3 (Weak acceptance), recognizing the concrete physical motivation, substantial comparative evidence, and careful evaluation safeguards.

Reviewer KzU2 rated the paper 2 (Weak Rejection) but recognized our physically motivated architecture, supporting ablations, and reference-verified circuit design. In response to their concerns regarding physical comparisons and generalization, we conducted additional experiments to address both points.

We have addressed the reviewers’ questions through additional experiments, numerical checks, and manuscript revisions. The main results are summarized below.

Reviewer concernHow we addressed itKey result
Comparison with physical methodsCompared K=1/K=2 truncation under matched conditions.Physical methods are more accurate, while PhOracle has 8–10× shorter evaluation time, including setup.
Generalization to new source conditionsEvaluated the frozen model at source energies absent from training.Median APE is 4.42%, comparable to 4.43% for the paired in-distribution control.
Generalization to new collision structuresRetrained with entire collision signatures excluded from training and validation.Median APE is 10.18% on the held-out signatures.
Fairness of baseline tuningCompared all learned models under matched eight-run tuning budgets.PhOracle retains the lowest benchmark errors.
Reliability of small-probability labelsAdded probability-stratified checks and high-precision recalculation.Correcting the labels leaves validation median APE unchanged and Log-MSE essentially unchanged.

We also added error-tail analysis and circuit-discovery variability statistics in response to the reviewers’ reporting requests.

These additions strengthen the evidence for our central contribution: a systematic framework for investigating how accurately and efficiently GBS probability evaluation can be learned and generalized, with demonstrated utility in quantum-photonic design. We thank the reviewers for helping make the evaluation more complete and the revised paper clearer.

AuthorsOpenReview ↗

Follow-up Response and Final Remarks

Thank you for your follow-up. We appreciate your willingness to consider the other reviewers’ assessments and to support acceptance if favored by them and the AC. Although you expressed limited confidence in assessing GBS applications, your questions identified central issues and helped us strengthen the paper through additional experiments and clearer explanations.

The nine-rail setting was a deliberate starting point for this first systematic study. Our aim is to provide the community with a benchmark and a model baseline for learning GBS probabilities. The results already demonstrate useful acceleration in circuit design, and we plan to extend the study to larger systems.

The other reviewers recommended acceptance (ratings of 3 and 4), recognizing the novelty of the approach and the strength of the empirical evidence. We hope publication will bring this research direction and its benchmark to a wider community, allowing more researchers to build on this foundation.

With this community value and the additional evidence in mind, we would be grateful if you could also consider raising your score as discussions with the AC continue. Thank you again for your thoughtful engagement and continued consideration.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.