acceptodds
Under review as a conference paper at ICLR 2027

LiftDiff: Guided Discrete Diffusion for Quantum Code Construction

Abstract

The realization of fault-tolerant quantum computation relies on Quantum Error-Correcting Codes (QECC). Of particular interest are codes that combine a high encoding rate with sufficient distance. Existing algebraic constructions, such as lifted product (LP) codes, build a large quantum low-density parity-check (qLDPC) code from a classical base matrix. In this work we introduce LiftDiff, a generative approach to the discovery of QECC. A discrete denoising diffusion model is trained to generate base matrices over the group algebra , which are then expanded into LP codes. The reverse diffusion process is guided by an energy function that evaluates quantum code properties, including girth, degree, sparsity and decoding performance, steering the sampler toward matrices that yield LP codes with favorable parameters. Our method discovers codes with parameters such as , , and , including a code that achieves the highest among the baselines. Moreover, several discovered codes outperform the Gross code across a broad range of the tested physical error rates under both code-capacity and circuit-level noise.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Discrete Diffusion, Quantum Error Correction

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_by7ROpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

The paper introduces LiftDiff, a guided discrete diffusion framework for constructing quantum LDPC codes by generating classical base matrices and lifting them into quantum codes. The method uses energy-based guidance to optimise desirable code properties, discovering several competitive codes with favourable rate–distance trade-offs and logical error performance.

Critical strengths

Interesting framework: The combination of discrete diffusion, energy-based guidance, and lifted-product constructions provides a promising approach to automated quantum code discovery.

New code constructions: The method discovers several competitive codes and demonstrates encouraging decoding performance under code-capacity and circuit-level noise.

Critical weaknesses

Limited discussion of related work: The literature review could better contextualise the contribution relative to earlier frameworks for constructing quantum codes from classical codes, particularly Roffe et al. (1804.07653) and the broader graphical framework of Chancellor et al. (1804.07653). While these works use different constructions, they also explore systematic quantum code design from classical structures, and a clearer discussion would help distinguish the proposed framework's novelty.

Limited methodological comparisons: Although the paper compares the resulting codes against existing constructions and studies guided versus unguided diffusion, it lacks a systematic comparison with alternative search or optimization methods under comparable computational budgets, making it difficult to isolate the benefits of diffusion itself.

Questions for authors
  • Could the authors address the two points raised in the weaknesses (i.e. discussion of related work and role of diffusion)?
  • Are there any interesting applications to early small scale noisy quantum architectures, where the methods provides a principles framework to work around the constraints of such architectures?
Reviewer_LDiPOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents LiftDiff, a guided discrete diffusion framework for quantum error-correcting code construction. The method generates sparse matrices within an existing lifted-product construction and guides sampling using objectives related to decoding performance, graph structure, and implementation cost. The experiments report 32 code instances, some of which outperform the selected baselines in parameter-based metrics or under specific noise models and decoder settings. Representative codes are also evaluated under circuit-level noise.

Critical strengths
  1. The paper applies guided discrete diffusion to quantum code search, incorporating performance objectives and structural constraints within a valid code construction. This provides a promising approach to learning-based quantum LDPC code design.
  2. The method finds several competitive code instances, evaluated through distance calculations and simulations under both code-capacity and circuit-level noise. Some codes show improvements over the selected baselines.
Critical weaknesses
  1. The comparison with the Gross code does not fully account for implementation costs. The [[247,19,14]][[247,19,14]] code achieves a higher kd^2/n, but increases the check weight from six to eight and has a slightly lower encoding rate. For finite-length quantum LDPC codes, check weight affects syndrome extraction and error propagation, so this metric alone does not establish a practical advantage. Appendix B provides circuit-level results, but the associated circuit depth and ancilla overhead are not sufficiently characterized.
  2. The contribution of diffusion to search efficiency remains unclear. The method uses an existing lifted-product construction, but lacks comparisons with random or local search within the same construction space and under matched distance-evaluation or runtime budgets. The 32 selected codes demonstrate that the method can find good codes, but do not establish whether it does so more efficiently than direct search.
Questions for authors
  1. Do the new codes retain an advantage over the Gross code after accounting for syndrome-extraction circuit depth and ancilla overhead?
  2. Within the same construction space and search budget, how does LiftDiff compare with random or local search in its success rate at finding codes of equal or better quality?
Reviewer_1KbwOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

LiftDiff uses guided discrete diffusion to search for quantum error-correcting codes. The model generates compact base matrices over a cyclic group algebra, which are expanded into CSS codes through a lifted-product construction. Sampling is guided by structural objectives and a differentiable belief-propagation decoder. The paper reports several competitive codes, with exact MILP distance certification for selected candidates.

Critical strengths
  • Searching over base matrices is a sensible choice. It keeps the representation compact and uses the lifted-product construction to ensure that the resulting checks commute.
  • The paper reports concrete constructions with a useful range of rates, distances, and check weights. The reported exact distance certification is particularly valuable. For example, the [[247,19,14]] code has kd²/n of about 15.08, which improves on the selected baselines under the stated check-weight restriction.
  • The decoding results go beyond comparing code parameters. In the code-capacity experiments, [[260,24,10]] achieves lower observed block error rates than the Gross code over much of the tested range while encoding twice as many logical qubits. Including circuit-level experiments is also a welcome addition, and Appendix B acknowledges where the results are mixed.
Critical weaknesses

1. The benefit of the diffusion model is still unclear.

The most important missing baseline is direct sampling from the same PEG-plus-random-polynomial distribution used for training, followed by the same screening procedure. I would also like to see what happens when the authors optimize the guidance energies directly from random or PEG initializations. These comparisons would help establish what the learned denoising process contributes.

This seems especially relevant for the headline results. Table 2 uses m = 2 and target support column degree dv = 2 for these configurations. If that degree is enforced, both entries in every support column must be one, so the support is fixed and the training variation comes from the polynomial assignments.

Figure 3 is a useful starting point, but it leaves this question open. The trajectory divergence in Figure 3b shows that guidance changes the samples. Figure 3c suggests a shift toward more desirable parameters, although its distances are upper bounds, so that plot alone cannot establish an increase in true distance. The number of trials and the cost of individually tuning the reported configurations are also unclear. Budget-matched comparisons across several seeds would make this part of the paper much more convincing.

2. The decoding guidance needs a clearer description and an ablation.

Equation (17) measures BCE against the sampled physical error. In quantum decoding, a correction can succeed even when it differs from that error by a stabilizer. This makes the BCE a surrogate whose usefulness needs to be checked against actual logical error rate.

I also could not reconstruct the continuous BP updates from Appendix A. The text says that edges are weighted, but does not give the equations for real-valued check matrices or explain how nonbinary syndromes enter the updates. Since this decoder supplies gradients to the generator, these details matter. An ablation separating decoding guidance from the structural terms and the final refinement step would help show which parts are responsible for the improvements.

3. The reported diffusion schedule needs clarification.

Equation (3) retains some dependence on the clean matrix whenever all beta values are below 1/2. The reported 100-step schedule, with beta at most 0.02, therefore does not reach an exactly uniform terminal distribution. This is difficult to reconcile with the description of a cosine schedule on alpha_bar ending at 1/2. This may be a reporting issue, but the actual schedule, terminal corruption level, and sampling initialization should be stated clearly.

4. The claim about outperforming Gross under both noise models is too broad.

The code-capacity gains are encouraging under the stated X-error model and BP-OSD decoder. The circuit-level evidence is more limited. Appendix B states that [[260,24,10]] remains within the Gross code's confidence intervals at every tested point. The [[247,19,14]] code improves over part of the circuit-level range, but mostly tracks Gross in the code-capacity experiment. I would qualify the abstract to reflect these conditional gains.

I also have a question about the zero-failure confidence bound in Figure 4, detailed below. It affects how much can be concluded at the lowest physical error rate.

Questions for authors
  1. How does LiftDiff compare with direct sampling from the training distribution and direct optimization of the same energies under a common budget? How many configurations and candidates were tried in total, and how consistently are the headline results recovered across seeds?

  2. Could you provide the exact continuous BP update equations and an ablation of the decoding guidance? How well does the physical-error BCE correlate with logical error rate across candidate codes?

  3. What diffusion schedule and terminal initialization were used in the experiments? What is the resulting terminal value of alpha_bar?

  4. Could you provide the shot counts, failure counts, and confidence-interval and normalization formulas for Figure 4? The zero-failure marker for [[247,19,14]] appears near 7 × 10^-9. Appendix B reports at most 10^5 shots per basis. Even pooling both bases as 2 × 10^5 shots, the usual one-sided 95% zero-event bound, divided by kT = 19 × 14, is about 5.6 × 10^-8. A different normalization or shot count may explain the discrepancy; I could not resolve it from the manuscript.

1 comment

Thank you for conducting the additional experiments. The schedule and the zero-failure point presented in Figure 4 are now clear, and the decoding ablation provides valuable insight.

My primary concern is that the new baselines do not include the settings that produced [[247,19,14]][[247,19,14]] and [[260,24,10]][[260,24,10]]. For these two settings, how frequently does each method achieve the same distance and check-weight targets across different seeds, given a comparable search budget? I also recommend softening the abstract’s claim regarding outperforming Gross, considering the mixed results at the circuit level. I will maintain my score at 2 for now.

Comments on the paper

1 comment

Dear Reviewers and AC, We thank the reviewers for their thoughtful feedback and constructive suggestions, which helped us improve the paper. We have addressed the raised concerns and updated the paper, with all new text highlighted in blue. We are glad that the rebuttal allowed us to clarify previously unclear aspects of our work, introduce additional baselines and ablations, update the abstract, and discuss interesting potential applications. Unfortunately, due to time constraints, we were unable to complete the additional experiments requested in the final round of discussion.

We appreciate the reviewers' feedback and believe that the resulting revisions have strengthened the paper and helped clarify the motivation and contributions of its different components.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.